{"_id":"q-en-kubernetes-91fe6ceffd0a40ec3ded85dd062721e869736715f793aa77d338c920cf20d98f","text":"CA Advisor runs inside docker containers, but I don't believe it's correct to show the Kubelet and Proxy processes as inside the docker box on the minion. Also, shouldn't the internet-firewall-really load balance across minion proxies instead of be connected to just one SPOF minion, treating that connection more like a logical connection across all minion proxies?\nThis is the desired state, which is admittedly not the current state. Yeah, this is a good point. k8s currently relies on its environment for load balancing. Maybe or can speak to the desired state here.\nWe should just review the diagram and make it correctly reflect the state of 1.0."}
{"_id":"q-en-kubernetes-c2e5391174f153a1c9277af876ed0c0c2fa7a8f69a8e1b77e99161308c31e686","text":"I can see that this message is flooding my minion logs so I was curious what should this config file looks like. I was trying to find some example, but with no luck. Also is this config file created during the runtime? What actually creates it? Shouldn't we just give up when the config file does not exists? Also there are better ways (maybe) to do polling if the file exists or not. I meaning you can use inotify () instead of an infinite loop + sleep. I'm happy to make a patch if you think it is a good idea. PS: I know that you put the timeout 5 seconds before retries.\nThanks for the report. I think this is just a bug-- we should check whether it exists and not pollute the log. I'm pretty sure we don't currently even use that file. Also we should NOT be reading the proxy's config from a file in /tmp where anybody can write it. Bad default location.\nProxy config sounds like an /etc thing, same for kubelet's container manifests from disk\nAre we interested in polling for the configuration files changes? Or do we want to read the settings just once, as an ordinary unix binary?\nYes, polling is a requirement. This is a live config file like the kubelet's machine files."}
{"_id":"q-en-kubernetes-7240f013dec109f7f6f3db86b7f71f4f168b8139d5a4174b363b9b39856ae102","text":"Had a report of this over email. Filing this issue so we don't forget to investigate.\nCan't reproduce."}
{"_id":"q-en-kubernetes-da9d81cdbfc64fa83c7a8102d516c422ecb439ff0b0723a5f46e6b4cdef1149c","text":"The hash function generates names with upper-case letters, but the validation only allows lower-case. First, we clearly need a test for this. Second, we need to fix it. Should we allow upper-case characters (which the RFC does allow for DNS labels) or should we fix the hash? Allowing upper-case seems attractive, except that the RFC also says that case doesn't matter, but should be preserved \"wherever possible\". So we really have three choices, I see. a) only allow lower-case b) allow upper-case and always use caseless comparisons everywhere c) allow upper-case, but normalize to lower-case since it was your hash code\nRe: test - sounds like a config source -config integration test, which I'll add. Re: hash - would argue we should fix the hash to generate lowercase for now. I think, but will double check, that the collision space is still improbably low for 63 lowercase chars. I can do that shortly.\nAnd a validate test on all sources.\nLowercase would require base32, rather than base64, no? On Aug 3, 2014 8:20 AM, \"Clayton Coleman\" wrote:\nYeah, although I don't believe there's a difference to the outcome between base32 and base64(lowercase) given an appropriate hash function.\nWell, base32 is reversible and a modified base64is not :) On Aug 3, 2014 9:00 AM, \"Clayton Coleman\" wrote:\nRemind me again why reversibility is important? We discussed it in another issue but I can't find the context.\nI'm going to change the generated name to which even with 2^8 files per host and a thousand hosts should be well below 0.% chance of collision\nIts only important if we want to use it. We don't AFAIK. On Aug 3, 2014 9:34 AM, \"Clayton Coleman\" wrote:\nFix makes them unique, somewhat traceable (first 15 safe chars of the name), and generates base32 output."}
{"_id":"q-en-kubernetes-c5eabf1d549f41dc6ff35936ea16374bcf7ff68f6c6211b1fe87c9dc27d00d31","text":"It's causing problems and is kind-of GCE dependent anyway, we should just use the kubecfg binary, and document the environment variables it uses.\nI think its beneficial to keep something, but remove the GCE dependent pieces, i.e. \"external-ip\" piece.\nI put together some instructions here:\nWe can remove this when kubecfg is deprecated .\nReopened since it still needs to be removed. /cc\nDoh. Didn't mean to declare this fixed in that PR.\nClosing this (again) as it is a dupe of"}
{"_id":"q-en-kubernetes-9d9eb5f8f6da0df47984d5529d5299f1f0821654a05b1fd79fc26d6b03acedb8","text":"Salt was failing to install on the master vagrant instance. The change noted by the FIXME solved it.\nIt's possible the required merge in the upstream repo occurred, will take a look.\nI just attempted to install a new cluster from upstream/master and was not able to reproduce the issue described. Was there an issue with curl-ing that specific version of salt-bootstrap? I guess more details on what the failure you encountered appeared as would help know how to proceed.\nI will submit a PR to revert back to using the default non-versioned bootstrap URL. Salt-stack PR: Fixed the errors in install that were previously encountered around salt-api not being found.\nSee , thanks!"}
{"_id":"q-en-kubernetes-009e7492cbeef180b9883cfe661e04a8dca4a46ed7fabb95f8d66a2876df42c7","text":"The Kubelet REST end-point has turned into a complicated switch-case code block that is harder to maintain overtime. The Kubelet should move to a http.ServeMux pattern and register handler functions to a pattern.\nI am working on a PR to submit in next couple of days that will make this change, so feel free to assign to me if possible.\nTagging because I asked for this in one of his PRs\nThis one is also under my radar. If no one had time, feel free to assign it to me too.\nI have a branch in progress. Feel free to offer feedback on it when I submit. Sent from my iPhone"}
{"_id":"q-en-kubernetes-ba201f6cc37fd09086cd86a1a7a1d3d232b0107e0e285e64b6f2a09676e0b920","text":"In the guestbook example, the redis slaves are instructed to connect to the redis-master using the proxy provided by the minion on which they reside. The minion is designated by its hostname \"kubernetes-minion-N\". The container cannot resolve this name to an IP address and so slave connections fail. 1) create vagrant cluster 2) create redis-master pod 3) create redis-master service 4) create redis-slave replicationController Verify redis-slave pod locations: list pods ID Image(s) Host Labels Status redis-master-2 dockerfile/redis kubernetes-minion-1/10.245.2.2 name=redis-master Running -3839-11e4-a922- brendanburns/redis-slave kubernetes-minion-1/10.245.2.2 name=redisslave,replicationController=redisSlaveController Running -3839-11e4-a922- brendanburns/redis-slave kubernetes-minion-3/10.245.2.4 name=redisslave,replicationController=redisSlaveController Running Log onto a minion containing a slave (note: vagrant hostnames are minion-N, not kubernetes-minion-N) Find the slave container docker ps CONTAINER ID IMAGE COMMAND CREATED STATUS PORTS NAMES brendanburns/redis-slave:latest /bin/sh -c About an hour ago Up About an hour k8s--slave.---3839-11e4-a922- docker logs | tail [7] 09 Sep 17:27:08.428  Connecting to MASTER kubernetes-minion-1: [7] 09 Sep 17:27:08.430 # Unable to connect to MASTER: No such file or directory [7] 09 Sep 17:27:09.435  Connecting to MASTER kubernetes-minion-1: sudo nsenter -m -u -n -i -p -t $(docker inspect --format '{{ .State.Pid }}' ) /bin/bash [ root ]$ ping kubernetes-minion-1 ping: unknown host kubernetes-minion-1 [ root ]$ cat /etc/hosts 244.1.4 -3839-11e4-a922- 127.0.0.1 localhost ::1 localhost ip6-localhost ip6-loopback fe00::0 ip6-localnet ff00::0 ip6-mcastprefix ff02::1 ip6-allnodes ff02::2 ip6-allrouters [ root ]$ cat nameserver 10.0.2.3 cat /proc/1/environ HOME=/rootPATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/binHOSTNAME=-3839-11e4-a922-0800279696e1REDISMASTERSERVICEPORT=10000REDISMASTERPORT=tcp://kubernetes-minion-1:10000REDISMASTERPORT6379TCP=tcp://kubernetes-minion-1:10000REDISMASTERPORT6379TCPPROTO=tcpREDISMASTERPORT6379TCPPORT=10000REDISMASTERPORT6379TCPADDR=kubernetes-minion-1SERVICE_HOST=kubernetes-\nTim - are you the right guy to handle this? I think we need to resolve the name of the service host and only give IPs to the minion in env variables. We can't assume the containers have the same DNS set up as the host system.\nTo be clear -- this example is for the checked in Vagrant stuff? If that is the case would be the guy? I'm trying to catch up on IRC discussions from earlier today.\nI can look at this, but there are two possible solutions. The first solution is in the vagrant cluster (or any other local vm setup) we could run a dns name server on the kubernetes-master that can resolve each minion by host name. We would then pass a -dns option to the docker daemon to specify that name server to use in each container. The second solution is that we pass IP address instead of host name and make no assumption or override on the containers specified name server. Is there any obvious negative to this solution? Has much thought been done on the expected Nameserver used by containers spawned by our service? - if you can weigh in, I would be happy to resolve using either approach. My preference is solution two.\nSo, the vagrant case is a little special as there is no sane DNS set up by default. Another suggestion (3) would be to revert the stuff and just devolve to using ip addresses directly. I would do (3) before (2) before (1). But I'm cool with any of them. As for the nameserver of the containers that are spawned -- docker currently just takes the of the host and copies that into the container. It doesn't do that for . that is the core of the issue here. If/when we do offer some level of split horizon DNS (I'd love to) we can start exposing an extra layer of DNS on top of whatever is configured for the host. That is part of the \"ip per service portal\" stuff that is being talked about a lot right now.\nIf I do (3) there is a piece of code in pod storage that mangles my hostname where I return 10.245.x.x and it asks the cloud provider to resolve the IP address for \"10\". Since other providers appear to also be returning IP as host, I will look to change this so kubernetes makes no modifications to hostnames that it's given. Look for PR tomorrow.\nyou're right I'd missed that point. This really only applies to the Vagrant cluster and so a fix to the vagrant setup is sufficient. I also agree with your priorities. Injecting a line into /etc/hosts (if that's reasonable) is the simplest solution. Either reverting or allowing a switch to pass the minion public IP rather than hostname is next. Adding a DNS service to the vagrant master and forcing DNS resolution to the master host is most complicated and least preferable. Thanks both of you for helping me think this through better. Mark\nI don't grok what the difference between 2 and 3 is? I sent a PR to docker a while back () to insert arbitrary lines to /etc/hosts. I still think it is a reasonable thing to want to do.\n- difference between (2) and (3) is that in the past, the vagrant cloud provider was returning the IP address as the literal hostname for the minion. I had changed this after noticing that it had some ugly side-effects in the CLI, where a call to list pods would report pod ip addresses as IP / nil. The issue (which I now believe is a bug that I want to change independent of what we do here) is that the code at this line mangles the hostname returned by the cloudprovider: This makes using the IP address as a hostname (10.245.2.2) to have the ugly side-effect of the code then asking the cloudprovider for the IP address of hostname \"10\", and then the cloudprovider having no idea how to proceed. Independent of what is done in this issue, I want to submit a PR tomorrow to stop that code from manipulating the reported hostname that was provided by the cloudprovider before asking that same cloudprovider to return the IP address using a potentially different input value. so with that backlog (3) means revert vagrant cloudprovider to use IPAddress as hostname, which means when we pass in what we believe is hostname in this environment, it will just work because it is in-fact the IPAddress. (2) means do not pass in hostname but pass in IPAddress by changing more core k8s code. all that being said, there is an option (4) that I am intrigued by, - after following your PR through its final conclusion, it looks like as of docker 1.2.0 you can change the /etc/hosts file of a running container. this introduces an option where the kubelet injects the /etc/hosts file of minion into the running container so it has the /etc/hosts values of the minion. I would need to play with that support a little more to know pros/cons more. I think longer term, we need a plan on what nameserver our containers are expecting to reference, and if they are inherently different in some way. I have run into a number of times the issue where doing things like running docker in docker causes the nested container to default to the nameserver 8.8.8.8, which happens to be blocked on the Red Hat network ;-) For now to unblock I will probably pursue (3) and fix the hostname bug referenced here at the same time. It does make the CLI uglier when working with this environment, but not a huge deal, at least it will function.\nThanks both for the complete explanation of what's happening and for the suggested fixes. I think in the long run using hostnames in SERVICEHOST and ensuriing either through IP injection into container /etc/hosts or by providing DNS are the best options. I'm not sure if there might still be cases where resolution of SERVICEHOST would still be best done by presenting the IP address or not."}
{"_id":"q-en-kubernetes-fb3b457ee434f4c000eee30c833933190e7e6916431cae811f57cb551a433ff1","text":"Kubernetes use kubernetes/pause image to create network container. The problem is that whatever signals the pause container received, it will exit with code 2. Instead, with a normal exit, it should exit with code 0; while it exit with port binding issue, it should exit with code -1.\nThe port-binding issues will occur before the pause image gets invoked, but the pause image itself should exit 0.\nI filed to docker to capture port-binding issue there. Port-binding is not my concern since we validate the port to avoid the conflict anyway. The concern I am having is PreStart we plan to introduce between docker container creation and start ().\nThere seems to be another issue - quick testing shows that if there is a conflict with a port allocated OUTSIDE of docker (e.g. the kube-proxy), Docker will accept the container and run it, but the host port doesn't work at all, and there's no way to tell except trawling logs. , Dawn Chen wrote:\nIs there anything we could do to help go through, to eliminate the need for the network container?\nGood catch. But following your steps above, my quick testing shows that docker container is running, but doesn't work at all. Also daemon logs doesn't have any error information. Will file another issue to docker: 1) # netstat -lnt Active Internet connections (only servers) Proto Recv-Q Send-Q Local Address Foreign Address State tcp 0 0 0.0.0.0:22 0.0.0.0: LISTEN tcp6 0 0 :::4194 ::: LISTEN tcp6 0 0 ::: ::: LISTEN tcp6 0 0 :::8080 ::: LISTEN tcp6 0 0 :::22 :::* LISTEN 2) # docker run -p :80 -d dockerimages/apache 6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11 3) # docker ps -a CONTAINER ID IMAGE COMMAND CREATED STATUS PORTS NAMES dockerimages/apache:latest \"apache2 -DFOREGROUN 6 seconds ago Up 5 seconds 0.0.0.0:-80/tcp jollynewton 4) # docker logs AH00558: apache2: Could not reliably determine the server's fully qualified domain name, using 10.244.3.11. Set the 'ServerName' directive globally to suppress this message AH00558: apache2: Could not reliably determine the server's fully qualified domain name, using 10.244.3.11. Set the 'ServerName' directive globally to suppress this message [Tue Sep 30 00:03:16. 2014] [core:warn] [pid 1:tid ] AH00098: pid file overwritten -- Unclean shutdown of previous Apache run? [Tue Sep 30 00:03:16. 2014] [mpmevent:notice] [pid 1:tid ] AH00489: Apache/2.4.7 (Ubuntu) configured -- resuming normal operations [Tue Sep 30 00:03:16. 2014] [core:notice] [pid 1:tid ] AH00094: Command line: 'apache2 -D FOREGROUND' 5) # cat | grep [] +job log(create, 6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11, dockerimages/apache:latest) [] -job log(create, 6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11, dockerimages/apache:latest) = OK (0) [info] POST /v1.14/containers/6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11/start [] +job start(6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11) [] +job allocateinterface(6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11) [] -job allocateinterface(6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11) = OK (0) [] +job allocateport(6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11) [] -job allocateport(6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11) = OK (0) [] +job log(start, 6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11, dockerimages/apache:latest) [] -job log(start, 6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11, dockerimages/apache:latest) = OK (0) [info] GET /containers/6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11/json [] +job containerinspect(6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11) [] -job start(6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11) = OK (0) [] -job containerinspect(6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11) = OK (0) [info] GET /containers/6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11/json [] +job containerinspect(6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11) [] -job containerinspect(6f7309782be63f95dbc3f87bded69637e3299cd255668f1f171dc5eef28f5a11) = OK (0) [info] GET /v1.14/containers//json [] +job containerinspect() [] -job containerinspect() = OK (0) [info] GET /containers//logs?stderr=1&stdout=1&tail=all [] +job containerinspect() [] -job containerinspect() = OK (0) [] +job logs() [] -job logs() = OK (0)\nI agreed we should help docker/docker go through, so that network container should be gone forever which makes our system, especially kubelet much simple. But that doesn't help with docker/docker and a new issue found by following.\nhave you tried this on docker master because it should have been fixed by , just want to prevent duplicate issues\nI just rebuild docker from HEAD, and tested it with above steps I posted. The problem is fixed: 37a039245cd9faba6ae900cc463313eae77c9c9dce589616ec992a7e8eecb458 2014/09/30 16:29:20 Error response from daemon: Cannot start container 37a039245cd9faba6ae900cc463313eae77c9c9dce589616ec992a7e8eecb458: Bind for 0.0.0.0:8080 failed: port is already allocated But another issue I filed (docker/docker) is still exist. Thanks!\ngreat, just wanted to make sure! and yes I can see if want to change the existing functionality for docker/docker"}
{"_id":"q-en-kubernetes-c13ad6ae8245ad270ca4d0b7b122e9cd5dbf6367a9eed33f85c1a7904dfa7c4f","text":"It's not well tested nor widely used; so it should either be removed or have tests and uses ."}
{"_id":"q-en-kubernetes-7c5448078f172c3dba4d1f6f6ed5d66f21d2fc0ea9dd38dbb1ed9b4cc49eea15","text":"Right now, we have a mixture of user documentation and developer documentation in the top-level directory. We should make it easier to find a just user documentation. Step 1 would be to move developer documentation into a subdir like docs/development or something like that. User docs would stay in the top-level docs dir.\nmove to"}
{"_id":"q-en-kubernetes-12a3ede7d945c55e14c6bf835d8548e1819c9a378123160dd4714cd3f7d4519c","text":"On Fedora20 I have the following error: [fedora ~]$ sudo kube-proxy.service - Kubernetes Kube-Proxy Server Loaded: loaded (/usr/lib/systemd/system/kube-proxy.service; enabled) Active: active (running) since mer 2014-10-29 15:15:21 UTC; 24ms ago Docs: Main PID: (kube-proxy) CGroup: /usr/bin/kube-proxy --logtostderr=true --v=0 --etcdservers=http://fed-master:4001 ott 29 15:15:21 fed-minion systemd[1]: Started Kubernetes Kube-Proxy Server. ott 29 15:15:21 fed-minion kube-proxy[]: I1029 15:15:21. ] etcd DEBUG: get registry/services/specs...:4001] ott 29 15:15:21 fed-minion kube-proxy[]: I1029 15:15:21. ] etcd DEBUG: Connecting to etcd: attempt...=false ott 29 15:15:21 fed-minion kube-proxy[]: I1029 15:15:21. ] etcd DEBUG: http://fed- GET ott 29 15:15:21 fed-minion kube-proxy[]: I1029 15:15:21. ] etcd DEBUG: recv.response.fromhttp://fe...=false ott 29 15:15:21 fed-minion kube-proxy[]: I1029 15:15:21. ] etcd DEBUG: Hint: Some lines were ellipsized, use -l to show in full. kubelet.service - Kubernetes Kubelet Server Loaded: loaded (/usr/lib/systemd/system/kubelet.service; enabled) Active: activating (auto-restart) (Result: exit-code) since mer 2014-10-29 15:15:21 UTC; 29ms ago Docs: Main PID: (code=exited, status=2) ott 29 15:15:21 fed-minion systemd[1]: kubelet.service: main process exited, code=exited, status=2/INVALIDARGUMENT ott 29 15:15:21 fed-minion systemd[1]: Unit kubelet.service entered failed state. docker.service - Docker Application Container Engine Loaded: loaded (/usr/lib/systemd/system/docker.service; enabled) Active: active (running) since mer 2014-10-29 15:15:21 UTC; 28ms ago Docs: Main PID: (docker) CGroup: /usr/bin/docker -d -H fd:// --selinux-enabled ott 29 15:15:21 fed-minion docker[]: 2014/10/29 15:15:21 docker daemon: 1.3.0 /1.3.0; execdriver: native; graphdriver: ott 29 15:15:21 fed-minion docker[]: [] +job serveapi(fd://) ott 29 15:15:21 fed-minion docker[]: [info] Listening for HTTP on fd () ott 29 15:15:21 fed-minion docker[]: [] +job initnetworkdriver() ott 29 15:15:21 fed-minion docker[]: [] -job init_networkdriver() = OK (0) ott 29 15:15:21 fed-minion docker[]: [info] Loading containers: ott 29 15:15:21 fed-minion docker[]: [info] : done. ott 29 15:15:21 fed-minion docker[]: [] +job acceptconnections() ott 29 15:15:21 fed-minion docker[]: [] -job acceptconnections() = OK (0) ott 29 15:15:21 fed-minion systemd[1]: Started Docker Application Container Engine. When I run netstat -tulnp (cadvisor)\" I don't see anything in output.\nExit code of 2 implies glog.Fatalf to me. What are the contents of the logs for kubelet and kube-proxy? What was the command line of kubelet?\nPlease reopen if you have more details."}
{"_id":"q-en-kubernetes-ddb3c77615a6f61140a3f455ef45c381d55feb4351606a018aa9cb9944c9a41d","text":"introduces PullPolicies (PullAlways, PullNever, PullIfNotPresent), but by default, the policy is PullAlways, which bite docker repository again today. The related #google-containers IRC conversation can be found at: We should Introduce a new policy to cap the number of attempts. Maybe getting rid of PullAlways completely? cc/\nI would like to change the default one to PullIfNoPresent for now. We can continue discussing the right policies.\nSuggestions: PullIfNotPresent should be our default? PullAlways needs to be nerfed. Suggestion: max N pull attempts per pod config change, with <attempt#*10 second delay between each pull attempt, where N is a small number like 3 or 5.\nI think both PullIfNotPresent and PullAlways (if we decide to still have it) should be nerfed.\nAgreed, when \"not present\", we should behave the same as PullAlways, which should be nerfed.\nNot finding another issue, so recording here for posterity: we really need a TTL on images to make PullIfNotPresent maximally useful.\nSo if a registry is down and the cluster restarts, pods will eventually give up (if we're nerfing pull always)? Seems like we should drastically reduce pull frequency after failures, but giving up means you can get into weird various scenarios. Esp if the registry is hosted in the cluster. Why not just back off?\nI vote for some exponential back off. , Clayton Coleman wrote:\nWe think we do have backoff now, check and close it if so."}
{"_id":"q-en-kubernetes-b0f40bfe0350c828d92f94d76b7846c99a5446ea0a208e561c5b1550f68762be","text":"From kubelet logs: E1117 23:17:17. ] Sleeping: Unable to write event: event \"e2e-test-etune-minion-1.c.seventh-circle-487.internal.\" is invalid: involvedObject.namespace: invalid value '' The kubelet code says: I'll update when I figure out where the record.Event is coming from. The code is:\nEvent in question is: I1117 23:25:42. ] Event(api.ObjectReference{Kind:\"Minion\", Namespace:\"\", Name:\"e2e-test-etune-minion-1.c.seventh-circle-487.internal\", UID:\"e2e-test-etune-minion-1.c.seventh-circle-487.internal\", APIVersion:\"\", ResourceVersion:\"\", FieldPath:\"\"}): status: '', reason: 'starting' Starting kubelet.\nI think the namespace needs to be kubelet. Note that one bad event blocks other good events from getting sent. Probably better to change that.\nI'll file a new one for bad event blocking good.\nI have a similar issue related to event. When generating event for image, what kind of namespace I should use? Image is not APIObject kubernetes managed, and it could shared crossing several containers even belongs to different pod. Currently I plan to just use default when creating ObjectReference for image related event. If we use kubelet for event related to minion, are we going to make it reserved one?\nwhat events are you planning to add for images? \"default\" seems like a good enough namespace for minion events for now."}
{"_id":"q-en-kubernetes-6a98eaf0134f550b05a03ee7991e3bdcaaccc7953bd821682f7b5b34771c6a6b","text":"For example, creating a service with id 'redis-master10012345678999' fails and the returned message contains the following: \"message\": \"service \"redis-master10012345678999\" is invalid: name: invalid value 'redis-master10012345678999' 'invalid value' is too general, users should understand exactly what is wrong with the input. for example - is it too long? does it contain characters that are not allowed?\nAgree - there is now a place to stick the explanations in the ValidationError - we need to link together validation logic and that string. , abonas wrote:\nI checked and it seems there are only two errors that do not have detailed explanation. I'll send a PR to fix this.\nThis is one:\nThanks for the pointer Brian! I'll send second PR to fix this one.\nStill second PR to be sent."}
{"_id":"q-en-kubernetes-bdb8aba44fdfc136ba5914129c5151c5b123e60df26d898975d42cdde5aeebc3","text":"Kubelet now only reports ContainerStatus. PodStatus / PodCondition is observed at master side based on a set of ContainerStatus. There is no other way to get PodIP except net container for master side. This is very confusion to the end user, , , ... Now we are finalizing on v1beta3 API, including PodStatus & ContainerStatus. We can make net container completely invisible to the end users by finishing the following steps: 1) Expose network related errors at API level 2) Generate PodStaus directly from Kubelet\n3) once PodStatus generated by kubelet, PodIP can be removed from ContainerStatus.\nI think reporting PodStatus rather than just container status is most important. How hard is that?\nUsed to be hard since there are many dependencies when I first report Status back. Now it is totally feasible. I already started to write PRs for this. :-)\nOne more PR required to close this one."}
{"_id":"q-en-kubernetes-e082935199b3867216aadc7ab3099030787c94aab618524e13109de89e7d352f","text":"In case it isn't clear why this happened, it's because gcloud ships with an older version of kubectl (currently version 0.7.0). We keep it as up to date as we can, but it will always lag a release or so behind head.\nFixed by the revert."}
{"_id":"q-en-kubernetes-8d9e539297dc070efbb3a72a40e6c4dd30fcde81ce3b97c5e3f1b6dd525ae415","text":"I'm trying containervm with kubelet built from head and this YAML file: The YAML file is v1beta2 and doesn't set a name to the Pod. When starting kubelet, it doesn't start the \"nc\" echo server and complains about this on the logs: Chatting with on IM she mentioned it might be due to a change from that now requires non-empty pod names (replaces random generation with validation for non-emptiness) but that means all existing YAML files (even the ones with an older version) will stop working... Is that expected? Thanks! Filipe\nYes, the latest kubelet call validation on all sources: etcd, file, http, etc. now, and PodName is a required field. One potential fix for this is also invoke conversion code before call validation, like what we have done at apiserver. Thoughts?\nJust so I get unblocked... Do you have an example YAML file that will set the pod name and will work? Right now this is what I'm using: Should I bump up version to v1beta3 and include a new field somewhere? Thanks! Filipe\nThe type of that yaml is not pod, and the kubelet code doesn't know how to read pods from a file. So, I think change needs to be reverted before you can make the next ContainerVM release. The v1beta2 vs v1beta3 should not have anything to do with it, unless I'm misunderstanding.\nOr will it work to do this:\nStill, not sure if you want to release a containerVM version that doesn't accept the old format?\nThanks I made it work with the YAML above! I still have errors in the logs: But the docker containers are being brought up... It's weird that is the YAML field that is missing, considering the source seems to say it's a deprecated field that will be removed at some point... I agree with your point that this needs to be fixed at head before the next containervm release. I opened to increase test coverage of standalone kubelet since we have been seeing many changes break that model recently. Cheers, Filipe\nid in v1beta2 maps to name in v1beta3 Tim's change made name required. The internal representation is like v1beta3. But v1beta3 is not officially stable so we don't want to use it in the GKE or containerVM docs.\nThe issue is caused by , and affect both v0.8 and v0.9 release for standalone kubelet to support old format ContainerManifests.\nUgh, I thought ID was really required. I spoke with Dawn a bit, and here's my thinking. 1) update all our docs and examples to include ID 2) in file and HTTP config sources, check for Name = \"\" and apply a default, maybe log a warning that this is deprecated 3) wait for the ContainerManifest -Pod conversion to really make it required"}
{"_id":"q-en-kubernetes-7c037c8cde80ec7f20aa2f250e008d71bb0af3d2acb8876fcb8161235bc3e00f","text":"Part of issue . Re-write test in Go, generate unique names for top level resources, ensure test is idempotent, clean up all resources at end of test.\nWhat should we do about this test? The current test in only brings up the guestbook but doesn't even confirm that it is up or tries to connect to any of the ports... Was the point of that test case to test /? Does it make sense to rewrite this test in Go or do we already have more test coverage from e.g. ? If it makes sense to rewrite it in Go, what should we accomplish with this test?\nHow about a functional test that inserts some names and tries to retrieve them i.e. enough to know that the service has basic functionality?\n+1 to an actual functional test. The service is basic enough, it should be straightforward to do the same POST/GET from go that the js on the guestbook homepage does.\nand I talked about this and for a while and think that they should remain as tests. This test needs to be improved: it needs to be turned into an actual test (we don't do anything with the service).\nhas not passed on gce or gke-ci since around 11am this morning (around when this PR was merged). Can you look at the Jenkins logs and try to figure out what is going wrong?\nThe problem is with guestbook itself not with the test. The frontend cannot connect to redis data base so it returns error response to the user. I created issue for that\nI have been 100% absorbed for weeks with my own e2e test so I've not paid much attention to this. Does the test connect to the service via an external load balancer?\nNope. It's using proxy. For example:\nIs this a reasonable use of the proxy service? Today told me that the proxy service is fine for contacting cluster-level services (note the dash Brian) but for regular applications we should use some other mechanism (e.g. external load balancer). ?\nMy first idea was to use ELB but it requires to create also firewall-rules using. I'm not sure if we want to do it.\nPR has been reverted by The test is passing for me locally. I'll updated . Is there any chance that using proxy instead of ELB can cause the problem?\nPersonally, I think -- for now -- using an ELB is fine -- by setting the use external load balancer field in the System specification. I am working on a PR to fix an issue relating to the use of application endpoints using the proxy server (to address issue -- which can bite you if one of the fields of a JSON response contains a field called ).\nIt sounds to me like there are 2 issues: should the guestbook example make the guestbook app accessible to users? should the test test correctness? I think these are 2 totally different things. The right answer to (1) is clearly external IPs. That's what we want to teach users to do. (2) likely doesn't have the same requirements. Conflating the 2 cases might not be helpful. However, if you want to do that, use the external load balancer, not the proxy.\nYes, I am torn about this too. Another thing to consider is that a test that uses an external load balancer is going to take longer because it takes much longer to create a service with an external load balancer (almost instant with ELB vs. perhaps up to a minute with a load balancer) and then there is a teardown step as well when you delete the service (which deletes the forwarding rules etc.). As a point of principle we should decide if it is an approved use of the proxy mechansim for use in tests in order to make them more lightweight vs. using an ELB.\nSo there is agreement to use ELB. What about firewall rule? Port 8000 should be opened, I can do it by using gcloud, however it's not provider agnostic operation. Is there any other option for that?\nSadly not: I suggest just using gcloud / making the API calls for the firewall and keep this as a GCE specific test. Which is indeed a shame.\nReopening since the PR was reverted in"}
{"_id":"q-en-kubernetes-8857d4913f590dc5c0dae632efd4f6f9d7848af1d63f93740d749815422076cf","text":"I stumbled across this while experimenting with kubeconfig load order/merging. There are various helpers constructed from a Factory that use , however this always uses the same default ClientConfig because doesn't use the passed in version param: Am I missing something or is this a bug? /cc\nLooks like a bug, clientforversion should accept \"\" (which can provide a default) and SetKubernetesDefaults has the final fallback protection (by using latest.Version). If version != defaultConfig.Version, we should copy defaultConfig, set version, then apply kubedefaults. The reason we take version here is if a resource is specified via file which only exists in a newer version, ie user says: --api-version=v1beta1 -f RESTMapper can handle that newresource is only valid for v1beta3 and apply the right override, then ask for client. --api-version is expected to bubble up via default config."}
{"_id":"q-en-kubernetes-bb0cafb0475cc6e5be0cdac2147f4b242ac75aa090a22ab5c84cfc4d4c244da0","text":"We should at least print the local client version before we try to create a server client for version checking.\n+1. It took me quite a while to figure out that kubectl version -c was the magic incantation to only print the client version.\nSorry, didn't realize it was assigned until after I had the PR ready :P If you also had something ready feel free to ignore mine."}
{"_id":"q-en-kubernetes-deb75b38aec5bebbcea339d029a1dfeb2da737ca01d1c960040375471f28ccca","text":"Part of umbrella issue . cc But this is my favorite shell test; it's so fast!"}
{"_id":"q-en-kubernetes-8fe8d41011716a5e2c28a734b3bece1d608434eb521eddfc89e8a0849b497adc","text":"I've reported this before.\nFrom a Makefile I get exit status 255.\nJust to be clear (sorry for the basic question), are you saying that when the service file is malformed the crash is unhelpful for debugging? Or that creating any/all services using is currently broken?\nThe YAML file is not malformed. I think there was something perhaps not quite right about firewalls/rules etc. and this caused to crash unhelpfully."}
{"_id":"q-en-kubernetes-b0f3dc2cf2d68fb8ee1e89f48638b658c3ae4124a5a63a21f7ced93b1e767d8a","text":"I thought I had worked this out during development, but from an IRC report from off a 0.10 GKE cluster, SkyDNS isn't starting. A bit of log trawling: Maybe I closed too early?\ngist with additional info:\n(FWIW, this came up the next time, and we test this setup all the time using \"Services should provide DNS for the cluster\" ) but even if this is a low grade flake, I take a bit of a personal affront.)\nThe cluster where SkyDNS didn't come up was created via the console at It was a cluster size of 2 on us-central1-b."}
{"_id":"q-en-kubernetes-dbcf83004ea3f0459daec6a4d94a4bece404fe8f97225f1a2fba429c343bf057","text":"I've been pulling my hair out trying get kubectl talk to my local cluster. It seems that kubectl merges the local file with values pulled down from the cloud. I've done the following: The I just set is silently overridden. Questions: is this documented anywhere? should kubectl merge in my GCE cluster when I set my cloud-provider to something else? should kubectl override my previously set current-context?\nworkaround is to use flag to override the config file value.\nbut not a good UX long term\nSomeone else hit this as well: . Better fix is to have the local-up script set a context, which I will try to add today. As for the overwriting behavior, it's documented in , but based on what you describe, the intended load order is not working; local .kubeconfig should overwrite global. cc\nThat sounds infuriating. I'll take a look this morning."}
{"_id":"q-en-kubernetes-7466b54034457325e304e8d94b4e7685272fe09d95978484a7adf32fd05db2b4","text":"I think the need for this argument to the factory methods has passed - it means that callers who want to provide configurable behavior have to do it through the factory itself, which is acceptable. We can remove it whenever.\nI removed most of the unused arguments. But, There are a few methods still using the argument. Namely, , , and . Before proceeding any further I wanted to ask what should be done about these methods? Deprecated/Refactored/Removed ?\nThe factory methods were the most important. The pattern before was people using the flags by name (cmd.Flags().Get(\"foo\")), but now factory methods must use pointer flag vars, so there's nothing that should need the command there. The printer methods are optional units that can be used when you add the flags for them to your command. Unless we refactor the definition of those flags to use pointers and return a printer struct (something I had started a while back and never finished), it's not worth changing them right now. The other consumer of factory is OpenShift - we wrap the factory to add our additional types (implementations of things like describers, action verbs, etc): - but we will pick up and remove use of cmd when we rebase.\nThanks for the explanation. I've cleaned the as well. Print method APIs are untouched as you suggested.\ncc\nLooks like this was done by PR"}
{"_id":"q-en-kubernetes-c692c87cef94fb1baca9eba906c310bcdc8b4d6325b69948d0a021b1972f05ef","text":"A significant number of builds fail with the error If you look at , this is the leading cause of flakes (build passes on 1.4, fails on 1.3 or vice versa).\nI am going to look into this one.\nI kept getting emails about shippable failures... Below is the snippet of the log that includes new log messages: The host field of the pod was not assigned and the check failed after 30 seconds of polling. This means either the scheduler didn't schedule the pod or it couldn't schedule the pods due to, say, conflicts. The pod in question was from I have a wild guess. Since usually only one runtime fails (1.3 or 1.4), could it have been port conflicts (hostPort: 8080) that failed the tests?\nIs this still a problem? Or can we close this issue now?\nSome bugs were discovered and fixed, but in general, timeouts are still a problem in integration tests. We can close this one though because there are other issue () tracking problems like this. is currently blocked on a reflector benchmark to stress test the watch performance."}
{"_id":"q-en-kubernetes-92291ba27b38eeaf8a2eb7cf1c2fd4ca17817007f4731e08ba809be4fd605c4a","text":"No other types have helper methods, and this does weird things with pointers.\nIt's a clean build/test without that helper method. See ."}
{"_id":"q-en-kubernetes-714ad5690812390e7d9a20665971045e80400b49fb24258316a1a9c3a074af93","text":"Many comments in there say \"if not specified, default is ...\" - but we previously asserted that there are no defaults for the internal API. We should strip those comments out, and leave them in versioned structs only.\ncc"}
{"_id":"q-en-kubernetes-399b2503d95e5517e810fb8fc23f0fca9215d555088a5ac25779615594a9bf09","text":"This is an offshoot of . cc Here's a log from a single showing that it's getting it's context screwed over by the other clusters coming up at the same time, even those its local environment never changed: Here's the relevant part of the new script: ( is in a separate commit and allows the override of on the GCE so that this script can work.)\nI think this can be resolved by having kube- set the context by exporting KUBECONFIG, so a subsequent run doesn't overwrite it.\nAfter some digging with it's a bit trickier than I thought. If we want to be able to run kube-up in parallel from the same branch/working directory, needs to create a separate kubeconfig for each run. But we also want kube-up to set config such that you can immediately run against the created cluster. Probably best to defer until is resolved.\ncc And the script is sitting in"}
{"_id":"q-en-kubernetes-1ae5c7614ddb98a42a715ebf40752bbc4cc609990675ef1d30825067e3fee6aa","text":"As per we should add a common label to all of our add-on pods: DNS, logging, monitoring, etc. The proposal is to add \"\": \"true\" for now.\nFWIW \"true\" has to be quoted in YAML\nCan you close this one once it's completed."}
{"_id":"q-en-kubernetes-8f0f8905f49f915362932082b98bd7c82f49f0968ad531a7f06a4941081c21b7","text":"It appears that the provision-network script that is run on the kubernetes-master can cause salt provisioning on the minions to fail or block in a forever retry timeout error loop. I think it's because of the service network.restart that happens after openvswitch is installed in provision-network. I only say this because when I disable installing docker on kubernetes-master in vagrant, I do not have an issue. - is there a different way we can structure this on the kubernetes-master? For example, could we run the script in or something?"}
{"_id":"q-en-kubernetes-c7cdd53074381ab5057ebc6ac3d6b49ac153717489df01e93d2a3ba62a136f72","text":"This works: but this does not: We should somehow emphasize the need for the trailing slash for now, and then work out what is going wrong (client side scripting?) to fix this (our fault in the proxy or Grafana's?)\nLikewise for Kibana.\nFor the Kibana case, I observe a sequence of correctly proxied GETs e.g. to: but then I see: which fails because I assume what was really needed is: and a curl against that gives a pile of plausible Javascript.\nAfter a little thought, I think what's going on here is that web clients can't tell that is a directory without the trailing slash. One thing we could potentially do to help is to redirect (301) from to , since I don't think the former is useful for anyone.\ncc\nHow are we generating that URL? Can we append a \"/\" at the end while generating the URL? Doing a redirect or auto-adding \"/\" in the proxy for all requests could break other requests.\nThe URL is not generated anywhere. It is assumed that '/proxy/service/<service name' would work. , Nikhil Jindal wrote:\nI discussed this with and To clarify, proxy server does not require a trailing slash. It works as expected (gives the same response with or without a trailing slash) The client code, on the other hand, expects it to be rendered on a path with a trailing slash and breaks when that is not the case. Ideally, we should fix the client code. For now, we decided to update proxy to send a 301 redirect from a path without a trailing slash to a path with a trailing slash. I will send a PR for this.\nBy \"client code\" do you mean things like Elasticsearch and Kibana?\nYes. The code being rendered on those paths.\nI agree with what said, I just want to clarify that it's not necessarily a \"broken\" client, just one with different expectations. Adding the 301 redirect seems like an easy way to satisfy such clients and reduce user pain.\nNot fixed because of\nOn GCE, the redirection is still not happening because the request is upgraded before we get a chance to redirect On localhost, the redirection works fine but we are unable to load the css and js files because of\nIs proxy still supported in v1beta3? I vaguely remember that we wanted to remove it. I tried this: but got NotFound error.\nAccess to cluster-level logging components depends on the proxy so avoid making ELBs. Elasticsearch and Kibana are accessed via the v1beta3 service proxy.\nAlthough the trailing slashes are no longer required."}
{"_id":"q-en-kubernetes-e83aef6ed0412aae3ae7527f665df46f7b642df7deb3335cb962e44d6695abb9","text":"On latest pull, the Vagrant kube- scripts completes successfully, but the Salt setup in the minion keeps going forever. End result is a cluster that is not useable. Looking at '/var/log/salt/minion' on the minion box, see lots of entries like: 2015-03-04 20:09:11,952 [ ][INFO ] Starting a new job with PID 2015-03-04 20:09:11,954 [ ][INFO ] Returning information for job: 2015-03-04 20:09:12,027 [ ][DEBUG ] Decrypting the current master AES key 2015-03-04 20:09:12,028 [ ][DEBUG ] Loaded minion key: 2015-03-04 20:09:12,217 [ ][INFO ] User vagrant Executing command with jid 2015-03-04 20:09:12,217 [ ][DEBUG ] Command details {'tgt_type': 'glob', 'jid': '', 'tgt': '*', 'ret': '', 'user': 'vagrant', 'arg': [], 'fun': ''} No other help looking at the master's salt log. This is Vagrant 1.7.2, with VirtualBox 4.3.20 on Mac OSX (Yosemite).\nIs this repeatable for you if kube-down and kube-up?\nYes, this is reproducible on my laptop. But, I was able to get it up by doing a kube-push after the failed kube-up. , Eric Tune wrote:\nSelf assigning and will try to reproduce. Sent from my iPhone\nI was able to reproduce. It appears that the Kubelet fails to start properly on first start:\nThe push basically caused the Kubelet to restart, and then it does work. Without the push, systemd would have eventually restarted the Kubelet. We need more investigation as to what was causing the Kubelet to fail on initial startup."}
{"_id":"q-en-kubernetes-17b7eb889bfb0d0f34c61424511db8cf312d22a6958eeab3e7ebc261ea16baf6","text":"I tried to spin up GCE kube environment using kube-up and got the following error: The script would then hang on a subsequent command. The verify-prereqs for kube-up should require yaml support for python and fail early. To rectify, I had to do the following:\nThanks for the report Derek!\nYeah, this broke our Jenkins a while ago, and I thought I fixed it, but didn't. The real fix is in"}
{"_id":"q-en-kubernetes-20b7e73eec176d141ed7021484976d2204427591b40ebcc573eca325f6704635","text":"If the cleanup of the replication controller takes too long after a successful test run, then the clean up performed in the AfterEach can actually cause the test to fail. Should only perform the clean up if the rc exists and its replica count is non-zero."}
{"_id":"q-en-kubernetes-6be393fdfa676676e2c2a919450eea6e64e7299e61f3fdbea996cf6d1af38cd8","text":"We've seen this a few times on GCE now, and I saw it flake on now as well (cc\ncc Victor integrated cadvisor into the kubelet. I wonder if a fake version of cadvisor gets integrated into kubelet in the case of builds.\nI noticed this first on GCE. (FWIW, GKE uses the same binaries as GCE.)\nJenkins shows the following four changes for the GCE build where this test started consistently failing: Commit by mattmoor Enable usage of a \"json key\" for authenticating with Commit by dbsmith Add a system modeler to scheduler Commit by dawnchen Convert RestartPolicy to string for v1beta3. Commit by abshah Build statically linked binaryies. With the change to go 1.4, we probably were generating dynamically linked binaries accidentally.\nIf I had to guess, I would take a stab and say it was , . cc\nFeel free to revert by 5pm if no fix.\nI believe and are correct, makes the build not use cgo which places a fake cAdvisor in the binary and that fails. Looking into the fix.\nis out, it is a revert of the old one along with a plan for not needing static binaries."}
{"_id":"q-en-kubernetes-c863156a4e33fc97f13e0b76d7c72461cca123b7f2e4e0958a06f940e1cd99d8","text":"cc\nWhen deleting/stopping using file, you don't specify resource (in this case 'pod'), as it's inferred from the file.\nThanks and sorry for confusion\nOk, but should not work as well."}
{"_id":"q-en-kubernetes-48c5dfb8683a81615afc4ea08a4324bb162203e3714a3b2fd8f08b4f5a9b8f4e","text":"A \"kubectl describe\" on a pod ID give a lot of informations about that pod but don't you think it should also print the pod's env var ?"}
{"_id":"q-en-kubernetes-dbffed5438f566ecf1b631adacf4d68392348225a8f199c6d9953211f55b24b3","text":"This is the script that runs to capture event output during e2e runs: It's failing after the recent pflag merge:\ncc\nOr something. I might be completely off. But something broke it recently enough, and I know like his event log.\nThat command is actually invalid. is a flag on the command, not the root. Fix in"}
{"_id":"q-en-kubernetes-6bdc2b90ff152c92e3bbe26723a057422401366398c588c04109e25e8a7ccf9b","text":"I used the ubuntu-cluster setup scripts to build a mini testing environment. The first time I inputted a wrong config and when I reconfigured the environment by running the scripts again an error occured. I looked into the code and found the scripts hadn't deal with old configs. And of kube-proxy configs are not replaced.\nLooks like this was fixed by ."}
{"_id":"q-en-kubernetes-cec696c80d7bb22acf8809e7c4a3d65bfaf721f6124d2319a7a6475d4f30f213","text":"After 6 minutes or so of using 50% my 4-core MBP CPU and chewing up a ton of memory, make release gives up with a timeout.\nThis is on my master branch (no local changes) after rebasing to head this evening. It was working fine earlier today so seems like the likely culprit. /cc\nIt's certainly my changes. I'm on a low function device, feel free to revert. This is only an issue for low memory systems, I suspect. :/ On Apr 1, 2015 10:24 PM, \"Robert Bailey\" wrote:\n(This is functioning fine on Jenkins and my desktop, FWIW.) On Apr 1, 2015 10:36 PM, \"Zachary Loafman\" wrote:"}
{"_id":"q-en-kubernetes-04040bdb5cd938b726629688668f9baf9bede652c53f08edb44f58e6b143c345","text":"Restarting might also be acceptable. It's a pretty painful user experience right now if this happens-- it fills the logs with this message.\nYes, it should just panic() , Daniel Smith wrote:\nYeah I guess if it's out of FDs it can't really open a network connection to file an event...\nkube-proxy should send an event when it starts up, just like kubelet does with its method. This works around the fact that you can't send an event when you are out of FDs, and it still provides a lot of debugging value.\nAgreed with 1) kube-proxy should panic when hitting out of fd 2) restarted kube-proxy should pick up the crash reason like what we do for internal agent 3) crash reason should be propogated to upstream as an event.\nMoving to the right sig.\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. Prevent issues from auto-closing with an comment. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/lifecycle frozen\n/remove-lifecycle stale\n/area kube-proxy\nwe're going to close this because, after 3 years or so , decreasing interest in the userspace proxy and nobody other then some openshift folks, are using the linux kube proxy .. can reopen if necessary ping :)\n/close\nClosing this issue. <detailsIn response to : /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-683f0af5e9cbef9ac9e42be8d3345c188822e579c19423287c94afa9fbba590f","text":"This was broken by and then (hopefully) fixed by . But we need a test for it. cc"}
{"_id":"q-en-kubernetes-1b08731726b30375f5b0183cb678680fc867aa4d0b55fd733e8704ed7e56ca62","text":"kubernetes]$ make hack/build- Detected go version: go version devel + Wed Mar 4 20:55:55 2015 +0000 linux/amd64.Kubernetes requires go version 1.2 or greater.Please install Go version 1.2 or later. !!! Error in 'kube::golang::buildbinaries \"$ exited with status 2 Call stack: 1: kube::golang::buildbinaries(...) 2: hack/build- main(...) Exiting with status 1 make: * [all] Error 1\nHello, it would be really nice feature to have and it seems easy to fix/implement. Are there any plans on fixing the build?"}
{"_id":"q-en-kubernetes-b7b85da4ef1fcf3b1c9916279f506d55083905ef54f59a6d8f39cac395ba8dbc","text":"The master sets its endpoints into the 'kubernetes' service, but endpoints can only be IPv4 today. So ChooseHostInterface() should either not return an IPv6 address, or we should allow endpoints to be IPv6.\nI think that this is likely the cause of .\nI thought that function explicitly checked for v4. got any bandwidth for review? On May 4, 2015 11:24 AM, \"Robert Bailey\" wrote:\nIt doesn't - and on my mac, it fails to find this address when I add that logic. Not sure if that's FlagPointToPoint or something: ----- Original Message -----\nPEBKAC for the issue I mentioned above, is opened to fix the choice.\nYeah, I can take a look at later this afternoon, just as soon as I've unbroken ELB's :-) smarterclayton@ let me know if is more urgent than that.\nwas smaller than I thought, so committed priority inversion I reviewed it :-)"}
{"_id":"q-en-kubernetes-0cb771279e2d0e1af0305b6952611afffca3feab7817e435889dc80d9ca21ce4","text":"This page: Links to this YAML: Which doesn't exist. There are several examples now in the NFS folder:\nThanks for reporting! Always happy to get docs up-to-date, especially to reduce friction. Please report / open PRs for any others. Also, if the examples / explanations aren't clear, we're happy to address issues for that as well."}
{"_id":"q-en-kubernetes-832c9b022e2d3879328d55b738068527b96cd04964ec313be5394fb24e6f52f7","text":"kubectl 0.17.0: curl http://localhost:8080/api/v1beta3/minions { \"kind\": \"Status\", \"apiVersion\": \"v1beta3\", \"metadata\": {}, \"status\": \"Failure\", \"message\": \"the server could not find the requested resource\", \"reason\": \"NotFound\", \"details\": {}, \"code\": 404 } master3 core # /opt/bin/kubectl get mi master3 core # kubectl 0.16.0: uses: http://localhost:8080/api/v1beta1/minions master3 core # ./kubectl get mi NAME LABELS STATUS 192.168.10.70 <noneReady 192.168.10.71 <noneReady in the same time curl http://localhost:8080/api/v1beta3/nodes works well.\nShall we just remove \"minion\" from kubectl now and throw an error?\nremove \"minions\"?\nYes, that should be part of . cc\nPresumably works?\nAssigning who's working on\nRef\nCould you please take a look at this?"}
{"_id":"q-en-kubernetes-cb019f3c5a5459d10f5c4e9d499a59fe3549588177be47b0dd8f7a52a6551849","text":"If a pod is not scheduled because it cannot find a fit, you run into really ugly event output like the following: This makes the output of horrific to work with. A better event would just read: The pod is known from the involved object already so no need to encode it all."}
{"_id":"q-en-kubernetes-7801592fe1a929421527a19ff1e4db70f69311885756bd569f596cfe383cc288","text":"The following lines are incorrect. Set your cluster as the default cluster to use: - Also the whole section is not so appropriate when it comes to issue. After all, certificate authentication works well without such token thing. Why trouble adding parm?"}
{"_id":"q-en-kubernetes-bf5172e25b0232eb28e1fcc7a2afb381c4a8c5c84602ffee854fe026df1028c9","text":"Hello, I am having issues with a deployment on AWS with version 1.0.1. I had a collision in security group when I launch a service and there already was a security group from an earlier deployment of the \"same\" cluster. I deleted the security group and then kubernetes managed to create the load balancer so thats good, but since then I am having these messages in a loop: E0729 14:48:07. 8 ] Could not delete route k8s-v1.0.1-10.244.2.0/24 10.244.2.0/24: error deleting AWS route (10.244.2.0/24): RequestLimitExceeded: Request limit exceeded. status code: 503, request id: [] E0729 14:48:07. 8 ] Could not create route -35f5-11e5-8163- 10.244.2.0/24: error configuring source-dest-check on instance i-: RequestLimitExceeded: Request limit exceeded. status code: 503, request id: [] E0729 14:48:07. 8 ] Could not create route -35f5-11e5-8163- 10.244.3.0/24: error creating AWS route (10.244.3.0/24): RequestLimitExceeded: Request limit exceeded. status code: 503, request id: [] E0729 14:48:17. 8 ] Could not create route -35f5-11e5-8163- 10.244.1.0/24: error configuring source-dest-check on instance i-: RequestLimitExceeded: Request limit exceeded. status code: 503, request id: [] How can I stop them at least to get under the limit request and try again ? Thanks\nI'm seeing the same thing, and AWS let me know they didn't like it: It seems like maybe when we describe the route tables Kubernetes is not seeing them as being consistent with what it expects internally. Any ideas on what the root cause might be here? I haven't looked at this piece closely yet, but will try to find and fix the issue.\nSome research - I believe this has to do with the name that Kubernetes has assigned to my instances internally. tells me that my node looks like this: Whereas describe-route-tables returns the amazon instance identifier: Based on the above, I'm guessing we should just use node.Spec.ExternalID instead of node.Name when doing the reconciliation. I'll start on a PR and testing this.\nI am not doing the calls, Kubernetes is for me :). It tried to do a load balancer and failed, after that I helped when i saw the logs saying that there already was some security group, deleted it and then it managed to do the load balancer. So i guess the routes making comes from that, My use case is not something out of the ordinary, I launched a reverse proxy and it tried to create an external load balancer on aws. I could try to create the routes manually in the aws console, just tell me what are the routes that kubernetes is trying to create. And you received a message from AWS on my account or is it one of yours ? Adam Guldemann\nYep, sorry, my flurry of details might have been confusing. What I was trying to say was: Kubernetes is doing the same thing on my AWS account and they actually sent me that message to let me know. The load balancer problem you saw is actually another issue that I believe has been merged and should be in the next release, see . No need to create the routes manually, they have probably been created by Kubernetes. The problem, at least as I understand it, is that Kubernetes tries to reconcile the routes it sees on AWS against what it knows internally. That reconciliation fails because of a mismatch between how AWS and Kubernetes identify nodes. I'm working on a fix for that and will have a PR pending some testing on my cluster. I'll keep you posted about the progress on that here.\nthank you for taking this one on - feels pretty serious if AWS are sending warnings out, so I appreciate it. Please tag me when the PR is up (and/or put up a WIP PR and I can test too!) LMK if there's anything I can do to assist.\nI attempted a fix for this on . I'm testing now and am considering adding some logging at a higher log level for future debugging.\nWe should probably add a rate limiter into the AWS client, like we did for GCE:\nagree. I can take a shot at that once I get my fix for route reconciliation done.\nI am having issues with redis / rethinkdb cluster on my AWS cluster, I am wondering if that could from the fact that it can't create routes or maybe deleting and recreating ? Here is a ping request from one container to another for like 5 minutes: .... 64 bytes from 10.244.1.10: icmpseq=962 ttl=62 time=0.693 ms 64 bytes from 10.244.1.10: icmpseq=963 ttl=62 time=0.670 ms 64 bytes from 10.244.1.10: icmpseq=964 ttl=62 time=0.636 ms 64 bytes from 10.244.1.10: icmpseq=965 ttl=62 time=0.732 ms 64 bytes from 10.244.1.10: icmpseq=966 ttl=62 time=0.590 ms 64 bytes from 10.244.1.10: icmpseq=967 ttl=62 time=0.533 ms 64 bytes from 10.244.1.10: icmpseq=968 ttl=62 time=0.544 ms 64 bytes from 10.244.1.10: icmpseq=969 ttl=62 time=0.605 ms 64 bytes from 10.244.1.10: icmpseq=970 ttl=62 time=0.636 ms 64 bytes from 10.244.1.10: icmpseq=971 ttl=62 time=0.603 ms 64 bytes from 10.244.1.10: icmpseq=972 ttl=62 time=0.548 ms 64 bytes from 10.244.1.10: icmpseq=973 ttl=62 time=0.632 ms 64 bytes from 10.244.1.10: icmpseq=974 ttl=62 time=0.596 ms 64 bytes from 10.244.1.10: icmpseq=975 ttl=62 time=0.436 ms 64 bytes from 10.244.1.10: icmpseq=976 ttl=62 time=0.516 ms 64 bytes from 10.244.1.10: icmpseq=977 ttl=62 time=0.660 ms 64 bytes from 10.244.1.10: icmpseq=978 ttl=62 time=0.532 ms 64 bytes from 10.244.1.10: icmpseq=979 ttl=62 time=0.544 ms 64 bytes from 10.244.1.10: icmpseq=980 ttl=62 time=0.588 ms 64 bytes from 10.244.1.10: icmpseq=981 ttl=62 time=0.553 ms 64 bytes from 10.244.1.10: icmpseq=982 ttl=62 time=0.604 ms 64 bytes from 10.244.1.10: icmpseq=983 ttl=62 time=0.380 ms 64 bytes from 10.244.1.10: icmpseq=984 ttl=62 time=0.531 ms 64 bytes from 10.244.1.10: icmpseq=985 ttl=62 time=0.569 ms 64 bytes from 10.244.1.10: icmp_seq=986 ttl=62 time=0.606 ms ^C--- 10.244.1.10 ping statistics --- 987 packets transmitted, 802 packets received, 18% packet loss round-trip min/avg/max/stddev = 0.323/0.547/4.269/0.165 ms I see 18% packet loss, any ideas where that could come from ?\nIf I ping google from the same container 0% packet loss to compare: 64 bytes from 8.8.8.8: icmpseq=193 ttl=54 time=0.988 ms 64 bytes from 8.8.8.8: icmpseq=194 ttl=54 time=1.014 ms 64 bytes from 8.8.8.8: icmpseq=195 ttl=54 time=0.952 ms 64 bytes from 8.8.8.8: icmpseq=196 ttl=54 time=1.069 ms 64 bytes from 8.8.8.8: icmpseq=197 ttl=54 time=1.038 ms 64 bytes from 8.8.8.8: icmpseq=198 ttl=54 time=1.094 ms 64 bytes from 8.8.8.8: icmpseq=199 ttl=54 time=1.083 ms 64 bytes from 8.8.8.8: icmpseq=200 ttl=54 time=1.005 ms 64 bytes from 8.8.8.8: icmpseq=201 ttl=54 time=1.042 ms 64 bytes from 8.8.8.8: icmpseq=202 ttl=54 time=1.014 ms ^C--- 8.8.8.8 ping statistics --- 203 packets transmitted, 203 packets received, 0% packet loss round-trip min/avg/max/stddev = 0.887/1.027/1.166/0.055 ms\nClosed, but obviously please reopen if the problem continues And I think we're tracking the ping problems separately in ; I have my fingers crossed that we will fix them with this path also.\nwe're def seeing this right now on AWS. We spun up about 40 nodes using the default k8s aws installation, and are at times crippled by this... Even to the point where pods fail to start because k8s is unable to mount their EBS volumes due to rate limits. Any ideas re workarounds? Anything we can do interim? FWIW, we've spoken to AWS premium support, and they are pretty adamant about not increasing their API quota. They claim that they should be sufficient, and that they won't even raise those for their biggest customers...\nI'm curious if you had any luck with this? Unfortunately 1 year later we are experiencing the same problems with EBS volumes. Makes me wonder if the Persistent Volume API makes a ton of calls for EBS volumes. As a side note, we are also having problems with terraform and even packer when running in the same account as our kubernetes cluster, which leads me to believe it is indeed the kubernetes cluster"}
{"_id":"q-en-kubernetes-907072de4daa6911db8f6a81adc78d97e42d008af334d61c5840b8a732fcb546","text":"When mounting several shares over NFS to container and when one of the directories is deleted from NFS server, something in the node is deleting content from other shares. I suspect that something goes wrong with \"cleanupOrphanedVolumes\"? not sure tough. Use case: php-app is running in container, uses php-sessions for remembering login. php sessions are stored in file and shared through nfs-share between containers. also app code is shared and several other directories to store different data (data dirs are mapped under code, code is in separate share so generated document files don't get to git tree), all is managed by separate integration server (in container), that uses same shares to prepare the code. This happened first time in GCE with Cluster API version 0.17.something, now it acts same on 1.0.1 and 0.18.2. Problem is repeatable: First of all you need to install package \"nfs-common\" to all cluster nodes, otherwise you can't map NFS shares at all: now you need nfs server. I created small instance for that with /etc/export stating: also I made directories under /export/test/: next you need to start it up (replace nfs server ip with your own): now start up the pod: verify, that the directories are mapped: now the funny part: remove directory from nfs server the zubuntu pod shows: nothing happens on nfs server yet, but just launch the stop command: and now nfs server shows: you see, that one file was left, but not for long. this is because root in client does not have root access in nfs-file system, but just after giving rights, the file is gone: now you can test \"black hole\" by creating new files. they dissaperar from the server in few moments. the dissapearing continues until cluster node is restarted.\nso the problem looks to me is that containers delete files though they are not told to do. I am not quite convinced is the culprit here, since the directories are still alive. what the first version you believe this problem started to show up? Would it be possible for you to take a tcpdump on the nfs server so we can look at the nfs traffic? Thanks!\nI started to play with the Google Container Engine about in middle of June and the version of the cluster was 0.17. something. I don’t have that cluster anymore. So I don’t know what happened or did it worked before. About tcpdump. What ports and what traffic do you think you need for that? Also in what timeframe? Maybe you can even give me command-line for that? Jüri.\nI am setting up an environment to simulate your tests, stay tuned\nyes, I can reproduce your issue. And I was wrong about . You are more likely to be right on that suspect. Blame me if I point you to the wrong suspect. Here in this line , doesn't return any error if a volume fails to be torn down, which is what happening here: the first volume is not able to be torn down, yet the error message is suppressed, so later on thinks volumes are unmounted and it is safe to remove the directory\nSomething is fishy. Volume cleanup should be happening against the volume dirs, which is totally separate from pod dirs. Pod teardown should un-bind-mount the NFS from th epod dir, so pod cleanup can happen, even if the volume can't umount...\nI didn't touch much of the volume code, but think the directory structure is . Removing pod directories would remove the volumes as well. We can delay the pod directory cleanup until all volumes in it are properly unmounted. Another user reported a similar case in\nin this particular case, the mountpoint cannot be properly unmounted (because it was deleted), and I believe the current NFS mount is , so it won't be easy to unmount even after delay. unmount error might be a good heuristic. In this case, NFS stale handle error is probably good sign nothing is good for wait for.\nAhh yes, it was only GCE PD that was doing something tricker here , Yu-Ju Hong wrote:\nI see. Thanks for the explanation. I didn't read the original post carefully. If we let propagate the error upstream, kubelet may always skip pod directory cleanup because of one failed teardown. That seems undesirable. I agree that the NFS stale handle error alone should be enough for us to skip waiting. In short, should distinguish the NFS error and report a successful teardown in this case. should check if there are any uncleaned volume directories, and avoid removing the specific pod.\ncool, thanks. When a patch is available, I'll give it a test on other volumes too.\nThis was completed by in PR can you take a crack at this?\nok, glad to do it.\nThanks!"}
{"_id":"q-en-kubernetes-cab9179de0369f40f8cfa54fff0d46c239e117acfc7e0e5362f896602ccc17de","text":"I follwed the Getting Started guide from and tried deploying a vagrant based cluster on RHEL 7, first by building the source code and then when that didn't work, using the curl script. In both cases the cluster seems to be set up and I can do a vagrant ssh to the master and a minion, but running kubectl get pods or for that matter almost all kubectl commands return: error: couldn't read version from server: Get https://10.245.1.2/api: dial tcp 10.245.1.2:443: no route to host I read somewhere that this could happen if your config was faulty, so I tried config info and this is what I get: config view apiVersion: v1 clusters: cluster: insecure-skip-tls-verify: true server: http://127.0.0.1:8080 name: local cluster: certificate-authority-data: REDACTED server: https://10.245.1.2 name: vagrant contexts: context: cluster: local user: \"\" name: local context: cluster: vagrant user: vagrant name: vagrant current-context: vagrant kind: Config preferences: {} users: name: vagrant user: client-certificate-data: REDACTED client-key-data: REDACTED password: vagrant username: vagrant I'd gotten this to work in a different machine successfully with v1beta3. Unfortunately I don't have access to that machine any more so don't know what the configuration had been set to. Any help would be much appreciated. Cheers --Rupa\nLooking at the config-default .sh file in the vagrant folder, I see that MASTERIP and KUBEMASTER_IP is set to 10.245.1.2. However pinging the IP address doesn't work. When I do a vagrant ssh master, and an ifconfig, the eth0 inet address is not 10.245.1.2. either.\nI am having the same problem. Running with ends with: are you seeing this on latest master? Edit: this just recently occurred when rebasing my working branch to upstream, I'm not sure exactly when this occured however as I have been working on the networking, but yesterday morning this worked fine.\nThere was a bug in the v1 release that caused network manager to control the eth1 device when it should not have been doing so. The fix for that was merged in HEAD a few weeks ago, and the cherry pick to go in the 1.0 release was made last night here: As for e2e, it had passed conformance tests, but I will see if I can reproduce the issue you are seeing now to determine the cause if I can reproduce at HEAD.\n- I was able to reproduce your error just now. There must have been some kind of regression in the last day. I see the kubelet starts on the master, and docker is running, but doing a shows no attempt to start the api server container. I will need to do some digging into what caused the regression in the last day.\n- think I found the offending PR that broke the setup here:\nthanks for checking this out, Is there anything I can do to help sort out the regression? Edit: read , looks like you have it covered already. Thanks again!\nSee fix above.\nOnce there's a fix, downloading the latest source from MASTER or using the curl script again should both work ?\nLatest source from MASTER should work fine. Please confirm. , rupsee wrote:\nConfirmed that latest master is again working, sorry it took so long to get back to you.\nI'll try it and post here tomorrow, it's the end of my day here now. Thanks a bunch for the quick turnaround.\nI still see a similar issue with a somewhat different error message, \"connection refused\" instead of \"no route to host\". Here's what I did: git clone cd kubernetes make quick-release export KUBERNETES_PROVIDER=vagrant ./cluster/kube- At the bottom of the master and minion updates I see this: xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx Wrote config for vagrant to Each machine instance has been created/updated. Now waiting for the Salt provisioning process to complete on each machine. This can take some time based on your network, disk, and cpu speed. It is possible for an error to occur during Salt provision of cluster and this could loop forever. Validating master Validating minion-1 ...... Waiting for each minion to be registered with cloud provider error: couldn't read version from server: Get https://10.245.1.2/api: dial tcp 10.245.1.2:443: connection refused xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx Then I tried: [root kubernetes]# get nodes error: couldn't read version from server: Get https://10.245.1.2/api: dial tcp 10.245.1.2:443: connection refused\nhere's part of what I get from ifconfig from the master: .................................................................................................. cbr0: flags=4163<UP,BROADCAST,RUNNING,MULTICASTmtu 1500 inet 10.246.0.1 netmask 255.255.255.0 broadcast 10.246.0.255 .................................................................................................. eth0: flags=4163<UP,BROADCAST,RUNNING,MULTICASTmtu 1500 inet 10.0.2.15 netmask 255.255.255.0 broadcast 10.0.2.255 .................................................................................................. eth1: flags=4163<UP,BROADCAST,RUNNING,MULTICASTmtu 1500 inet 10.245.1.2 netmask 255.255.255.0 broadcast 10.245.1.255\nHere's what \" config view\" returns: apiVersion: v1 clusters: cluster: insecure-skip-tls-verify: true server: http://127.0.0.1:8080 name: local cluster: certificate-authority-data: REDACTED server: https://10.245.1.2 name: vagrant contexts: context: cluster: local user: \"\" name: local context: cluster: vagrant user: vagrant name: vagrant current-context: vagrant kind: Config preferences: {} users: name: vagrant user: client-certificate-data: REDACTED client-key-data: REDACTED password: vagrant username: vagrant\nAnd finally,I'm on a RHEL 7 box - don't know if that matters. lsb_release -i -r Distributor ID: RedHatEnterpriseServer Release: 7.0\n- I am closing this issue since vagrant should be stable now on HEAD and in the release branches. Your connection refused error is most likely not being connected to the vboxnet device on your host to your VM that Virtual Box launches. To check, open Virtual Box and look at the Network Adapter setting for kubernetes-master. You should see something like the following: ! Make sure your host is connected to that adapter (in my case its vboxnet5) The other alternative is to just ssh into the master box and use kubectl from there: $ vagrant ssh master $ kubectl get nodes Either way should work well. Reopen if you continue to have issues.\nHi Derek, thanks for getting back. I should have closed this one ... It was because I hadn't set the KUBERNETRS_MASTER environment variable. None of the documentation talks about it though. I'd to figure it out from the cluster scripts in the back folder. Cheers Rupa On Sep 11, 2015 12:40 AM, \"Derek Carr\" wrote:\nI meant the hack folder, Swype seems to think it knows best! On Sep 11, 2015 6:43 AM, \"Sreerupa Sen\" wrote:"}
{"_id":"q-en-kubernetes-6bc8df82589815e2e610f5ccd928cca589126610fdc4779d12361c6b9ee84a34","text":"When I'm accessing persistentVolumes and persistentVolumeClaims, I'm getting nil instead of empty list of items, as happening in other entities when there are no items. { \"kind\": \"PersistentVolumeList\", \"apiVersion\": \"v1\", \"metadata\": { \"selfLink\": \"/api/v1/persistentvolumes\", \"resourceVersion\": \"\" } }\nYes, this seems like a bug. I see an empty list when I access pods: And items is nil when I access persistent volumes:\nItems shouldn't be tagged with omitempty:\ncc\nPersistentVolumeList and PersistentVolumeClaimList appear to be the only resource types with this problem.\nPlease see for fix."}
{"_id":"q-en-kubernetes-53453bd5ca3e0fd97097a9d5375a1c84f5c3ea86da4fd838c0b744a110f4390f","text":"Both the GKE and Parallel tests recently had a run where they timed out. (after 90 minutes!) This is like 10x longer than the parallel run should be taking.\nAnd the build project, too:\nany ideas?\np0\nyup, investigating this is next on my queue, right after I figure out why the PR jenkins is failing so many runs.\nThis happened again: In fact I've seen at least few cases like that: \"at some point tests just stop and completely nothing is happening to the end\" The same happened here - completely nothing happened during last hour of the run,\nWhile the failure rates of the individual test jobs is not very high (e.g. kubernetes-e2e-gce-parallel : MTTF (Last 7 Days) 7 hr 27 min ... the fact that we gate auto-merging on about 7 of these jobs means that a merge fails almost every hour due to this.\nI looked at 4 of the builds in kubernetes-e2e-gce-parallel, comparing the test cases marked \"N/A\" (i.e. no test result given) instead of their usual \"PASSED\" or \"SKIPPED\" report. So far, there's been only one test case that's been consistently N/A in all four of the builds I looked at: I'm not sure what might be happening, but I'm tempted to disable this test in parallel runs now while we investigate what's actually wrong.\nOK, that test has been disabled. I'm hoping that the PR builds recover (along with some git fixes as referenced in ). Now on to figuring out why this is causing us problems...\nHad a successful failure in build 8902. Grepping out the Ginkgo node number shows this test being the problem: So it looks like it got stuck in ? Is there anything in here that should hang without deadline?\nI had another run get stuck in exactly the same place: What's worrying is that another parallel build (http://go/k8s-test/job/kubernetes-e2e-gce-parallel/7960/) timed out again early this morning, but this networking test was not enabled. This suggests that there's still something else triggering the problem.\nI may have a partial fix for this in .\nHave we seen this since merged? Can we close this bug?\n- I think we had a lot of other problems during that time. I would suggest leaving it open during the weekend and close it on Monday if we don't see any problems.\nI don't think we've seen this in a while. Please reopen if we're still having mysterious timeouts in Jenkins.\nWe're seeing this again in . It looks like someone put \"Networking should function for intra-pod communication\" test back in the parallel run recently. I think that we should revert whatever that PR was.\nThat test had been fine. More worrying is that my fix from seems to have vanished.\nOK, looks like that was removed in , referencing issue . I'm not sure if the same issue is happening again (where operations mysteriously hang forever) or if it's something else. any thoughts?\nI think your answer is here: The timeout field, and all references to it in clients were removed, based on the fact that clients should use request.Timeout() instead, but the clients were not updated to do that. I will send a PR shortly.\nIt seems like very few of the e2e tests are using this method, but we aren't seeing issues on most of the tests. I wonder if there's something in particular we're hitting with the networking test?\nCould this be this magic number in ? I guess this means the checks only happen for about a 75 seconds. If one or two of the nodes are spinning up very slowly, this will trip over . Just a thought. Either way I think this needs to be configurable, I'll work on that angle.\nI agree this \"15\" should probably be bigger. However, I'm pretty sure, that this shouldn't cause the test hang forever - it may cause test failures, but hanging forever test shouldn't happen...\nAnyway - I've sent out out for review - we will see if that helps...\nOK - it won't help, because it doesn't work :( - I asked if there is any reasonably easy way to do it...\nAnother instance: Again - the \"Networking should function for intra-pod communication\" is among tests that didn't run...\nassigning this issue to you since you're working on a PR\n- sure, although I'm not sure if this will solve the issue...\nyou can assign back to me if it doesn't :)\nUnfortunately my PR didn't help. This just happened again: - reassigning back to you\nAnother instance:\nAnd again:\nAre folks still seeing this? I'm having a really hard time getting it to occur on-demand.\nOK, I did find another timeout - The list of suspicious tests does not include the networking test this time: Kubectl client Kubectl patch should add annotations for pods in rc [Conformance] Kubectl client Simple pod should support exec through an HTTP proxy Pods should be submitted and removed [Conformance] Secrets should be consumable from pods [Conformance] Will have to keep searching for builds timing out to see if any new pattern emerges.\nLooking more carefully, is another timeout. (I really wish it'd fail on my debug PR...) Suspicious tests: Job should scale a job down Kubectl client Simple pod should support exec through an HTTP proxy I'm gonna start suspecting that Kubectl HTTP proxy test now... are we doing anything in this test that might be blocking?\nhas a very curious failure: only 3 tests ran at all. It seems we are running the cluster-sanity checks in for every Ginkgo node, so there are probably cases like this run where it'll race and fail in mysterious ways (or possibly even deadlock)? We should probably use and (as documented ) to run the sanity checks, cleanup, and log dumping only once, rather than once per Ginkgo node. This may solve the mysterious timeouts.\nBumping down to P1 because this seems pretty rare nowadays (but I'll still address the issues I noted).\nIt seems like this issue is way too general: jobs timing out without a clear reason should have an action of making that reason clear. Is there any action we can take there?\nI think the fixes I mentioned in will address the mysterious timeouts.\nHappened here:\nHoping this one is gone for good, but I'll track and see.\nReopening; saw this again with\nFinal lines of that before being killed: Another run which timed out is kubernetes-pull-build-test-e2e-gce/. Final lines from that: Plus my earlier suspicions in make me think the kubectl proxy test is lacking a necessary timeout somewhere.\nMoving discussion of that test case to .\nWith open, can we close this in favor of ?\nClosing for now. If fixing doesn't fix, I'll keep looking..."}
{"_id":"q-en-kubernetes-70ca5b9393197b9c681e23f75e1621b021d66b70b0941bb143c6bb4d3bdbccb3","text":"I0921 20:36:33. ] Could not connect to D-Bus system bus: %sdial unix /var/run/dbus/systembussocket: connect: no such file or directory"}
{"_id":"q-en-kubernetes-95af4df06321abec34d22b614a4f94b038ee2b47038f19fa65ee5d7b6d8c32e5","text":"From an e2e test failure: The command should retry resource version mismatches like scale and rollingupdate. cc\n/cc fyi.\nFrom perspective , Isn't the failure intentional (i.e. optimistic mutate) ? or... maybe I'm (probably) missing something. I would think its the user's job to retry... ?\nThe desired behavior that we settled on previously is to detect version mismatches at the client and error. User can force replacement using --force. See , and .\nWhat is the mismatch here? If it is just resourceVersion, kubectl should retry. It should only bail on a meaningful conflict. We're working on eliminating flaky tests (and flaky parts of the system). Also, that log message isn't very useful.\nper discussion on , we can't determine the correct behavior when the resource version changes without more information, so best to catch version mismatch at the client and bail, and let the user replace the object with --force if they want. We can also make the error message less verbose.\nI don't see any mention of resourceVersion in . It shouldn't be serialized in the annotation, because the user shouldn't specify it in their config. The resourceVersion-based precondition is intended to guard against concurrent read-modify-write operations. kubectl apply should be able retry the entire diff&patch in the case of this type of conflict. That said, it's not obvious to me from the log message what the actual problem was in this case.\nI'm sorry, my mistake. You're right that resourceVersion is not in the annotation, and that we should retry on resourceVersion failure. I was confusing this with the api version change issue discussed in . This is actually a much simpler issue, in the sense that it's an optimistic lock conflict. If we encounter a lock conflict, then we should fetch a fresh copy of the object, and try again to apply the new configuration supplied by the user. If we don't get a merge conflict when we resubmit, then all's well. If we get a merge conflict because the object has changed out from under the user, then we report it in those terms, and the user can decide what to do (e.g., replace using --force). Does this fit your expectation of how it should be handled?\nThis is causing test flake:\nyes, your understanding is correct. Do you have time to fix this? We're trying to get all our tests fixed. Additionally this sounds like a legit system bug, not just a flaky test problem.\nMaybe. Swamped right now, but can fit it in given enough notice. When is the fix needed?\nasap. We'll find someone else.\nThis should just be a generic GET/Update retry loop, eg:\nGo for it. Jack Greenfield | Google , Prashanth B wrote:\nChao, do you have cycles to take a look?\nI thought Patch is retried in the server. I'll investigate.\nPatch is retried at server-side five times before an error is returned: , so I expect this issue to occur very rarely. In fact, it has passed last ~400 kubernetes-gce-e2e tests. If we want to make this test less flaky, I'd recommend increase the number of the retries at the server-side. What do you guys think?\nIt'll fail if an actual conflicting change was made, though. Do we know what the nature of the conflict was? Maybe first we need to add some log messages to figure out what the conflict was next time it occurs. , Chao Xu wrote:\nThis happened last week:\nyou are right. I thought an actual conflicting change would result in a different error message but it's not. I'll add the logs.\nIn I a log entry if there is a meaningful conflict. That said, I don't think we have a meaningful conflict in this particular e2e test. After taking a closer look at the implementation of , I think the submitted is precise, so we shouldn't retry at the client side if server reports a meaningful conflict. [edit] And if the conflict is not meaningful, we should just rely on the server-side patch handler to retry.\ndoes the milestone on this need to be changed to v1.2?\nHow often is the test flaking?\nLast time I checked, I'll check it again.\nLast time I checked the results was two days ago, in the past 48 hours, the test has consecutively passed another 124 builds ( to ).\nI only see 3 reports of failures since October, so I'm going to say, no, it doesn't need to be fixed in 1.2.\nany progress on this now? I returned back to the kubectl apply issues, and hope I can fix all apply related issues before 1.3.0 release.\nI'll try to check the flake rate again. IMO if the flake rate is non-negligible, we should increase the retry limits of Patch, or more aggressively, just remove the limit.\nInfinite retry still scares me, mostly because we may have failure modes that could race for a very long time. , Chao Xu wrote:\nI don't think anyone has worked on this. Though the original failing test/logs weren't recorded, it appears to be a failure of \"should apply a new configuration to an existing RC\": There is a retry loop in the Patch handler in apiserver, but if resourceVersion is specified, the retry loop will always fail in the case of a conflict. (We should detect that and bail out, but I don't think we do.) In the case of such a conflict, apply should diff again, create another patch, and try again. I agree with that it shouldn't retry forever. We should use the same retry limits we use for other commands.\nwe have an infinite loop in GuaranteedUpdate (), maybe we should apply a retry limit there as well?\nAgree. Just note that the test flake doesn't specify resourceVersion.\nSince this is causing flakey test failures, it is something we should focus on. Is this something I should pick up?\nFYI: has a PR out for this under review\nMaking this a P3 (no release block) since I looks like we haven't seen many instances since Feb 2016"}
{"_id":"q-en-kubernetes-f390f9655b3a3a2bdc3cb72f1f0fbadbba843c0fc19be84f91c4dc99ab97598a","text":"We should be testing the reliability of a long-running cluster. These tests kill/restart components or nodes in the cluster, defeating the purpose of a soak cluster.\nExample tests: NodessResize Services.restarting DaemonRestart NodessResize Reboot Etcd failure\nAgreed. Some of them already seem to be skipped: GCESOAKCONTINUOUSSKIPTESTS=( \"Density.30spods\" \"Elasticsearch\" \"Etcd.SIGKILL\" \"externalsloadsbalancer\" \"identicallysnamedsservices\" \"networkspartition\" \"Services.Typesgoessfrom\" ) Q\nFYI\nIt'd be hard to prevent regression as more tests are . It might be better if we explicitly define a set of DISRUPTIVETESTS, and include them wherever needed, e.g., GCEPARALLELSKIPTESTS and GCESOAKCONTINUOUSSKIPTESTS. That way, people who are adding tests will be aware of the category. WDYT?\nAgreed, that seems like a good idea. Want to send in a PR? Or at least open an issue?"}
{"_id":"q-en-kubernetes-5aa59b92896882cb1f1faca8e724525235f30c0a0b8b5362a85db1fdeef9947c","text":"The desciption of --all-namespaces is listing the requested object(s) across all namespaces. Namespace in current context is ignored even if specified with --namespace.\" But when I used the following command: error: namespace may not be empty when retrieving a resource by name no namespace set on resource pods \"counter\" The error messages are unclear, and the second message isn't labeled as an error even.\nAnd why couldn't use --all-namespaces flag?\nYou're asking for a specific pod name (counter) in all namespaces, which is a nonsensical query. The pod name is not a grep-like filter, it's a specific name. Yeah, the errors could be better. Describe could work, but potentially generates a LOT of information and API calls.\nBut I found the following API is supported now. Get The result is: has the more information. Do we need to support the following kubectl command(using the above API) or just return clear error messages? get pods counter --all-namespaces\nI'd be ok with this being implemented via the field selector.\nThe issue is still here :). What's our expected result? I'd like to contribute this.\n, after check the code, the URL was built by ; do you mean we should update it; or avoid to use , instead, build url () directly.\nAfter checking the code, it's better to add fieldSelector to , is created for fieldSelector. Will submit a PR based on .\ngetting an object with a particular name across all namespaces seems like something we shouldn't encourage... it breaks the assumption of namespace independence.\n, so we'll only update message for this issue? for example:\nDisagree. Need easy way to troubleshouting the problem \"why pods is stuck in Pending status\". It's tiring for every pod to find who created it and in which namespace. In the logs there is only the name of the pod.\nNow, with field selectors we can do To further my understanding, I'd ask:- doesn't itself break independence? One can search by label across all namespaces, it feels funny for by name to be special."}
{"_id":"q-en-kubernetes-6b0ebb305fa6f354e94ea080d3ec8b3f747326739d46689612ca744492c9132e","text":"From gen-swagger-docs should read the local swagger spec rather than fetching it from That way it generates release api-reference when run-gen-swagger-docs is run in release branch and it generates api-reference for HEAD code, when the script is run on HEAD.\nAssigning to as per discussion.\nBlocks"}
{"_id":"q-en-kubernetes-4c43c7fef986a718c460babbe7d4adf5fa347b29ebfc5017571678a77dac017f","text":"The of the persistent-volumes walkthrough contains the volume mount which seems to be outdated for the used nginx conatiner which is by default serving from the folder. As a result, the data wrtitten in in the node will not be returned by the at the end of the walkthought. Proposed change: Update the path\nThanks for bringing up this issue, Would you be up for sending a pull request with the fix?"}
{"_id":"q-en-kubernetes-e74b76d296fd4bae2768426a77e20fe61b63a0ea9deffdfca568f2c34779640e","text":"Per our email discussion while getting 1.1 out the door. It looks like this never happened, though, which means we're not running autoscaling tests against GCE on the 1.1 branch, (we were running it on GKE with the 1.1-features job, but I'm removing that job in ). I'd once again like to (strongly) suggest that we mirror test suites running against release branch to jobs running against master, so that we avoid this problem of accidentally not running tests when we think we are. I'd much rather have a perhaps-unnecessary slow/HPA/whatever suite than continually try to figure out where slow/HPA/whatever tests are running, depending on if I'm looking at release jobs versus jobs running against master. cc\n+1 We should be running the autoscaling tests in slow or continuous."}
{"_id":"q-en-kubernetes-9d53901b20a3ce5248db51f577ad265c56a786f908ebe1fc18ad2aee2b1cef2a","text":"Hi, Right now in the installation script for Ubuntu there are hardcoded SSH_OPTS. In my case I am using a custom SSH port, and, unless I am mistaken, I have no way to customize it, either globally or on a per-host basis. Would it be possible to do that? I can try to submit a patch if necessary. Thanks.\nYes, please. Just send a PR to add in and make it configurable."}
{"_id":"q-en-kubernetes-292213e5846f2bd05d52e0c96b28a8696121992d21f62e361bb9fafd6583182c","text":"cc\nReopening to remember to check if helped\nNo flakes in recent history\nThis just failed on PR Jenkins for me:\nI've been hitting this one repeatedly in https://doc-0l3o4-0hqt8-s-\nAccess Denied for that link\nJenkins has already purged the link. Anyone who sees this - please open issues WITH THE ERROR MESSAGE if this pops up again. A Jenkins link is limited-lifetime.\nthis just re-occurred, here's the log:\nCan you take a look, with prio, please. Of note:\nI just randomly started debugging this because it caused a PR I was reviewing to fail e2es. I think the problem is the test is unaware that kube-proxies can fall behind each other. This is very likely because the \"relist timeout\" imposed on the watch is randomized server side, also one node could be low on cpu etc. Take a single failing case: To break down what happened: Created webserver-f0m07, tested it was up Picked some node (104.197.253.149) and tested reachability Scheduled hostexec pod on some node (e2e-gce-master-3-minion-w6ut, 104.197.244.14) Ran the ss command and checked for nodeport Evident from the following console logs: 09:05:09 Jan 27 09:04:45.111: INFO: Waiting up to 5m0s for pod webserver-f0m07 status to be running 09:05:09 Jan 27 09:04:45.119: INFO: Found pod 'webserver-f0m07' on node 'e2e-gce-master-3-minion-bcr0' 09:05:09 STEP: trying to dial each unique pod 09:05:09 Jan 27 09:04:45.278: INFO: Controller webserver: Got non-empty result from replica 1 [webserver-f0m07]: 1 of 1 required successes so far 09:05:09 STEP: hitting the pod through the service's NodePort 09:05:09 Jan 27 09:04:45.354: INFO: Successfully reached http://104.197.253.149: 09:05:09 STEP: verifying the node port is locked 09:05:09 Jan 27 09:04:45.376: INFO: Waiting up to 5m0s for pod hostexec status to be running 09:05:09 Jan 27 09:04:57.437: INFO: Found pod 'hostexec' on node 'e2e-gce-master-3-minion-w6ut' 09:05:09 Jan 27 09:04:57.441: INFO: Running '/jenkins-master-data/jobs/kubernetes-pull-build-test-e2e-gce/workspace --server=https://104.197.246.4 --kubeconfig=/var/lib/jenkins/jobs/kubernetes-pull-build-test-e2e- exec --namespace=e2e-tests-services-oht70 hostexec -- /bin/sh -c for i in $(seq 1 10); do if ss -ant46 'sport = :' | grep ^LISTEN; then exit 0; fi; sleep 0.1; done; exit 1' 09:05:09 STEP: deleting service nodeportservice-test in namespace e2e-tests-services-oht70 09:05:09 STEP: stopping RC webserver in namespace e2e-tests-services-oht70 09:05:09 [AfterEach] Services node the webserver pod landed on: node the hostexec pod landed on: Comparing kube-proxy logs from these 2 nodes we see that the second kubeproxy (hostexec) only actually processed the OnUpdate event for the node port at 17:05:17, which is actually waay later than the first kubeproxy sees it (17:04:40). To rule out clock skew I check when the kubelet on the first node saw the hostexec, and this matches up to roughly the time the kubelet on the second node saw the webserver pod (~17:45). To figure out where the skew comes from just check time between OnEndpointsUpdate log lines, you'll see the hostexec kubeproxy taking 2s at times (~17:04:29), while the webserver kube-proxy just whizzes along at 0.5s per update. So I think the problems here are: 's no way to know that all kube-proxies have executed the nodeport actions for a service (i.e \"service readiness\") -proxy maybe processing OnUpdates in-efficiently, but that's not the core problem. test should either have a much larger timeout (the hostexec sh command above has seq(1 10)) and/or make sure it lands on the right host it has already verified preconditions for. The \"service readiness\" problem is a little deeper and will probably help ingress too.\nchatted with Prashanth. He is investigating this.\nAnd again today: gs://kubernetes-jenkins/logs/kubernetes-e2e-gce-parallel/ In the last 20 days this has failed ~9 times in that suite, only counting runs where only a few tests failed. 0.8% fail rate.\nthis is now assigned to you - plans for a fix?\nfor review\nPR for fix is in."}
{"_id":"q-en-kubernetes-e921a1329e162a603540fb4dce228e0e7a7b55de85a818dddd7891c563012592","text":"When I try to bring up my cluster (with a custom wrapper script that sources some other environment files,but calling in the end) using the AWS provider, the script fails with following reason: This happens because the correct environment variable set is . Changing the line to use this value fixes the issue."}
{"_id":"q-en-kubernetes-dff5a657be89122e51d4f3eb3f8988656cad8de0e3ba7b736bbec0e7c70e87f4","text":"Have only seen this once, personally.\nThis flaked in Jenkins as well. [job/kubernetes-test-go/5351]\ncan you take a look at this? I think has a lot already :)"}
{"_id":"q-en-kubernetes-87225d8f4bc8e526bc045df03189198434a2c9e524932dfbbdbe1aba72725fa0","text":"there is a test that you marked as skipped that needs to be reenabled or otherwise taken care of (see ): \"Security Context [Skipped]\" Can we instead classify these as ? See for context. Happy to answer any questions if anything doesn't make sense.\nThis is work toward .\nI like the reclassification! I would be happy to relabel the test. I'll wait for to merge though\nhas been merged; if you send a PR, I will review it. Thank you!"}
{"_id":"q-en-kubernetes-afb0861e202f8fce3b6c8d79cb48e19d0d8d5c78740b4366b66f69e1018d5f1b","text":"cc:\nIt was broken by . I'll send a fix in a sec.\nSorry & thanks for fixing."}
{"_id":"q-en-kubernetes-898fb42ff3be6ddf5751636f0c5233242d4bfcd4663d2731ea511851184e64ac","text":"My reading is that the first attempt to create e2e-gce-minion-template looks like it failed with \"Backend error\", but it seems that the template was actually created, so subsequent attempts failed because it already existed.\nFrom kubernetes-e2e-gce/ claims to have deleted: But it apparently wasn't. Looks like we need to improve handling of this: cc There were previous attempts to make instance template more robust, such as in .\nAnd, after this failed, the subsequent test run failed the leaked-resource check.\nHa, an old TODO strikes again. :)\nIt's actually pretty disturbing that claimed when it wasn't. There may be an internal bug here, since the name should be available for use if delete occurred, in theory.\nDo you think someone from your team could take a look at this one?\nIt sounds like the hypothesis is that the delete failed in , but gcloud reported 'deleted'. But the failure message from the first create operation in is one I haven't seen before: reads to me as \"something went wrong, I don't know the state of the world\". It looks like the template was created OK (ish), but we try recreating with the same name which then fails (as expected). Is there evidence that the delete actually failed in ? My inclination would be that we should have a retry loop that looks more like: (That is go-bash; neither go nor bash. Just be grateful there's no salt in there!)\n- can you take a look at this one this morning?\nI was out this morning but I'll see if I can look into it later this afternoon.\nping\nThis just happened again: I'm going too apply P0 label to the PR that is supposed to fix it.\nAh, this is the same as . We created an internal bug for this too, ref ."}
{"_id":"q-en-kubernetes-c681678808babfbe7431923b7db420cefa8f168cf23d7fd835b30d0e6ca097d9","text":"See example log below. I'm pretty sure the SG in question was from an ELB, but we should just auto-retry anyway."}
{"_id":"q-en-kubernetes-44bcfeddc8a0280d868b3901f7f513a7e1565a2dce681aab6546cd53db5f9e78","text":"In 1.0 and 1.1, headless services that specified a mismatched port/targetPort were tolerated (the targetPort was ignored). tightened validation to reject these services. That causes problems when upgrading from currently-working clusters.\n2 candidate to avoid breaking backwards compatibility\noh hell. Yeah. I don't think this is worth a breaking change. We should either roll that back ASAP or change it to nuke any user-provided value and replace it with the value. I think the latter is potentially dangerous and also not worth the risk, so I think we should rollback and just update the docs for targetPort. objections?\nWho has time to do this? It doesn't look like an auto-rollback is doable.\nAlready opened a PR in\nYes, I agree this isn't worth a breaking change, But could we give an warning log that targetPort isn't equal to port and would be ignored?\nThere's no facility in validation for warning. I don't think spamming the server log in response to API requests is that helpful here\nYes, that sounds reasonable. I just think that targetPort was ignored is a little unfriendly to user."}
{"_id":"q-en-kubernetes-ef950703bbc523928d3eaae485e68af79bed5a970f7cb3a77ab67617ae2b27b3","text":"This might be fixed, haven't reproduced on latest upstream source. But in any case, I thought I'd create the issue. bug = i assume any failure in e2es shouldnt be b/c of a panic. ... I noticed a failure on Openshift today, which I think alludes to a possible issue in replica sets.\ncc\ncc\nYup, looks like a bug, and it looks like it's still there. The isn't checked, so can be nil, which will cause the panic on L102. cc\ni think the iterators should be declarative here i.e. ? thats a better fix .... (update : nvm, i think its fine as is, just add a nil check: Because, the declarative won't work for the entire test, since the function for checking pod name responses requires a full )\nCc\nThanks, Jay. Even with this change, the test would have failed, so we should still debug the cause.\nIt seems we don't check err here: if return an error, will be nil.\nYup.... that is fixed in the PR, right?"}
{"_id":"q-en-kubernetes-886b06652982581fd7e25d590373ed434d5bcc6b0395de55126cecbb82a25826","text":"Getting this error... looks like we are quoting : Working on a hotfix...\nLinking to\nFix is"}
{"_id":"q-en-kubernetes-6f0ac5bc0e5382a6ce45dad10850f935dd1f0d3ace15d7812b713d43f420d4c7","text":"From Slack, we got a report that \"my service controller is trying to delete a ELB ( im running k8s in ews ) thats is not there, and for that reason its failing over and over again with the message \" It turns out the user had deleted the ELB manually. We should tolerate this better.\nI think this was actually caused by a failure to tag the cluster, which triggered another bug where we can't pass an empty filters list to the AWS API"}
{"_id":"q-en-kubernetes-f85722742125f7ea9b877706460d59003186444bb7882bd19fd7f4e8c120208f","text":"I assume isn't actually a fix for this, just a way to get faster logs?\nActually is only where the test failed. To be honest, I didn't have time to look into it at all.\nany progress on this?\nhttp://pr- is another failure.\nOnly this one is an actual error. And... the logging logic doesn't make much sense. It logs \"err\", but it only gets there if err was nil. I can't see any other logs besides the junit logs, which don't provide any extra info. Am I missing something?\nThere is a concurrency bug in the test. We signal the scheduler to stop by but we don't synchronize at this point to verify that the scheduler has stopped. Indeed, since blocks on in the absence of any work, we could sit around a very long time, schedule one more thing, and only then get the signal to exit. We need to synchronize here, but once we do that, we may need to prevent the deadlock that occurs while blocks forever waiting for . I don't know that this is the bug, but the bug is there and it fits the symptom.\nThis is similar to the bug I fixed with . The seems susceptible to this by design. Not only does it provide no way to verify that the goroutine has been stopped, it makes it hard to avoid deadlock because the provided may block forever. We should probably re-do this interface to avoid the problem, although in practice I bet it only pops up in tests. I see two choices: invest in fixing this now, or comment out all the test that come after the assumption \"now we've stopped this scheduler\" and fix it later. WDYT?\nWhen you say Are you talking about everything after step 7?\nYes.\nI guess it is fine to comment out everything after step 7 for now, if you don't see any easy alternative. The key is that we test that the right scheduler is scheduling the pod, and while steps 8/9 do that, earlier parts of the test also do it and I think they're adequate. Please file an issue related to the problem you described, so we can fix it eventually."}
{"_id":"q-en-kubernetes-388477b7b71a1ca060f820d6295e607b1410db3be88b2f2825a1396c929dcb98","text":"in there is a generic error variable defined that is used when a failure occurs and no other error is surfaced, it is a catch-all when something goes wrong. Think this should be reworded to something that is more reflective of the true issue and not so confusing for users. This error mostly comes when a builder can not be created and it is typically not true that the volume type is unsupported, but rather some other underlying issue, such as missing secrets or endpoints or etc... propose something like: or\nare you working on this?\n- yes, I would like to submit the PR for this"}
{"_id":"q-en-kubernetes-35fd153d614877189212c9effd62b4185551ad89b0a5afd2d50910e4d13004f7","text":"GCE PD docs so unused blocks can be freed back into the cloud and performance gains realized. Rummaging through the source of the GCE PD and mounter stuff in Kube, it does not appear that that's being done (the old shell script version of safeformatand_mount does, but I don't think that's used). I haven't confirmed that it's the case at runtime for sure, though, although I will be able to shortly.\nI LGTMed the PR and then realized it needed more work. Reverting it and assigning it to ciwang. We need to make sure that we only use the \"discard\" mount option when the fs supports it. Assigning to ciwang for now. If anyone else wants to work on it feel free to grab it.\nAssigning to This fell through the cracks as ciwang was working in other areas. This did not make 1.4 either moving to next-candidate.\nIssues go stale after 30d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. Prevent issues from auto-closing with an comment. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten /remove-lifecycle stale\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-testing, kubernetes/test-infra and/or . /close"}
{"_id":"q-en-kubernetes-58924bc188741050d42aef28bcde689e9e536597bd1a8023545ea7c340476df0","text":"There was a naive attempt to do this in , but _output it hardcoded throughout the code base so we need a more comprehensive solution.\nDo we still need this ?\n/sig release\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. Prevent issues from auto-closing with an comment. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten /remove-lifecycle stale\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-testing, kubernetes/test-infra and/or . /close"}
{"_id":"q-en-kubernetes-0e001d2949b98669c91261c4dff5862e4b2b89e42c08b2ec522e0f2bf5577576","text":"The e2e test is flaky. It often fails with the error in the test suite. Sample failures: Marking it P0 since it frequently blocks the merge queue.\ncc\nIn the flaky test, the pod is deleted with grace-period=30, but there is no guarantee that the pod is going to be deleted from etcd in 30s, thus the flake. This piece of test is copied from the pod e2e test: In that test, the total timeout is actually larger than 30s, because there is some extra logic (< 30s) between the time when the deletion is requested, and when the timer is started. Given that there is no flake reported for test in , I will fix this flake by extending the timeout to 60s.\nThis flake should have the same root cause as .\nETA on fix? This is the most frequent cause of test flakes and submit queue blockage.\nsent out as a temporary fix.\nHmm, happened in: Which looks like we're just not handling the watch cancellation properly in the test.\ndoes the timeout really fix the watch cancellation? (didn't look too deep at the pr)\nI think so. see bprashanth's last comment, the timeout occurs again.\nAs of 3:20pm Apr.8, last failure occurs at 3:13:18pm, Apr. 7, Mike's fix is merged at 15:04:42, Apr. 7."}
{"_id":"q-en-kubernetes-e8ab9ffac8b2165072d2c1fbe420239717b9c3f34b2625dbc28e6435b3ca7b75","text":"When trying to use kubectl from hyperkube I'm met with \"unknown flag\" errors. It looks as if the flags aren't being passed through to kubectl but hyperkube is trying to handle them itself (and subsequently giving an error when you give it flags it doesn't understand). You can reproduce simply by running: And you'll see the error:\nI have a fix up in PR ."}
{"_id":"q-en-kubernetes-c9313de66f3508f8886cd07268b253d5e2cfa321359afb89838b6984a0afe883","text":"The recently merged PR a version cache in dockertools, which makes the test  flaky. The reason is that injected error into fakedockerclient and then expected to see the error from right after that (See ). However, the error was consumed by the goroutine which periodically updates the version cache (See ). The docker daemon version is rarely updated, and we only need version once for each container creating, during which we also do container create, start, inspect #etc. 1) Do we care about the cost of the call so much that we have to introduce the complexity? 2) Kubelet periodically calls , can we rely on that to update the cache? Anyhow, at least, we should disable the cache or have a fake one in the unit test. :) /cc\nI once considered move this cache to runtime level, while thus may need to change runtime interface. What if only cache apiversion?\nWe can only get via , so there will still be the data race. :) One option is to update the version cache in , now kubelet periodically call to check the runtime status. Another better option is to cache the version in layer, now is under our control, we can refactor it with an additional argument indicating whether to use cached value or get directly from the daemon. If we cache the version only for performance, I prefer the second option which is cleaner and more efficient. :)\nOK, will look into this tonight\nThe unit test itself should not start a goroutine to update the cache (unless you're actually testing the goroutine). We should disable it anyway, as we often do in our unit tests. It also sounds like we don't need a goroutine dedicated for this. If we can tolerate the additional overhead from time to time, we can set a expiration time (e.g., 1 minute), and only call the runtime if the entry expires.\nTotally agree. :) Maybe a new function in wrapping the update logic, or wrapping the logic in the .\nexpiration time sounds good in this case\nThis is happening quite frequently - increasing priority to P0."}
{"_id":"q-en-kubernetes-1bf3aa6fad4034e758c14fedc4b5cde3693f0c03678440275414d57febd5fa84","text":"Turns our that when running a container in host network, its that is passed to dockertools.SyncPod() will always contain empty IP, as a result, the environment becomes empty. Maybe we want to use the instead of in this case to ? Since the apiPodStatus has the valid PodIP, which is when the container runs in host network. Or just populate the podIP with hostIP in when the container runs on host network. Ref cc\nWhy not just set the same with before ? Ideally, we want to remove from SyncPod() in the future if possible, it may be better to get it from . WDYT?\nDo you mean \"set the podStatus.IP to apiPodStatus.IP before SyncPod?\" IMO, we should call only once. Why don't we populate set the before calling ?\nWe also use . However, because we don't support host network bandwidth limit, it is fine to leave the IP unset here. However, I think it would be good to have IP always correctly set in .\nI must admit, I have been thinking about this in the background since created this issue, and the discussion hasn't up for me. I finally sat down and spent some time getting my numerous fixes for podIP via downward API back into my brain, and now the problem seems very obvious to me. So, after some thought, I believe that for the docker runtime, the problem is that , which is responsible for generating the which is used to supply the pod IP, doesn't check to see whether the host network is being used and return the correct IP address if so. However, does, which is why the obvious fix seems to be to use the value from this method. I think the right thing to do instead is to change the following call stack to correctly determine the IP when the host network is being used.\nToday doesn't pass the pod spec, so it cannot know if the pod is running on host. This could be solved if we serialize the pod spec into docker labels. BTW, I think we have a similar issue here for rkt.\nAs said, GetPodStatus doesn't have full pod spec today. And I don't think docker manager has knowledge about the host ip now.\nI think we should think about adding the plbuming. I have not had a good experience storingthis information in multiple places :)\nThe code was written so that it'd work in the absence of the pod specs (which we don't always have). I agree with you, but adding the dependency today without some form of checkpointing (e.g., container annotations) of the pod specs may be problematic now. We can revisit this after we store pod specs in the annotations. What do you think? :)\nI'm cool with revisiting but if we do that we should drop this to p2\nReopen the issue to capture\nIs this fixed?\nhi, also interested in knowing if this is fixed.\nI don't think anyone is tacking on this one right now since people are mostly working on the new CRI implementation for different runtimes. cc\nthanks. it also looks like there might be a specific issue with rkt here. Looking at and , it looks like rkt does not handle the case of status.PodIP when using the host network.\nThanks for the heads up. I will take a look to see if we can quickly work around this.\nIssues go stale after 30d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. Prevent issues from auto-closing with an comment. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten /remove-lifecycle stale\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-testing, kubernetes/test-infra and/or . /close"}
{"_id":"q-en-kubernetes-5a056696a8b17624ee6d97139095a97ae226afc8b36ebaf5f287bd25b6fc9d03","text":"I'm not sure where the defaults come from but they sometimes drop all packets causing container IPs to become unreachable. ARP fails. I'm not sure if this is related but I also noticed on this machine that the default pfifo qdisc drops all packets. pfifo is the default qdisc for an htb node so it's used in the unconfigured htb qdisc: This is because the default value for queue size is zero. On other machines I've seen the default value set to 1 packet and 1000 packets. When I set the limit explicitly the container becomes reachable again. The broken configuration is applied by kubelet to implement traffic shapping and there is no option to turn this off. This patch fixes the kubelet on broken machines by disabling traffic shapping: cc\ndoes this happen with --configure-cbr0 and packet shaping too? The kubenet implementation just uses the existing kubernetes shaping code that --configure-cbr0 does.\nIt seems the *fifo qdiscs default their packet limit to the txquelen of the interface (man tc-pfifo). It also seems most virtual interface types (bridge, etc) set a dev-txqueuelen of 0. It further seems that relies on the default packet limit. I could swear this worked before with kubenet, and it must have worked with --configure-cbr0 when the code was originally . I'll check around to figure out what the default limit should be for interfaces that don't specify a tx queue length.\nNote that to test with --configure-cbr0 you'd also need --reconcile-cidr.\nUpstream CNI fix is here: and we'll also work around it in kubenet for now.\nFWIW, the upstream CNI fix got merged for CNI 0.3.0"}
{"_id":"q-en-kubernetes-d2258d12727adc502d6af29b9a737c8c906c7ec8108d6e52316062052f3408e1","text":"At least twice now, we've created a namespace and seen that the corresponding service account default secret token is never created. (The service account itself is, but not the default secret.) Namespace creation + apiserver logs logstash-ingester: namespace creation triggered service account creation but no secret logstash-broker: namespace creation triggered service account creation and secret creation Resulting service accounts Resulting secrets Controller-manager logs for broken case Controller-manager logs for working case Version Info:\nDupe of , fixed in\nActually, not a dupe. This may be a consequence of other refactors to controller logic - possibly we are missing / dropping events out of the queue. I would expect it to get fixed on the sync interval though.\nWe are also seeing this happening from time to time. In our case, it wasn't fixed in the next sync interval but it took actually a little bit more of 24 hours.\ncan we resolve this and remove it from the 1.6 milestone?\nfixed in in 1.4\nWe just hit this exact issue running 1.5.2. Maybe the fix didn't get the root cause or there was a regression since 1.4? This should be reopened. I waited about an hour to see if the sync interval would resolve this but it did not. I fixed by deleting and recreating the namespace. I can share logs, but they are identical to what the OP noted. Version info:\nWorking with this seems like it's not fixed in v1.5.2 -- what should we do here? We don't have permission to re-open the ticket, but filing the same ticket with the same logs (just fresher dates) seems passive-aggressive. :-) What's your preference?\nno problem, just getting back to this. can reopen this issue.\ncan you recreate with the controller manager running at and include controller manager log lines from ?\nFresh repo, offending namespace \"foo96\", with \"--v=5\" since we already have that logging level in place on my dev stack. This took 96 repros of a \"create namespace / monitor for token creation / delete namespace\" loop. Our script's commands, starting just after 16:36:43: Controller manager log lines from The first reference to the offending namespace (foo96) and the one other entry afterwards: All controller manager log lines citing the offending namespace: For comparison, the previous successfully-created namespace: Status as of 20 minutes later:\nAlso, here's the apiserver logs + controller manager logs, for more action/reaction context: Namespace: foo96 (problem) Namespace: foo95 (success)\nthat's helpful, thanks. it looks like while the token controller is processing the creation event for the service account, it does a live lookup against etcd and fails to find the service account. what version of etcd are you running with? are you running HA etcd? do you have quorum read enabled?\nWe run etcd version 2.3.7 (previously saw this on 2.2.x as well). We run etcd HA, with a 3 node cluster. We do not run the apiserver with the --etcd-quorum-read flag enabled.\n(Note: My specific logs above from this morning are on etcd 2.2.5, but yeah we've seen repro on 2.3.7 as well.)\nyeah, the non-quorum HA etcd config is the issue. I can look into updating the token controller to tolerate that, but I'm positive there are similar issues in many controllers that assume live lookups are reliable\nAhh excellent, glad to have an explanation! So -- now that we know what to google for, we found -- where you and clayton addressed what appears to be this same problem for service accounts, and specifically noted that you wanted to avoid recommending quorum reads. Why is the answer different here? Is it because the downstream impact of non-quorum reads is too diverse to reliably retry every relevant action? (Totally fine if that's the answer, just trying to make sure we understand the reason for what appears to be an entirely opposite conclusion to the same type of problem.)\nI still think controllers should be built to tolerate weak reads, and I'm willing to fix the token controller bug here, but three factors make me think this position is a losing battle: quorum read got a lot cheaper with etcd3 there are strong opinions that etcd quorum read should be on () which means that very few people are writing controllers designed to tolerate weak reads all the CI tests I'm aware of run with a single etcd or with quorum reads on\nPerfect, works for us. As for fixing this particular token controller bug, I no longer have skin in that game -- your call entirely. :-)"}
{"_id":"q-en-kubernetes-ae8bb1341a4214e1ef13fd4f4194d54f9693f2beafba7641cbf9e981c2373024","text":"When users try to label/annotate multiple resources, we should tell them to add . Current output: Expected output:\nand too\nInstead of \"--all is false\", say \"multiple resources were provided, but no label selector or --all flag specified\"\nDescription updated based on suggestion\nI think \"resource(s) were provided, but no name, label selector, or --all flag specified\" is better. I will send a pull request for this as my first PR."}
{"_id":"q-en-kubernetes-921194a3ee9035489edb0f46fa5ed9bbb84000a216afd5f16934a80b7eed9035","text":"Currently a Docker bug may cause KubeProxy to crashloop thus making the Node unable to use cluster network, which in turn makes Pods scheduled on this Node unreachable. To mitigate this issue we need to surface the information about problems with KubeProxy, that will allow scheduler to ignore given node when scheduling new Pods. ProblemAPI is going to address this kind of problems in the future, but we need a fix for the 1.3. Either Kubelet or KubeProxy needs to update a NodeStatus with the information that Node networking is down. I suggests that we reuse NodeReady Condition for this, as this would make the rest of ControlPlane work out of the box. We will replace this fix with the ProblemAPI use in the future. Another option, which requires more work, is creating a completely new NodeCondition type to handle this case. This would require making Scheduler aware of this new Condition.\nDoes kubelet monitor kube-proxy health?\nKube-proxy today runs as a static pod (daemonset), kubelet treats it the same as other static pods. Yes, Kubelet reports kube-proxy podstatus back to master, that is all. The detail information on kube-proxy is in crashloop at\ncc/ Brandon, this issue explains my long position against running kubelet, kube-proxy, etc. critical daemons as docker container. I am ok to packaging them into docker image, and running as linux containers, but they shouldn't depend on docker daemon.\nI think we don't want to run them as docker containers as well, we want to run them as rkt containers.\nThe initial proposal I reviewed was docker container, and I did raise my concerns to and coreos developers then. Running it as rkt container should be ok since rkt is daemon-less today, and shouldn't have this chicken egg issue at the first place. But we still have the problem on the nodes without rkt as runtime. At the meeting, I suggested that kube-proxy update NodeStatus with a newly introduced NodeCondition, like what we plan to do with kernel issue at: The reason of not using NodeReady is that kubelet might override that. Another very hacky way is before crashing, Kube-proxy logs the read-only filesystem problem somewhere, like /var/lib/kube-proxy/... Kubelet will pick up the information, then update NodeReady Condition with detail error message prefixing with \"kube-proxy:\" kube-proxy's responsibility to make the information up-to-date.\nCan I have more description of the bug? Is there a related docker issue? Is there a repro? Does this only affect kube-proxy or will it affect other pods on the node? We need to support running critical applications (including system daemons) in pods in order to allow easy deployment and integration with vendors (e.g. storage and network vendors). cc since we were just talking about integrating through addons. What features are we missing that we can't support this right now? Is it just that our default container runtime is buggy and flaky and unreliable? Do we need the NodeProblem API? Do we need taints/tolerations?\n- I don't know about any way to repro it other than running huge clusters. Pretty much every run of a 1000 Node cluster has this problem on at least one Node.\nIf you want to know the detail, see It is a docker issue, and pretty rare.\nCan we regroup on this issue this week? Like tomorrow? , Dawn Chen wrote:\nI will happily make kube-proxy write to or offer up pretty much any API we deem necessary to detect this issue. To me it still feels like a docker health-check suite is needed, and one of the tests is that sysfs gets mounted rw. BUt in the mean time, how about this: make be the API. Any component which feels that it has a node problem can write a JSON file into that directory. For this case, kube-proxy would write a file: Kubelet can turn that into Node status reports. , Tim Hockin wrote:\ncc\nPR of introducing NodeProblemDetector was merged last week, and the pr to enable it by default for Kubernetes cluster was merged over the weekend. So far everything works fine. Even it is an alpha feature, GKE folks agreed to enable it for GKE. To workaround this particular issue before we have proper docker fix, we can easily extend today's KernelMonitor module of NodeProblemDetector to monitor kube-proxy log, and report a new NodeCondition called SysfsReadonly to the issue visible. But we don't have remedy system yet, and pr of converting node problem (events, conditions etc.) to taints is still under the debate / review: 1) Should NodeController respect this new Condition, and simply mark node NotReady? 2) Should repair system pickup this issue, and restart docker?\nDo we still observe the problem with docker 1.11.X? If no, I want to close this; otherwise, we are going to reconfigure NodeProblemDetector to make the issue visible.\ncc/\nAgree, I'd love to not \"fix\" this..\nI think we've seen it the day before yesterday during our tests - can you please confirm?\nI can't recall - we certainly did saw some failed nodes during the test, but I'm not sure we made sure it was a kube-proxy issue.\nBut what I can tell for sure that it's definitely less frequent - we started 2000-node cluster 10+ times during this week and we've seen it at most once (which makes it at least 99.995% reliable).\nNot enough data. GKE is still deploying the old ContainerVM and isn't running Docker 1.11.1 yet, unless I'm mistaken.\naah ok - I didn't know that...\nI'm going to tentatively close this until we show that it's a problem on Docker 1.11.2, now that Docker 1.11.2 is in on both GCE and GKE.\n(Obviously feel free to reopen if I'm wrong.)\nAnother occurrence:\nSo it's still happening, can we teach NPD to detect and surface this as node unschedulable?\nYes, we can. But I found that the error only occurred once with docker 1.11 till now. At least, this should be a known issue, right? XRef Do we want to fix this in 1.3 or just claim known issues in release notes? :)\nDoesn't this render the node unusable (i.e no services will work because kube-proxy is broken)? if so, we should surface regardless of frequency IMO. Whether we fix or paper over it is a different issue.\n+1\nBecause it's code freeze now, I can add a temporary short hack in node problem detector to detect this issue. Should we do that? :)\nRemoved the priority and it back to 1.3 milestone, so that we can discuss at burndown meeting.\nYeah i think codefreeze just means no features\nIt is very rare after we upgrade to docker 1.11.X. To extend KernelMontor in NodeProblemDetector to detect the issue is totally overkill, since we have to allocate more compute resource to NodeProblemDetector to process tons of kube-proxy logs. Here is what I suggested for this issue for 1.3 and 1.4+: Document this as known issue for docker releases: docker 1.10.X (), docker 1.11.X (). Also document it as Kubernetes 1.3 release known issue at is also updated admin guide for NodeProblemDetector, and the suggestion on how to handle this issue will be included there. For 1.4, I think we should change kube-proxy. Instead of logging the error then crashing, kube-proxy should report its problem to NodeProblemDector, which aggregates the node issues and propagate them to upstream layers, so that the admin / remedy system & control-plane can take proper action upon the issues. I am removing this from 1.3 milestone, and unassigning myself. cc/ to find the right owner from cluster team for next release here.\nOn a 2k node cluster, I just saw 4 nodes come up this way on Docker 1.11.2. I'm not saying it's going to stop the release, but it does stop bringup on large clusters pretty frequently. For some reason we went through a \"nice\" patch where it was rare, but today it was just bad.\nWe can have a quick fix in node problem detector to catch this kind of problem by parsing the kube-proxy log. The only problem is that, kube-proxy generates logs relatively fast. If we let node problem detector do that, it has to waste significant cpu even on the good nodes. A better solution is to make node problem detector expose an endpoint and let other system components report problems. However, we are not there yet for 1.3. If we'd still want to solve this for 1.3, we could: 1) Let kube-proxy have some kind of error log, and only print error into it. Then we can easily change node problem detector to parse the error log and report problems. By this, we can at least make these errors visible to users for 1.3. 2) Or let kube-proxy report the node condition itself. /cc\nComment at is still valid. We cannot do much for 1.3 release here, and extending NodeProblemDetector to parse the application logs is wrong move. Also NodeProblemDetector is not running on GKE nodes yet.\nI'm seeing people say they reproduced on 1.11.2, is there a 1.11 docker issue? I know that was closed. If I'm going to get my docker guys to dig, I need a docker bug report...\nCan you please ramp up on this and come up with a way to bring node and kube proxy teams together to resolve? Even if we don't land in 1.3.0, I propose this is at least a P1 for 1.3.1.\nFYI and others In I a hacky way of resolving those issues to our test framework. In the last run of gke-large-cluster run it seemed to work (there was 1 broken kube-proxy, this triggered restarting docker, which in turn fixed the problem). This is obviously not the solution, but: it seems to prove that 'docker restart' solves the problem it should stop the bleeding in our tests\n& I copied & paste what I wrote in another internal thread to give you some background: No, it is not a blocker for 1.3 release: It is not a regression from 1.2 release. It is way less frequently after we upgrade to docker 1.11.2 from docker 1.9.1 NodeProblemDetector is not designed for processing arbitrary logs. It is an overkill solution given such low frequency and relative high resource overhead. By default NodeProblemDetector is not running on gke node due to the extra overhead per node introduced to our customer. Meanwhile Since GKE already has repair system built, it should be very easy to extend GKE repair system to check if kube-proxy is in crashloop. If yes, then kubectl logs to retrieve the last terminated container log to see if the issue is caused by readonly root filesystem. If yes, the repair system can issue a docker restart.\nagreed not a blocker. This is likely only fixable in docker. We should keep it up the priority so we dig into and fix docker. can try next week to get our docker people talking to whoever on your side you want to get this fixed (in 1.11.3 or later probably)\nCan we add a node readiness or liveness probe to solve this? (I remember there is a issue for this, but can't find it now) One of the default behaviours of the probe is to make sure the key node components are running normally, and the behaviour could be extended overtime. Probably adding a node probe controller or node health controller to perform the probing. Or just use the health probe of GCE MIG, if its functionality is enough to handle this. This could be part of the remedy system. Just random thoughts, not for 1.3 and even may not work. :)\nAll are good ideas. But let's move that discussion to a separate issue, and use this one to focus providing the workaround for KubeProxy panic within 1.3 timeframe. Thanks!\nAlso what you suggested probe controller can be easily achieved by GKE repair system which I suggested above.\nAgree. needs quite a lot of work, definitely not for 1.3 :)\nKeeping in v1.3 until we decide on Docker 1.11 for sure.\nThe issue seems to be present on old and new docker alike, sadly.\nI opened a docker issue here Feel free to supply information on that issue.\nSo what's the right kube-proxy fix? , Lantao Liu wrote:\n- didn't we decided that we just want to discover this problem in NodeProblemDetector and then mark this node as not-healthy (whatever that means - adding a new condition or extending NodeReady or ...)? In my opinion if it would be clear to user why this node is not healthy, this should be good enough. FYI in I worked around this in our tests :)\nIt wasn't clear to me if kube-proxy should do something different or not. , Wojciech Tyczynski < :\nyeah this is a docker bug (), unsure if there's a pod level workaround, couldn't make out if it was specific to hostNetwork. the work around is restarting docker. It happened more frequently in 1.9, but still happens on 1.11 (node team empirical data). To surface it to the NPD we need to either send some http event to a NPD server (vaporware) or write to a special hostPath termination log type file (parsing kube-proxy logs form NPD is too expensive).\nThis feels like a hot potato but I am not sure if it still needs to be addressed or not. If it does, let's do so. I apologize for not being able to understand what the current status is. It looks like we have a . Reading comment from 12 days ago she states: While folks may want the node problem detector extended to parse logs to detect it, that is not a good fit for that system and it doesn't run on all nodes anyway. Also GKE can be extended to possibly repair this problem. Questions: Is this an issue that we need to still address for production? Or can we wait for docker fix? If we haven't fixed the kube-proxy panic (which we should) how do ensure we still provide a signal that this is happening?\nI am leaving this on 1.3 until we know the best thing to do.\nIf we can DIRECTLY tell the node problem system about a problem, I have no qualms about doing that. drop a file in a dir, for example. , Michael Rubin wrote:\nNode problem detector is not running on GKE cluster now. :) and should be a better fix for this. If we still want a fix in the node problem detector, letting kubeproxy drop a file could be enough for now.\nIf you can give me a spec of what dir to mount and what sort of file to write, I'll do it. , Lantao Liu wrote:\nFor temporary fix, something like should be enough. The kernel monitor of node problem detector could be ported to parse other logs. The only reason we can't do that for kube- is that the log is too spammy. should be enough for now. Only the kubeproxy panic error is printed directly into : Even though the default  of glog is , I checked the node in and cluster, normally only few lines are at Error level.\nI thought that Dawn said we DO NOT want to parse logs. Parsing logs is, in general, a total hack and very very brittle. we should have a real push API. , Lantao Liu wrote:\nI agree with I also understood thought we should not parse logs. Is that common practise for the node problem detector? It feels fragile to me and to be avoided.\nYeah, I agree parsing log is fragile, that's why I call it a temporary or a quick fix. :) The only good thing is that it doesn't need significant change on both sides, although I don't think it's a good solution, either. :( For a better fix, it will need some more design and work. Let me think about it a little more~ Anyway, I still think and is needed. As a key component running on each node, it's wired that we don't have a component to monitor whether it is ready or not.\nRe making kube-proxy \"not panic\", it currently fails in resizing the hash table for conntrack entries, for which it needs to write to /sys. I'm not sure if there are more locations (we moved /sys/module/br_netfilter loading into kubelet IIUC), or if conntrack itself needs write access to /sys. I think Dawn/Liu are more worried about the NPD not parsing kube-proxy logs, because its memory usage would balloon. If kube-proxy can detect the situtation internally, either by parsing its own logs (which is brittle) or poking someting in /sys for example (like the hash table resizing it's currently failing on), it can drop a token into a hostPath (hoping the docker bug doesn't manifest as ro hostPath). Kube-proxy will still remain dysfunctional, though we might be able to: the issue/mark node unusable a feedback loop that restarts docker We just need to be careful to cleanup the hostPath so it doesn't end up in a restart loop.\nMy understanding of 1.3's state based on a brief chat with : something like ~1 in 1000 nodes will be affected by this issue and if nodes are affected, pods scheduled on those nodes have {broken,little,no} networking. We hope Docker will eventually fix it. But because that's not the case for 1.3, and we don't want to make excessive changes to 1.3, perhaps that means that we should fix it expediently. Summarizing my understanding of previous discussion: expedient thing would be some mechanism to detect ( proposal like kube-proxy detecting and writing to /tmp/kube-proxy-claims-docker-issue-) and $something (kubelet, node problem detector, a bash script in a loop) that checks for existence of that file followed by some form of remediation {kill docker, sudo reboot, .....} to bash the node into working again. Maybe an ugly hack is okay for now and we hope that docker fixes it for the next release so in 1.4 we can rm the code? Or for 1.4 we can do something that is More Elegant And Less Offensive To Engineering Sensibilities?\nBounded-lifetime hacks are fine, but they need to come with giant disclaimers and named-assignees to cleanup the mess after a specific event. , Alex Mohr wrote:\nAgree, re: bounded-lifetime hacks needing care and we shouldn't do them indiscriminately. We and our users have a problem today -- and we've had a problem for the past N months without traction. I'd like the problem solved for 1.3 because I'd like us to have a great product and I don't think the state of broken nodes is there. Aside; we've spent N hours discussing the issue and could likely have worked around the issue in less aggregate people time than we spent discussing? Apologies for the leaked frustration: this issue seems to be stuck in either some form of analysis paralysis or cracks between teams or perfect-is-the-enemy-of-the-good state. Again, I don't care about implementation details. We also need more than just tech or a piece of code. Exit criteria is that users don't get broken nodes shipped to them.\nWe're not going to solve this problem for as long as we depend on a lagging docker release cycle, and they don't backport or do intermediate releases to help us. This is just a one off docker bug that's left us in a bad state. Such issues come up every release, get documented in the release notes and we ship anyway. The easiest fix (disclaimers and all) is to redirect kube-proxy stderr and parse it out from NPD. Even then, the node will remain broken unless we restart docker. the NDP will only surface the error.\nI'm working on a PR yesterday night, will send it out soon.\nFYI, a fix is here . In , I let kube-proxy update a node condition with specific reason, message and hit to the administrator about the remediation. The specific reason why we want to do this is discussed in the PR description\nFWIW I agree 100% with using node condition SGTM, but your PR should probably also modify the scheduler here to prevent the system from sending new pods to a node that has RuntimeUnhealthy condition. I realize there will still be a time window when we might send some pods to that node (before the problem is detected) but at least the time window will be bounded this way.\nYeah, we can do that. But before that, we should make sure that the node is really unusable without setting conntrack. In fact, we only setting conntrack to increase the max connection limit on the node (default 64k-256k) Without this, is the node really considered to be unusable? If so, we should prevent scheduling pods to the node. If not, maybe we should just surface the problem to the administrator and let the node keep working. I'll run e2e test without conntrack set to see whether there is any problem. Yeah, that's why I think we should not restart docker ourselves to try to remedy the problem. There may still be workloads running the node. We'd better leave this to the user to decide. :)\nis merged, and hopefully, it could solve this issue. I've also sent a PR to revert the walkaround in the test framework to verify whether the issue is really fully solved. Feel free to reopen this if the issue or relative issue happens again.\nthanks for the help!\nI just ran into this, and I am able to repro consistently. It might be due to the craziness of my experiment, but I still wanted to contribute the data point. I am experimenting with docker inside docker. All I am about to describe is itself running inside a privileged CentOS 7 container that has systemd and Docker installed. The CentOS 7 container is running on my dev machine (Docker for Mac). I have a kubelet running fine, and all control plane components running as static pods. When I attempt to run kube-proxy as a static pod (privileged) it crashes with the following logs: sysfs is mounted as rw: sysfs is mounted as rw on containers: Using busybox to further inspect:\nI'm seeing the same issue when running kubeadm-dind-cluster, where kube-proxy is failing to come up, because it is trying to write to /sys/module/nf_conntrack/parameters/hashsize. This is on the filesystem sysfs, which \"mount\" shows is rw (so it passes the R/W check in kube-proxy code), but the filesystem is not writeable. Within kube-proxy: I'm also running docker containers inside of docker containers. FYI, others have not seen this issue, when just running docker on bare-metal."}
{"_id":"q-en-kubernetes-f9cc6996cc57c60be4f48c68936328b146c9dc2bbf1bb7013e816c2703702419","text":"Slow conformance test has caused issues like We should make runtime/image conformance test run faster without breaking the test coverage, possibly by: Shorten consistently check interval . Remove redundant test cases such as ... /cc"}
{"_id":"q-en-kubernetes-cdbdab84d851624095d1ccede1b49e262098d3b6f4a3201986bfd86cae33cda9","text":"Ref and should sort revisions correctly (10 should be after 9, not 1).\nI can help with this bug. Looks like trivial fix in PrintRolloutHistory\nThanks\nare you still taking care of this one ? I am searching for some easy to do PR to just get familiar with the process and put my first PR in.\nI created pr some time ago to close this issue\ngreat thanks :-)\nyou can look for help-wanted issues or check and\nthanks"}
{"_id":"q-en-kubernetes-87a1a6cd7255a7d3af9b5badeea221b6e0f9a3874b0c406972669f22c6eff182","text":"vSphere volume spec fails with the following error, as its not been in api/validation. \"spec.volumes[0]: Required value: must specify a volume type\" Fixed in PR"}
{"_id":"q-en-kubernetes-67015524d03b3c3a046beb56e6abbbe143f04ab34a9cc2bb70bdff31e035c48c","text":"PersistentVolume integration tests fail sometimes: There are some interesting reports in the log like: Thans for report.\nIt turns out that volume integration tests have no logs at all, was not checked and I don't really know what's going on there. adds error checks to help me find out what's wrong there. Still, I suspect there are some troubles with etcd, not related to the test itself.\nThis flake affects also TestPersistentVolumeMultiPVs:\nExtended logging got merged just now, please post links to logs when it breaks again!\nMarking as P1 as per\nI have no real usable logs, but I think it could be fixed by , it hasn't flaked since it was merged. And I have (in merge queue) to further stabilize the tests.\nNot blocking this on 1.3.\nIt hasn't flaked for almost a month, can we close this?\nClosing this issue since previous fixes made the flake go away."}
{"_id":"q-en-kubernetes-74ce4fea7378dc634ccb83a22893b6255fa243ed3e0b4f0e57320154ecc66caf","text":"removed this test. We need to continue testing this. This wasn't about testing upgrades, the prefix has nothing to do with upgrades. It's just there so people can share etcd. We maybe don't need to double everything but we do need to make sure that keeps working.\nTo be honest - I think that what we was testing before didn't give us completely any information. We didn't testing what information we are storing in etcd and where. That said, if we had a hardcoded prefix somewhere in our code, we wouldn't discover it at all, and those tests would still be passing. We should have a test that is directly checking etcd contents. I think it should be possible to write a simple unit test for it. Will try to put something together next week."}
{"_id":"q-en-kubernetes-94f9f800309f4cdedddf2efc2c38e4c633d471e79f09976ab3145ec32dee399a","text":"If the first scheduling failed (whatever the reason is), the Error() func is called. However, the default Error function is wrong and in case of some errors may drop the pod and never requeue it for scheduling again. The problem is that if the Get() function: return an error (for me it was \"dial tcp 127.0.0.1:8080: connect: cannot assign requested address\", but it doesn't really matter), then we simply forget about a pod and will never schedule it again. We should retry those requests (with some backoff)."}
{"_id":"q-en-kubernetes-5922ad1ee5081362e41d63aecf4f82c8ff968e4c864a1f671bcdef7e00456963","text":"For example\nThis is with recent kubectl:\nI think this is a duplicate, but just in case.\nI will double check this.\nThanks - it isn't a big deal, because I obviously didn't intend to rolling update to an empty manifest :-) (I was experimenting with templating my manifests and made an error) But let me know if you need any more info to reproduce!"}
{"_id":"q-en-kubernetes-14cd86c3860004d87bc9d8599ab7960ef295aad162f6de5c31fc464723da6cf3","text":"Observed weird memory spikes for both the dns healthz and kube-dns pods in production. Kube-dns didn't oom because there's enough headroom, but healthz did. I can't explain the spikes yet. I tried a bunch of different repros: no , bad nslookup command, sharing the node with cpu hogs etc, but was unable to trigger it. The memory grows, but stays under 10Mi. As a short term mitigation, we stole some ram from kube-dns, because it was always comfortably within limits. Related b/\ncan be closed -- healthz has been removed\n-- close issue?"}
{"_id":"q-en-kubernetes-89a6771191a9d4789a41e22c8438d008748f4255d70087ba009452760b4511ad","text":"https://k8s- Failed: [] SchedulerPredicates [Serial] validates that InterPodAntiAffinity is respected if matching 2 {Kubernetes e2e suite}\nhttps://k8s- Failed: [] SchedulerPredicates [Serial] validates that InterPodAntiAffinity is respected if matching 2 {Kubernetes e2e suite}\nhttps://k8s- Failed: [] SchedulerPredicates [Serial] validates that InterPodAntiAffinity is respected if matching 2 {Kubernetes e2e suite}\nhttps://k8s- Failed: [] SchedulerPredicates [Serial] validates that InterPodAntiAffinity is respected if matching 2 {Kubernetes e2e suite}\nhttps://k8s- Failed: [] SchedulerPredicates [Serial] validates that InterPodAntiAffinity is respected if matching 2 {Kubernetes e2e suite}\nhttps://k8s- Failed: [] SchedulerPredicates [Serial] validates that InterPodAntiAffinity is respected if matching 2 {Kubernetes e2e suite}\nhttps://k8s- Failed: [] SchedulerPredicates [Serial] validates that InterPodAntiAffinity is respected if matching 2 {Kubernetes e2e suite}\nhttps://k8s- Failed: [] SchedulerPredicates [Serial] validates that InterPodAntiAffinity is respected if matching 2 {Kubernetes e2e suite}\nhttps://k8s- Failed: [] SchedulerPredicates [Serial] validates that InterPodAntiAffinity is respected if matching 2 {Kubernetes e2e suite}\nhttps://k8s- Failed: [] SchedulerPredicates [Serial] validates that InterPodAntiAffinity is respected if matching 2 {Kubernetes e2e suite}\nhttps://k8s- Failed: [] SchedulerPredicates [Serial] validates that InterPodAntiAffinity is respected if matching 2 {Kubernetes e2e suite}\nLooking into this...\ntrying to track this flake down. It looks like the pod is launched despite AntiAffinity in . In my local setup I saw the same, but noticed that the new admission controller was not activate. I have activate it and now get this:\nhttps://k8s- Failed: [] SchedulerPredicates [Serial] validates that InterPodAntiAffinity is respected if matching 2 {Kubernetes e2e suite}\nhttps://k8s- Failed: [] SchedulerPredicates [Serial] validates that InterPodAntiAffinity is respected if matching 2 {Kubernetes e2e suite}\nShould be fixed by"}
{"_id":"q-en-kubernetes-75c5d65065e1971e68f3d623cfc156efea5ba70b11e44a4a727242bc91b29a35","text":"https://k8s- Failed: [] Services should preserve source pod IP for traffic thru service cluster IP {Kubernetes e2e suite} Happened on a presubmit run in .\nSorry, it is a new e2e test I am still working on. It is not merged yet.\nDon't worry about it, our flake detection isn't quite perfect yet :)\n[FLAKE-PING] This flaky-test issue would love to have more attention...\n[FLAKE-PING] This flaky-test issue would love to have more attention.\nThis new e2e test has been merged. Let's observe it few more days and close this issue if nothing bad happen.\n[FLAKE-PING] This flaky-test issue would love to have more attention.\nThis does't look like a failure in new test, rather its a generic failure with serviceaccounts.\nThanks for the investigation. Yeah I think it is the serviceaccounts issue. I wonder when could we close this flaky issue. Seems like it is stable if serviceaccounts issue is fixed.\nYou can close it with a reference to the serviceaccounts issue - do you know the # for that one ?\n[FLAKE-PING] This flaky-test issue would love to have more attention.\nhttps://k8s- Failed: [] Services should preserve source pod IP for traffic thru service cluster IP {Kubernetes e2e suite}\nhttps://k8s- Failed: [] Services should preserve source pod IP for traffic thru service cluster IP {Kubernetes e2e suite}\nhttps://k8s- Failed: [] Services should preserve source pod IP for traffic thru service cluster IP {Kubernetes e2e suite}\nhttps://k8s- Failed: [] Services should preserve source pod IP for traffic thru service cluster IP {Kubernetes e2e suite}\n[FLAKE-PING] This flaky-test issue would love to have more attention.\nHave a fix for this, I'm now keeping running tests on my own cluster. Will send out PR once verify the fix works."}
{"_id":"q-en-kubernetes-bd12a4d0d123e25009d08863d65df818502d45139444b90d3cc95fd834eb7a2c","text":"On a cluster using GCI, I'm not able to create docker containers using the docker CLI. On further introspection, it looks like bridge is being deleted. I consider this a regression. K8s used to allow creating docker containers out of band and in-fact it is a feature that is taken into account as part of k8s node features. To further exacerbate this issue, exists on containervm. We need to match the behavior between GCI and container-vm and ideally stop deleting docker0 bridge. cc\nDuring GCI start up, got deleted to avoid overlapping cidr with . ref:\nthere's no need to delete docker0 afaik. It was deleted on gci master to fix the \"overlapping cidr\" problem, i.e we were supplying to kubelet on GKE master, and this resulted in 2 bridges with the same cidr. At the time I'd suggested passing the correct cidr. Don't remember where this discussion ended. Note that containers on docker0 with/without cni are broken anyway, we used to start docker with --bridge=cbr0, so would result in an ip from cbr0. This is no longer the case. We use kubenet to put containers on cbr0, so will result in an ip from docker0, which is not connected to nanything else.\nThis shouldn't cause docker crashes? you should still be able to create containers, just no networking. And networking is broken through docker CLI anyway because of the cni issue mentioned, or are you observing something different?\nwill return error. Have to add\nSince bridge has been deleted, docker run by default will fail unless one overrides the network ns to be host or none.\nso what we're really losing is the ability to docker run then talk to others (external internet, other pods) without net=host on GCI. This is useful and we should fix it, as mentioned earlier we probably don't need to delete docker0. docker run + others talking to us is broken because of the cni change, and will be broken even with docker0 (IMO this is fine).\ncc\nYeah. I'd prefer ignoring docker0 as long as it doesn't affect kubelet functionality. , Prashanth B wrote:\nVish pointed out that of docker dies, docker0 comes back. Given that we should be consistent. Either force docker not to create a bridge at all, EVER, or just ignore docker0. We may even want to consider assigning --bip= 169.254.123.0/24 or something, so we can reliably use 172.* for clusters eventually. , Vish Kannan wrote:"}
{"_id":"q-en-kubernetes-ce9585d659f1aa3e82e33b4391083ff4b7d6946234ff3a25101fb4ae4755e247","text":"When I tried to use the following command to build kubernetes: I encountered the following errors: I'm using Ubuntu 14.04 and the code from master branch. will submit a patch then to fix this problem.\nwhoa, I just landed in the exact same place, I am trying to build k8s on a CentOS7 box. Tried with removing the unused utils and works like a charm !"}
{"_id":"q-en-kubernetes-3b5c5f4907e997aaf8daa53f4c0daf41c3d1e2651d7a7af8d3943232c2d64d40","text":"liveness probe was introduced in Heapster seems to work correctly so we should either fix it or revert liveness probe.\ncc\nport seems to be wrong"}
{"_id":"q-en-kubernetes-6d7f7bc8962ffccb3a122ebb5d170bbec0359b559f947a84eb028587481ba4c8","text":"introduced a new requirement for node names to be DNS1123Labels. In case of hack/local-up-, an IP address is used as node name and thus endpoints aren't created for services, e.g. kube-dns doesn't work anymore. There are other projects that use IP addresses as node names for running simplified k8s dev environment: Probably there are other projects where it's convenient to have just an IP address as a node name. Verified against . On affected installs, service latency e2e test hangs, see e.g. Example ( some newlines):"}
{"_id":"q-en-kubernetes-d0585853c1eef197c56eb9e796f54145c38efffe62a23c566467505aa4e95a99","text":"During a manual upgrade test, described in , a kubelet failed to communicate via SSL to its master. This is currently blocking the 1.4 release. There is not a repro for this yet. The cluster was running 1.3, and was upgraded to 1.4 successfully, but upon downgrading to 1.3, the kubelets one one or more machines are not able to talk to the master. Current kubelet version is . (EDITED) The error when talking to the master is (lots like this): I think this means that the kubelet cannot verify the master's cert. I can't find the master's cert. Since there is not flag to kubelet to tell it where to find its config, I assume it is using the default location, which is . It seems wrong that clusters[0].cluster.certificate-authority-data is empty. That should have the cert for the master, right? I see this thing: but I don't understand what process generates , and if, presumably, from that template, why it isn't setting the env var CA_CERT.\nin case this is related to kubelet bootstrap changes in 1.4.\nCurrently node is :\nCorrection. The node is running 1.4.0-beta.5\nIf the kubelet hasn't set the then it will not go through the bootstrap code path. Is the kubeconfig provided through the bootstrap flag or the kubeconfig flag?\nNot set. For reference, args are:\nQuestion: Could you please tell me why your nodes are running with GCI image at the first place? The test is starting with 1.3 release, which is using old debian-based containervm image, and I thought doesn't change the image at all (separate topic).\nIt does change to gci on gce with the default upgrade scripts, i brought this up on an email thread. I'm guessing something was dropped when we translated salt grains (containervm) to gci env vars.\nWhat the default upgrade scripts you were referring to? I just double-checked with test cluster which was upgrade to 1.4 through , the node is stay with debian-based containervm, but Kubernetes version is bumped up to 1.4-beta.X\nOk, I am assigning this to to confirm. If upgrade logic from 1.3 to 1.4 changes the underneath os, that is wrong. I strongly believe this is just a simple operational error.\nMy understanding is that ignores base images altogether. , Dawn Chen wrote:\nHmm\nI've tried and the node image did get updated to GCI. Before upgrade to v1.4: After upgrade to v1.4: After downgrade to v1.3: /cc\nhelped me debug this. He found that the metadata for the instance did not include the CA_CERT in it. That seems wrong.\nRenamed from \"x509 failure for kubelet after upgrade test\"\nYeah, that's the error you'd get if the proper ca isn't loaded.\nWill we get this problem if we don't upgrade to GCI (i.e. we stay with CVM) in -N? I guess we can try it and find out.\nManual upgrades have been successful if we don't switch image types. , David Oppenheimer < :\nre: If you don't change the underneath image during the upgrade, the issue shouldn't occur."}
{"_id":"q-en-kubernetes-3e62f6c1056f3c575902ec60b40c73397e20805db4e60e6464c072402ee00112","text":"./hack/e2e- line 1192: local: `replica-pd=jenkins-us-central1-b-master-pd': not a valid identifier FYI - this looks related to your recent work? I don't see any obvious PR to blame, but I seem to remember you guys touching this part of the code recently? BTW, I deleted a bunch of leaked resources in the relevant test project on Friday (9/16), assuming that the test scripts would re-create whatever they needed. It's possible (but unlikely) that this has anything to do with that.\nHere are some example logs to aid debugging: https://k8s-\nI believe this to be the source of dysfunction: Which is turn caused by what you found, which causes the disk to not be cleaned up ahead of time: You can try this out in your local shell: git blames leads us to this recent commit: cc the author The root problem is unfortunatley very annoying: this is how bash tells you that variable names (in this case, $replica-pd) cannot contain dashes!\nfix is up in\nAwesome many thanks. I'll make sure it gets merged ASAP."}
{"_id":"q-en-kubernetes-97f44e21cc78f43605c4bf074e1d4149cbd6d9ee084c3fb8b7a55d290addbb40","text":"After PR , I did a bit experiments on kube-dns graceful termination and got some unintended result as below: I guess it is because the other container was terminated before . DNS query from clients failed because is the actual connection point. After a bit discussion with and two solutions for this are: another program wrap up the dnsmasq and capture the SIGTERM signal. out a way to modify the sig mask of process without touching the original codes.\nSomething I have been meaning to try: Put in Dockerfile Build and push image We should deliver SIGCONT rather than SIGTERM. I did a quick test and it seems to work. Otherwise trapping SIGTERM in the shell and ignoring it is well documented... , Zihong Zheng wrote:\nThanks, I will take a look and test these."}
{"_id":"q-en-kubernetes-e62d168d1db3d76f5f6327d88acb5a7be21698d985e9cda01df5be46fc45acd4","text":"Kubernetes version (use ): 1.3.7 Environment: Cloud provider or hardware configuration: Bare-metal Ubuntu cluster OS (e.g. from /etc/os-release): Ubuntu 14.04.3 LTS Kernel (e.g. ): 3.19.0-47-generic Install tools: KUBERNETESPROVIDER=ubuntu ./kube- Others: Docker version 1.12.1 What happened: appends the same copy of each time we start and stop Kubernetes cluster (using Ubuntu's & ). For example, if we start Kubernetes cluster the first time and restart it again, the content of is doubled: This cause the following error when starting Kubernetes as seen in : What you expected to happen: content should not be concatenated every time Kubernetes is restarted. How to reproduce it (as minimally and precisely as possible): Start Kubernetes, stop it, and start it again, using and . Anything else do we need to know: One possible fix is to remove the in the following line in :\nThis is a bug and easy to fix. I'll open a PR to fix it\nBug has been fixed."}
{"_id":"q-en-kubernetes-276209858b32894cc42d3f339c3506f4463f5a2951f4a4759747232d12c7a7b6","text":"https://k8s- Failed: [] Load capacity [Feature:Performance] should be able to handle 30 pods per node {Kubernetes e2e suite} Previous issues for this test:\nhttps://k8s- Failed: [] Load capacity [Feature:Performance] should be able to handle 30 pods per node {Kubernetes e2e suite}\nWe understand the problem here - I'm working on the fix.\nhttps://k8s- Failed: [] Load capacity [Feature:Performance] should be able to handle 30 pods per node {Kubernetes e2e suite}"}
{"_id":"q-en-kubernetes-ff687515ab25c4bdd4af394a75c676d411e02b095a7423ff0f91396e6bca07a8","text":"What keywords did you search in Kubernetes issues before filing this one? (If you have found any duplicates, you should instead reply there.): vSphere, \"message\": \"no endpoints available for service \"kube-dns\"\", kube-dns, etc -- similar issues include: however it's not quite the same Is this a BUG REPORT or FEATURE REQUEST? (choose one): Bug Kubernetes version (use ): Environment: Cloud provider or hardware configuration: vSphere OS (e.g. from /etc/os-release): Using provided vmdk Kernel (e.g. ): and Install tools: GOVC + the script What happened: Our environment utilizes the RFC1918 10.0.0.0/8 block, so I edited to utilize an address space we are not using. This is my current configuration: After the deployment is finished, I am able to login to the dashboard via the URL provided via , however, when I check the DNS server I see: When I list services, I see: When I start a container and try to ping kube-dns, I am unable to: I can ping external IP addresses: Here's the configuration on each minion and the master: What you expected to happen: I expect containers can reach ping eachother. How to reproduce it (as minimally and precisely as possible): I've reproduced this problem using both 172.16.0.0/16 addresses as well as the 192.168.0.0/16 addresses. Anything else do we need to know: I followed the documentation here: and the only change I made was to the network configuration.\nI just noticed that doesn't include any routes for the service IPs I defined (e.g. has no routing table entries on any of the minions, only ). I'm not sure how to fix it, but it looks like the DNS server can make requests to the upstream DNS, it just can't get responses back.\ncc\nSome additional info I probably should have included the first time:\nI just noticed that while from inside a container I can ping google: I cannot ping our internal DNS which is at IP . So I think I'm getting closer to where the problem actually is. I also think the 192.168.120.120 was not actually unreachable, just not pingable. For example: Returns right away, while doing the same to IP hangs. I'm not sure why when using our DNS in the of my master/minions doesn't seem to give them the ability to resolve internal assets on our network.\nCan you provide your output for iptables also\nHi imkin, Sure thing! Minion 1 - 10.248.30.38 Minion 2 - 10.248.30.40 Minion 3 - 10.248.30.37 Minion 4 - 10.248.30.39 Master - 10.248.30.36 Let me know what else I can provide, I haven't had much time to continue troubleshooting this week, but getting this to work would be a really great template for building our bare metal deployment. Thanks! ZC\nCan you try with DNSSERVERIP=\"192.168.128.120\"\nHi - I will edit that in the initial configuration and re-up my cluster this morning to see if that helps things work out of the box so to speak. Thanks! ZC\nHey I have edited my configuration as follows: I have re-run the script and everything came up and validated. I am still experiencing the same behavior, however. I now see this in the logs for the DNS service via the dashboard: I see that the service can't be reached by kubectl either (it doesn't even return kube-dns like it did previously): I also notice that on the minions, none of them seem to have routes for the higher IP ranges:\nThis looks to be the same as posting a PR shortly.\nAwesome!! Thanks - I'm grabbing the from your pull request and will re-open this issue if that did not resolve it (once it's closed by the merge).\nIn the second case [when your kube-dns had the ip 192.168.128.120] where was your kube-dns?\nbased on the logs you posted you were definitely impacted by the issue in but I am not sure if that is the only issue on going on your setup. Please let us know how the fix goes for your environment.\nThere is no route for 192.168.3.0/24"}
{"_id":"q-en-kubernetes-cf9984ddc8d3414505d43a827d639ed0a0feb51bbf3cc5f704b07594a951615a","text":"Is this a BUG REPORT or FEATURE REQUEST? (choose one): BUG REPORT Kubernetes version (use ): Environment: Cloud provider or hardware configuration: AWS OS (e.g. from /etc/os-release): Debian GNU/Linux 8 (jessie) Kernel (e.g. ): 3.16.36-1+deb8u1 (2016-09-03) x86_64 GNU/Linux What happened: Deployed a scheduled job from this manifest: After some time the jobs are no longer executed: What you expected to happen: Jobs are executed as defined in the manifest. How to reproduce it (as minimally and precisely as possible): Deploy a scheduled job and wait for some time.\nSome logs from the controller manager:\nPing. More than 2 weeks and no action at all? Anything else i can do?\nI'm also encountering this using both and as concurrency policies, fwiw.\nI'm hitting this as well. Kube\nDo the jobs start to reschedule if you delete old one? Try running . cc\nFrom kube-controller-manager logs: The job is scheduled for 1:00. I think the \"Multiple unmet start times\" message is a clue of where things start going downhill. Also, notice how at 1:00 the controller wants to create job (for a whole day, it had failed on ).\nDeleting (all) old jobs / pods fixes the problem temporarily. After some time the scheduled jobs face the same issue again.\nThere's a fix () that I've submitted against master to fix the Replace strategy, I'll mark it for inclusion in 1.4.\nThe other problem that I see here is this: It is possible that our is creating duplicates. I'll try to investigate that a bit more.\nsorry for the delay in response time, I must have missed that while going through issues. Next time you're experiencing problems/have a question wrt Jobs/ScheduledJobs, just ping me directly in the issue.\nNo problem and I will. Thanks!\nI'm still hitting this on\nFix in flight , I'll make sure to cherry-pick it back to 1.4\ndoes this ship in to latest kubernetes release?\nyes, this should be present in 1.5, 1.4 cherry-pick hasn't been merged yet.\nso this issue is still present in v1.4.6?\nyes - it is still present there. My fix was merged to 1.4 yesterday. You should wait for the next 1.4.x release.\nthanks!\nwhich according to should happen Dec 8th.\nany news on the release?\nSee !msg/kubernetes-dev/o0At2-4KKQ8/_kPVM1XyDAAJ\nso v1.4.7 will be released on that same day?\nLook like should be available later today: !msg/kubernetes-dev/00lJpT9pxDc/YZfDzCxMDQAJ"}
{"_id":"q-en-kubernetes-379dabb0d3852f2c76fc6671dc8637dfc359bc3af1d634613a62926d11f19208","text":"It fails with: cc\nLooking at the generated I see that FEDERATIONIMAGETAG is \"\" which is resulting in this error. I see that before , FEDERATIONIMAGETAG was being set to but that docker tag file does not exist for hyperkube. I see the following exist in my : but does not exist there. Trying to figure out why.\ncc\nmake sure you are running the scripts with env var exported. That is what controls whether gets created.\nWith , we are not generating the federation-apiserver binary and instead using the hyperkube binary.\nsorry, i copied the wrong line. The droid you're looking for: File conditionally written during process File consumed when templating federation control plane yaml\nRight. My is also empty. I am trying to run federation e2e tests. The command I am running is: which is what I have always used. I guess something changed in\nthat should definitely work. Please and then run the step, capture the output and email to me? I was literally doing exactly this process for another PR today, and I verified the image tag file was getting written correctly as well.\nI figured out this is happening because after make clean or a fresh clone, kubectl will not be available when we are trying to get the semanticimagetag_version() in here.\nAwesome! Thanks"}
{"_id":"q-en-kubernetes-ea165427d7925a52e0a3175d5df532a0d4e87bf9f1606c149ffe6c4d4639a9f0","text":"Implement unit (eventually, integration) tests for . This is an effort for Q4 2016 to be lead by and Refs\nHas been implemented very much thanks to ;)"}
{"_id":"q-en-kubernetes-530f1b156a6d906607bacfbd0f2d31ee7d1e4a6539ca9af4be5b6cbffe99dc57","text":"Failed: https://k8s- Run so broken it didn't make JUnit output!\nFailed: https://k8s- Run so broken it didn't make JUnit output!\nFailed: https://k8s- Run so broken it didn't make JUnit output!\nFailed: https://k8s- Run so broken it didn't make JUnit output!\nFailed: https://k8s- Run so broken it didn't make JUnit output!\nFailed: https://k8s- Run so broken it didn't make JUnit output!\nFailed: https://k8s- Run so broken it didn't make JUnit output!\nFailed: https://k8s- Run so broken it didn't make JUnit output!\n, it seems we forgot to add to ; I'll create a PR for that.\nFailed: https://k8s- Run so broken it didn't make JUnit output!\nFailed: https://k8s- Run so broken it didn't make JUnit output!\nFailed: https://k8s- Run so broken it didn't make JUnit output!"}
{"_id":"q-en-kubernetes-036fbe154616710f38e97e5f728b943492772b1362cb5fd735c4695cc31f0484","text":"Similar to\nis a dup of this. PR is just a quick try. Will spend more time on it to cover all cases.\nstatus?\nI sent PR which is waiting for reviewing from . Also, that PR is blocked by issue which is self-assigned by you but not fixed yet :)\nThat issue might be not hard to fix, but I didn't find how.So I need your help. IMO, we mignt need to change CI jobs configure in some files or images.\nAfter discussing with reviewer ( this might be better suited to merge but not for 1.7.0, as it touches a part of the code that is quite old and stable. In other words, we probably need to give this change some time to soak in the e2e system as opposed to put right away into the release."}
{"_id":"q-en-kubernetes-6f0afc6f5bb87e4e6a0796419b897e03b743cf6c06e15bb5c0128536770fe820","text":"<!-- Thanks for filing an issue! Before hitting the button, please answer these questions.--Is this a request for help? (If yes, you should use our troubleshooting guide and community support channels, see ): What keywords did you search in Kubernetes issues before filing this one? (If you have found any duplicates, you should instead reply there.): Is this a BUG REPORT or FEATURE REQUEST? (choose one): FEATURE REQUEST <!-- If this is a BUG REPORT, please: Fill in as much of the template below as you can. If you leave out information, we can't help you as well. If this is a FEATURE REQUEST, please: Describe in detail the feature/behavior/change you'd like to see. In both cases, be ready for followup questions, and please respond in a timely manner. If we can't reproduce a bug or think a feature already exists, we might close your issue. If we're wrong, PLEASE feel free to reopen it and explain why. --Kubernetes version (use ): all :smile: Environment: Cloud provider or hardware configuration: GCE OS (e.g. from /etc/os-release): all Kernel (e.g. ): all Install tools: Others: What happened: uses the cli a lot and does a (in the commonly used function ) before doing things like . updates can introduce breaking changes (for example in the recent 132.0.0 release it removed a bunch of flags). This has caused cross-the-board test failures like . But this is not limited to just tests. Any kubernetes GCE users who rely on the and commands could have been broken. What you expected to happen: Stop updating gcloud automatically. We could either warn user to have the required dependencies installed, or pin to specific versions of dependencies like gcloud. How to reproduce it (as minimally and precisely as possible): download 1.4.4 release tarball run and Anything else do we need to know:\ncc\n/cc\ncc I think it's reasonable to remove gcloud update calls from scripts, as I believe gcloud is updated out of band by k8s continuous testing. The scripts will still need to work with latest, so maybe they should check for a minimum version.\nGood idea.\nYeah, we definitely want to continue using latest in CI so that we know if things break, but I think it's reasonable to remove the update itself from the kubernetes cluster util scripts."}
{"_id":"q-en-kubernetes-ec43ec725d8e5511e7f6b41402f22fa4501e2298bef344d682b1ad93a1330242","text":"Is this a BUG REPORT or FEATURE REQUEST? (choose one): Bug Report Kubernetes version (use ): Client Version: version.Info{Major:\"1\", Minor:\"4\", GitVersion:\"v1.4.1\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2016-10-10T18:19:49Z\", GoVersion:\"go1.7.1\", Compiler:\"gc\", Platform:\"darwin/amd64\"} Server Version: version.Info{Major:\"1\", Minor:\"3\", GitVersion:\"v1.3.9\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2016-10-20T20:37:18Z\", GoVersion:\"go1.6.2\", Compiler:\"gc\", Platform:\"linux/amd64\"} Environment: Cloud provider or hardware configuration: GKE OS (e.g. from /etc/os-release): Local: El Capitan, Remote: GCI Kernel (e.g. ): Local: Darwin ... 15.6.0 Darwin Kernel Version 15.6.0: Thu Sep 1 15:01:16 PDT 2016; root:xnu-3248.60.11~2/RELEASEX8664 x86_64 Install tools: gcloud Others: n/a What happened: Provisioned 3 clusters each in different regions (us-central1, asia-east1, europe-west1). Each cluster is in 2 zones (w/ gcloud --zone... --additional-zones...) Provisioned Federation API Server and Controller (1.4.4) based on Verified clusters were ready. Provisioned federated service & deployed backends into each of the cluster. Verified the federated services were created, and all of the regional LB were created, and verified all the LB IPs were working w/ direct requests. DNS provider updated DNS w/ most of the service/zone CNAME records. But A records did not complete. Further investigation for the log points to multiple entries in rrdata w/ the same IP due to the multi-zone deployment: I1029 19:24:24. 1 ] Have recordset <(clouddns) \"buttonmasher-\" type=CNAME rrdatas=[\"buttonmasher-\"] ttl=180. Need recordset <(clouddns) \"buttonmasher-\" type=A rrdatas=[\"130.211.203.196\" \"130.211.203.196\"] ttl=180I1029 19:24:24. 1 ] Existing recordset <(clouddns) \"buttonmasher-\" type=CNAME rrdatas=[\"buttonmasher-\"] ttl=180is not equivalent to needed recordset <(clouddns) \"buttonmasher-\" type=A rrdatas=[\"130.211.203.196\" \"130.211.203.196\"] ttl=180, removing existing and adding needed. After creating the clusters w/ only one zone, the DNS provider worked properly. What you expected to happen: Cloud DNS Provider should update the DNS record properly w/ multi-zone clusters. How to reproduce it (as minimally and precisely as possible): 3 x multi-zone clusters Federation API Server and Controller clusters in to federation federated service and backends Anything else do we need to know: Using GKE and Cloud DNS\n/cc\nAlso hit this. Any workarounds?\nSorry, no direct workarounds available yet. You could convert your multi-zone clusters into multiple single-zone clusters, and federate all of those.\nFYI, although I'm sure you're aware - also affects AWS multi AZ clusters (1.6.2)\nThis seems bad. I'll try to take a look into it.\nFixed by"}
{"_id":"q-en-kubernetes-3b2e0eeab3e4b0fe28e300fb528156280a71aa0b819c586f9f1942c2b6e9effd","text":"I just found a cluster like this cc/\ndo you know if anyone else is currently working on a fix for this? I've already started looking into a fix, would you be willing to review my PR soon?\nI don't know of anyone who is working on it.\n(I see your PR, and I'll review it.)"}
{"_id":"q-en-kubernetes-27d72288a184094e1823c145c5bb129d6cec3b0353eddc1b781ecf2920a527a2","text":"https://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\n[FLAKE-PING] This flaky-test issue would love to have more attention.\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nshould this issue be moved into the 1.5 milestone and is it a release blocker? The test seems to flake fairly infrequently but we need to get the issue triaged as to whether it's a 1.5 release blocker. Please update this issue by 1PM PST 30 November 2016. Thanks! cc:\nSince the feature it is testing is introduced for 1.5, so it should have the 1.5 milestone? This test is flaking because it is not well written, hence I don't think it is a release blocker. And I already sent out a which got LGTM before but recently rebased.\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nThanks for triaging\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}\nhttps://k8s- Failed: [] DNS horizontal autoscaling kube-dns-autoscaler should scale kube-dns pods in both nonfaulty and faulty scenarios {Kubernetes e2e suite}"}
{"_id":"q-en-kubernetes-5adf70279297b5fe9bdca1de3004e2b9e28bafb4aafef6480213e83552448f49","text":"Is this a request for help? No What keywords did you search in Kubernetes issues before filing this one? http proxy, httpproxy, httpsproxy, AWS, kubeadm, kubectl, cloud-provider Is this a BUG REPORT or FEATURE REQUEST? (choose one): BUG REPORT Kubernetes version (use ): v1.4.4 Environment: Cloud provider or hardware configuration: kubeadm on EC2 instances on AWS OS (e.g. from /etc/os-release): Ubuntu 16.04.1 LTS Kernel (e.g. ): Linux ip-10-0-133-99 4.4.0-45-generic -Ubuntu SMP Wed Oct 19 14:12:37 UTC 2016 x8664 x8664 x8664 GNU/Linux Install tools: kubeadm Others: httpproxy, httpsproxy and noproxy set in /etc/environment, Docker config as well on startup scripts that run What happened: won't complete and gets stuck on it's first deployment. When one looks through the logs it shows that the pods are downloading fine but we are stuck due to the kube-controller-manager continually crashing as it can't contact AWS. Elsewhere in kubernetes I see references to and if the AWS cloud provider could use that it would be able to connect as well. From the POD if I manually set httpproxy and httpsproxy and no_proxy I am able to use the AWS CLI without any issues. What you expected to happen: I would expect to successfully complete and use the environment proxies set up on the system so as not to depend on a default Internet Gateway being connected to the VPC just for AWS access. How to reproduce it (as minimally and precisely as possible): Set up a public VPC with Squid set up as a proxy Set up a private VPC peered to the public VPC Set up a master node and run Set up a regular node and run will never fully initialize. If you remove the --cloud-provider aws` then it will complete and nodes will register. Anything else do we need to know:\nwhat do you have in proxy variables ? (env | grep -i proxy) it tries to connect to (tcp 176.32.118.53:443) directly. You're expecting to connect there via proxy ? then do you have something like in your no_proxy ?\nUnfortunately on the pod itself for the the environment variables aren't passed through. As this is created using I don't have control over the pod definition. So basically none of the *_proxy variables are set inside the pod in the docker image, however if I and export the variables and try to wget that URL it works. Basically I'm hoping we can update to respect and pass the environment variables down to the system pods if they exist, eg, add: Without it I can't stand up a cluster with AWS integration on a completely private VPC and for such a small change it would make a rather big difference.\nok, so request is actually to pass proxy variables into containers. interesting. I'll check how it can be done. BTW, for noproxy value in your example: things like 10.0.0.0/8 would not work on 99% of software that respects noproxy variable. it must be always exact IP match.\nyou're right that what I'm actually asking for is to pass proxy variables through either from a commandline setting on kubeadm or pick it up from /etc/environment or shell environment that launches kubeadm. I didn't realize it till I dug in to try and think about how one might go about fixing this. At first I thought we would just pick up the variables from the system environment but then I realized kube-controller-manager is deployed in a container and so we'd need to update the yaml file that launches it to contain the environment variables. As I'd like to go to production with this soon hopefully this is a small change that can be .\nThis issue was moved to kubernetes/kubeadm"}
{"_id":"q-en-kubernetes-426889248db5620ca07fac41b496d891421966e3ac8c0e7abb4b6fc6d694a224","text":"Kubernetes version: HEAD What happened: Using interface to delete Job throws an error . What you expected to happen: Job should be deleted. We've found that issue after dependency update on dashboard: We are using to delete resources. It works fine for all other resources. I've investigated that and it appears to be the core issue. Other groups have registered as a known type. however batch is missing it After adding I was able to successfully delete Job using our code. I can quickly create PR for that. Just need a confirmation that this is indeed a bug. cc\n+1 apis/autoscaling is similar\nI thought we only needed to register it in the external versions (). I'll double check. cc\nRegistering api.DeleteOptions in batch looks fine. In the future, I think it's more proper to register DeleteOptions and ListOptions as unversioned, like how we treat .\nDeleteOptions and ListOptions are expected to be in their group. They have to be in internal version and external version. Most of this code will be fixed in 1.6 (we'll standardize registering the \"common\" things). , Chao Xu wrote:\nI've created PR to fix this issue. Could anyone check that?"}
{"_id":"q-en-kubernetes-8fb3dd61d24c121f4676c41b25a663a9a477a13b64ab6c01bebf87cc2a6ec090","text":"Docker daemon complains \"invalid hostname format\" if the truncated host name ends with \".\" The truncation occurs .\nDocker checks hostname with the regexp (See ): It means that the hostname must end with an alnum, not or . This issue has been here for a long time. This usually happens to static pod, because we compose a looooong name for static pod. We didn't hit this just because we are lucky. :P Will send a PR to fix this. /cc\n: Needed for v1.5? please add milestone if so\nThanks for reminding. :) This is a bug fix in kubelet. Mark 1.5. /cc\nAre you working on a fix?\nYeah, the fix is here"}
{"_id":"q-en-kubernetes-dd03a69a67e6bc38eccba3ae9ea8cb7e3e19d9fe330b4a5be1d5cd7c9dc14ac1","text":"The error reporting functions here all repeat a URI without sanitizing it. An XSS attacker could use this to get the server to embed a reference to a malicious script. (As always, please note that kubernetes apiserver is NOT xss hardened and we do NOT recommend using the api directly from a browser.)\nThe error correctly sets the content type to , so this is only an issue using a browser that does content sniffing. Would it be sufficient to just add the header here?\nYes, that will work for browsers that honor the header. It's probably sufficient-- we could do that globally, we set Content-Type correctly everywhere, I think."}
{"_id":"q-en-kubernetes-769dd910c1537dacc371202c0fa5cf74ca58673ade516edca14d2971f96d4408","text":"Failed: https://k8s- Run so broken it didn't make JUnit output!\nSen what is the plan for migrating these to bootstrap?\nisn't it a manual run test?\nI guess I'll move it over with empty cron-string?\nWhy did run this test? Please don't run large and enormous tests - this requires coordination w SREs etcd and may potentially break GCE.\nI don't think we've run these test ourselves, someone ran it and failed(10 days ago) thus this test failure occurred."}
{"_id":"q-en-kubernetes-2fc397f88476d226bb42e5473e12b15e4e80da299ec2672fb4b931831893ed24","text":"Failed: https://k8s- Run so broken it didn't make JUnit output!\nBootstrap?\nWhy did run this test? Please don't run large and enormous tests - this requires coordination w SREs etcd and may potentially break GCE.\n/close hummm I just see the index is -1"}
{"_id":"q-en-kubernetes-63fa7f299df0a835444f5322257d02b3e772af5f6d05ad1050a4f3637d35996e","text":"Kubernetes version (use ): 6 Environment: GCE What happened: I created a cluster for Heapster testing purposes to debug failing test in It turned out that Summary API does not export system/misc stats in SystemContainers section (though there are kubelet and docker stats). Everything works as expected on nodes. See also https://k8s-\nWhat node OS were you using? We don't create the system (misc) container on systemd, so it's expected that GCI nodes don't report those stats.\nThe default configuration for 1.4, which is: master: GCI nodes: Container VM\nAre the nodes running 1.4.6 too? There was a bug with this in 1.4.5, but it should have been fixed in 1.4.6.\nYes, both are running in 1.4.6. If you want to reproduce from Heapster current head run . It will take some time and as a side effect you should have a cluster with the mentioned problem. This was catch by integration tests and I temporary disabled the check here\nif you have spare cycles, can you look into this?\nfriendly ping\nLooking at it now\nI have found a couple things so far. On the containerVM nodes, cadvisor is exporting a container with \"name\": \"/system\" On the GCI Master, cadvisor is exporting a container with \"name\": \"\" The has this: \"{% set system_container = \"--system-cgroups=/system\" -%}\" (/system is the default misc container name) So the real question is: Should cadvisor export \"/system\" instead of \"\" on GCI? Or should we be configuring GCI to use \"--system-\"?\nI thought we fix this very issue for 1.4 branch by a while back. Are you running into the same issue with 1.5?\nI think I told you the wrong thing earlier. We don't support the misc (/system) cgroup on systemd systems (I can't remember why, but it was an explicit decision - or might know). The is something different, I believe a base systemd cgroup (slice). As for the misc container on ContainerVM nodes, we've been testing for the misc container in 1.5+ in the summary api test (https://k8s-) for a while, and I haven't seen this issue. So if this is still reproducible, it's limited to 1.4. See this the bug was created for the post-fix.\nFound the explanation here: The should use the flag to know the location that has the processes that are associated with , but it should not modify the cgroups of existing processes on the system during bootstrapping of the node. This is because is the on the host and it has not delegated authority to the to change how it manages .\nIIUC this is intended behavior on systemd nodes. Based on the explanation provided, the kubelet doesnt have the authority to control \"misc\" processes on systemd.\n+1. \"misc\" is just a bandaid for cvm. , David Ashpole <:\nOk, so I'm closing this issue."}
{"_id":"q-en-kubernetes-38d52777560a5c6fb71d98b8b0953e61c4be2731ef4a377191b06cc53fcf6655","text":"https://k8s- Failed: [] Network Partition [Disruptive] [Slow] [] [ReplicationController] should eagerly create replacement pod during network partition when termination grace is non-zero {Kubernetes e2e suite}\nhttps://k8s- seems to be the first failure here. It seems to have gone in with a bunch of RC/RS changes () cc/"}
{"_id":"q-en-kubernetes-446565579dd5b2ae18b5f5c4d6bc54767727d41c59b8aef46949abc5084c3f3a","text":"Is this a BUG REPORT or FEATURE REQUEST? BUG REPORT Environment: Cloud provider or hardware configuration: GCE OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others: What happened: $ kubectl describe pods test-ipj93 Name: test-ipj93 Namespace: default Node: kubernetes-minion-group-yxra/10.128.0.3 Start Time: Mon, 28 Nov 2016 14:39:28 -0800 Labels: <noneStatus: Pending IP: Controllers: <noneContainers: busybox: Container ID: Image: busybox Image ID: Port: Command: sh -c while true; do echo (hostname)-(cat /proc/stat cut -f 4 -d\" \"); sleep 30; done QoS Tier: cpu: Burstable memory: BestEffort Requests: cpu: 100m State: Waiting Reason: ContainerCreating Ready: False Restart Count: 0 Environment Variables: Conditions: Type Status Initialized True Ready False PodScheduled True Volumes: data: Type: PersistentVolumeClaim (a reference to a PersistentVolumeClaim in the same namespace) ClaimName: test ReadOnly: false default-token-kjoqg: Type: Secret (a volume populated by a Secret) SecretName: default-token-kjoqg Events: FirstSeen LastSeen Count From SubobjectPath Type Reason Message --------- -------- ----- ---- ------------- -------- ------ ------- 57m 57m 1 {default-scheduler } Normal Scheduled Successfully assigned test-ipj93 to kubernetes-minion-group-yxra 57m 57m 1 {kubelet kubernetes-minion-group-yxra} spec.containers{busybox} Normal Pulling pulling image \"busybox\" 57m 57m 1 {kubelet kubernetes-minion-group-yxra} spec.containers{busybox} Normal Pulled Successfully pulled image \"busybox\" 57m 57m 1 {kubelet kubernetes-minion-group-yxra} spec.containers{busybox} Normal Created Created container with docker id ; Security:[seccomp=unconfined] 57m 57m 1 {kubelet kubernetes-minion-group-yxra} spec.containers{busybox} Normal Started Started container with docker id 47m 47m 6 {kubelet kubernetes-minion-group-yxra} Warning FailedMount MountVolume.SetUp failed for volume \"\" (spec.Name: \"default-token-kjoqg\") pod \"-b5bb-11e6-bab0-\" (UID: \"-b5bb-11e6-bab0-\") with: Get https://kubernetes-master/api/v1/namespaces/default/secrets/default-token-kjoqg: dial tcp: lookup kubernetes-master on 169.254.169.254:53: no such host 46m 46m 1 {kubelet kubernetes-minion-group-yxra} Warning FailedMount MountVolume.SetUp failed for volume \"\" (spec.Name: \"default-token-kjoqg\") pod \"-b5bb-11e6-bab0-\" (UID: \"-b5bb-11e6-bab0-\") with: Get https://kubernetes-master/api/v1/namespaces/default/secrets/default-token-kjoqg: dial tcp 10.128.0.2:443: i/o timeout 45m 45m 1 {kubelet kubernetes-minion-group-yxra} Warning FailedMount MountVolume.SetUp failed for volume \"\" (spec.Name: \"default-token-kjoqg\") pod \"-b5bb-11e6-bab0-\" (UID: \"-b5bb-11e6-bab0-\") with: Get https://kubernetes-master/api/v1/namespaces/default/secrets/default-token-kjoqg: dial tcp 10.128.0.2:443: getsockopt: connection refused 11m 11m 2 {controller-manager } Warning FailedMount Failed to attach volume \"test\" on node \"kubernetes-minion-group-yxra\" with: error getting instance \"kubernetes-minion-group-yxra\" 9m 9m 1 {kubelet kubernetes-minion-group-yxra} Warning FailedSync Error syncing pod, skipping: network is not ready: [Kubenet does not have netConfig. This is most likely due to lack of PodCIDR] 9m 9m 2 {kubelet kubernetes-minion-group-yxra} Warning FailedSync Error syncing pod, skipping: network is not ready: [Kubenet does not have netConfig. This is most likely due to lack of PodCIDR] 6m 14s 4 {kubelet kubernetes-minion-group-yxra} Warning FailedMount Unable to mount volumes for pod \"test-ipj93_default(-b5bb-11e6-bab0-)\": timeout expired waiting for volumes to attach/mount for pod \"test-ipj93\"/\"default\". list of unattached/unmounted volumes=[data] 6m 14s 4 {kubelet kubernetes-minion-group-yxra} Warning FailedSync Error syncing pod, skipping: timeout expired waiting for volumes to attach/mount for pod \"test-ipj93\"/\"default\". list of unattached/unmounted volumes=[data] What you expected to happen: GCE PD to attach and mount successfully post the upgrade How to reproduce it (as minimally and precisely as possible): Running the upgrade steps in the doc for upgrade of manual test from 1.4 to 1.5 Anything else do we need to know:\nRan the upgrade again. This time the .\nkube-controller-\nCurrent theory is because kubelet deletes and recreates the node API object during upgrade (because externalID changed), see the new object does not contain the correct status (it is written by the Attach/Detach controller to the old object but not subseqently updated). If this proves to be the issue, the proposed fix is to force the attach/detach controller to update the node status when the node's external ID changes. is working to verify this theory by executing this code path (without a full upgrade). Once he has validated the theory, he will work on the fix, verify it, merge it to master, and get it cherry-picked to 1.5.\n: is expected to fix this problem?\n: Yes"}
{"_id":"q-en-kubernetes-1c68361b9bf0a52e9d7a25ecd5b153749ee1e346cf6ed3e97529d309fc856401","text":"Slack user \"nalin\" reported an issue Looking at the code this makes sense. Will fix.\nFor anyone following along, the fix for this bug unfortunately didn't get cherrypicked because of a process issue, and is broken until 1.5.2. cc\nI just the new lines from As a temperatur fix ONLY. If anyone like me dont want to wait. Worked like at charm for AWS cluster setup. :)\nA quick note, just as confirmation: the above temporary fix does resolve the issue for AWS - even though it's a change to a GCE-related file (cluster/gce/configure-). It seems that the GCE files are used for certain aws functions.\nIn based AWS deployment, the script is shared between AWS and GCE.\nAlso, this fix was cherry-picked into the branch a while ago, so instead of manually adding the lines, you can also just checkout the branch and run from there."}
{"_id":"q-en-kubernetes-f96f989329e6fda0a97e9e4c684a7c54505fe3d7c8195a165e691ec2455ae62a","text":"Is this a BUG REPORT or FEATURE REQUEST? (choose one): bug Kubernetes version (use ): master Environment: Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others: What happened: When testing , I found kube-proxy can not record event with following errors. What you expected to happen: No errors How to reproduce it (as minimally and precisely as possible): cc\nI am still getting this with the latest build\n, thanks for your feedback, I'll check it."}
{"_id":"q-en-kubernetes-99f20a6006e53b856138dd520ed013802aa15e162b7578d9adad8b350decc8ea","text":"kubectl edit is not working on HEAD. It's working on 1.5."}
{"_id":"q-en-kubernetes-21ec941edfa0690a26ff395dea19961384462e3c6c950d124e082489371eeaf3","text":"<!-- Thanks for filing an issue! Before hitting the button, please answer these questions.--Is this a request for help? (If yes, you should use our troubleshooting guide and community support channels, see ): no What keywords did you search in Kubernetes issues before filing this one? (If you have found any duplicates, you should instead reply there.): subpath, emptyDir, mode, permissions Is this a BUG REPORT or FEATURE REQUEST? (choose one): BUG REPORT <!-- If this is a BUG REPORT, please: Fill in as much of the template below as you can. If you leave out information, we can't help you as well. If this is a FEATURE REQUEST, please: Describe in detail the feature/behavior/change you'd like to see. In both cases, be ready for followup questions, and please respond in a timely manner. If we can't reproduce a bug or think a feature already exists, we might close your issue. If we're wrong, PLEASE feel free to reopen it and explain why. --Kubernetes version (use ): Client Version: version.Info{Major:\"1\", Minor:\"5\", GitVersion:\"v1.5.2\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2017-01-12T04:57:25Z\", GoVersion:\"go1.7.4\", Compiler:\"gc\", Platform:\"linux/amd64\"} Server Version: version.Info{Major:\"1\", Minor:\"5\", GitVersion:\"v1.5.2\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2017-01-12T04:52:34Z\", GoVersion:\"go1.7.4\", Compiler:\"gc\", Platform:\"linux/amd64\"} Environment: Cloud provider or hardware configuration: Bare metal, SSD, but also VMs OS (e.g. from /etc/os-release): Debian GNU/Linux 8 (jessie) Kernel (e.g. ): 4.8.0-0.bpo.2-amd64 Install tools: Others: Docker v1.12.6 What happened: When I mount an volume using a subpath, I'm unable to write into it if my process is not running as . When I leave out the subpath, I can. What you expected to happen: I can write to the directory in both cases. How to reproduce it (as minimally and precisely as possible): Output is: Anything else we need to know: This issue is related to . I know this is caused by the Docker daemon creating those directories on a bind mount. Nevertheless, I expect Kubernetes to abstract this away, since this behavior is neither documented nor IMHO intuitive.\nCould someone from please take a look? Thanks! cc:\ncc/\nattempted fix created here\nI created a cherrypick PR for 1.6 here: since this bug is against 1.6. Could someone (who agrees this could be cherrypicked) please add the cherrypick-candidate label + milestone to the original PR , thank you! edit: sorry, it's against 1.5. Should it be cherrypicked back to that as well?\nI'm still facing this issue with Kubernetes version 1.12.8 on a AKS cluster when I mount a Azure Storage as azureFile share. Did anyone know if the issue is fixed for all types of mounts and not only just the emptyDir?\nsame issue on GCP @ k8s 1.14\nk8s 1.17 via eks as well\nSo what's the recommendation? Not using subPath?"}
{"_id":"q-en-kubernetes-fcaaa5c3446bc6cbea8f286101bf814a794ea21416d8194f6468c562e39fb1cd","text":"getNodeConditionPredicate() in makes our code hard to understand because it hides the node-condition-based filtering in the node lister, which is totally non-obvious. We should get rid of this function and have NodeController add NoSchedule taints for these situations instead. (Alas, we might not actually be able to get rid of getNodeConditionPredicate() completely until we get rid of the Unschedulable field of PodSpec, but at least we can get rid of all the other code in this function.) If there are pods that should still be able to schedule in any of these situations (e.g. DaemonSet pods?) we should add tolerations in admission control for them (e.g. see ). cc/ cc/ Sub-tasks according to the design doc: [x] Add node taints label and feature flag () [x] In the node controller, taint Nodes according to the Node Condition () [x] In DaemonSet, update DaemonSetController to Tolerant new Taints () [x] In admissionController, add MemoryPressure/DiskPressure toleration for no BestEffort pod () [x] In scheduler, disable the current behaviour of filtering out Nodes. Instead, pods will not be scheduled to tainted nodes if a toleration does not exist in its PodSpec () [x] Add e2e test for this feature () [x] Update doc for this feature ()\nI think you are referring Unschedulable field of NodeSpec?\nYes sorry, that was a typo.\n- please CC me to all PRs\nSee also\nNote that memory and disk pressure are checked in separate predicates, not in the NodeConditionPredicate. These should be represented as NoSchedule taints too.\nSetting NoSchedule taints corresponding to the node conditions that today cause pods not to schedule onto a node, should also allow us to use tolerations rather than special-case code in the DaemonSet controller (and thereby make it easier to have DaemonSet pods be scheduled by the default scheduler.)\nref/\nThe task list was moved to the issue description .\n/cc\n/reopen\nyou can't re-open an issue/PR unless you authored it or you are assigned to it. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign\n/reopen\n/reopen\nCurrently we already have toleration on dsc for taint related to node problems (e.g NodeNotReady and NodeUnreachable). But to me ds should respect the new taints (outOfDisk, NodeMemoryPressure etc..), any insights ?\n/reopen\nYes, it's one of sub-task of this project. This issue is closed by 'partially fixes' :(, I had updated PRs to 'part of #'.\n- I can handle the dsc tolerations if you want. cc to keep you on the loop.\n, sure, thanks very much :).\n- just saw that you moved node conditions into predicate func, once adding node taiting and the according tolerations to dsc, do we'll still need to have node conditions checking ? (since the we're tainting by node condition). Also, correct me if I'm wrong, but from what I understood: dsc tolerates new taints with effect noExec since we don't wanna evict any already-running pods ?\nWe won't, but it'll take time for us to get there. We'll roll out Tainting as alpha (obviously), later move it to beta and GA in 1.10 at the earliest. We can't remove Condition checking logic before that. dsc is a daemon set right? New taints are NoSchedule only, so running Pods won't be affected (at least on the control plane level). Kubelet can still decide to evict them based on problem it observes.\n- seems logic for the timeline. From sub-tasks Since new Taints are NoSchedule Only, Daemon set shouldn't have toleration for these ? It should instead respect the new taints when applied and not schedule on these nodes, or I'm I missing something ?\nThe task for DS in my mind: For all DS pods, it should tolerant MemoryPressure & DiskPressure For critical DS pods, it should tolerant OutOfDisk And if the taint by condition feature is enabled, we should check OutOfDisk by taints instead of predicates; the refactor is trying to make this simpler :).\n- You're right, I'll send a PR after yours get merged :)\nDo we have a to do list so folks can help out with I'd like to help when I get some time. I have a question about timeline. If alpha makes it into v1.8, will I be able to schedule to NotReady nodes (no CNI) if I reference the alpha tolerations? Or do you mean it won't be functional at all until v1.10?\nIt means you'll need to flip feature gate flag for it to work.\n, here is the task list: I'm working on related PRs for them; please ping me when you got time :).\n- i anticipate we will also add a CPUPressure condition which will also mean that no BestEffort pod can be scheduled to that node if the kubelet is running with a static cpu assignment policy. /cc\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. Prevent issues from auto-closing with an comment. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nI think it's done (as alpha). - do we want to keep this issue open?\nWe can close this."}
{"_id":"q-en-kubernetes-864550028401addaf71603a71d8e6fbe66d5e072ba1dc1a94c6030f36bc0f906","text":"https://k8s- Pod has a restart count 1 FYI, all relevant messages about this pod in the From the sync loop timeline, some containers died. Check the log again: The sandbox was GC'd right after it was created. This led to kubelet killing all the containers in the pod since the sandbox needed to be recreated. Need to check whether this was caused by the container state emulated by fakedockerclient, or by a real bug. /cc\n/cc FYI.\nkuberuntime GC routine has a minAge protection such that it can't GC any pod sandbox that was created no more than 30 seconds ago). This doesn't seem to have be respected from the log. I suspect that the fakedockerclient was not returning a proper timestamp.\nGC logic might be faulty. I see a similar failure in https://k8s- EDIT: The node name is The sandbox was created and removed within 2 seconds\n- by looking at the logs, can you tell if there's an easy repro?\nSo given that this happens also in non-kubemark suites, it seems like a release blocker to me. Anyway - - do you think it's releated to number of pod on a node, or it's just that it happens from time to time, and in scalability suites (no matter if it's kubemark or not) we have more of those tries (due to scale)?\nI think the timestamp comparison logic might be wrong in the sandbox garbage collector. This would only happen (in my theory) when the the snapshot of all containers/sandboxes on the node are taken when: sandbox is created but not running (started) yet. other containers in the pod haven't not been created. In this case, the fail-safe (minAge) kicks in, and apparently failed in the tests discussed above. This would more likely to happen when during batch creation of more pods. kubelet would usually heal itself by restarting all container in the pod (which has the undesirable side-effect of killing the user containers of course). I think other tests might not catch this because 1) they don't perform large batch creation, and 2) they don't check the restart count.\nOh the above comment was completely based on my theory. Reproducing this deterministically on a working node may need some hacking in the kubelet code (e.g., inject a long enough sleep after sandbox creation). I will try writing a unit test first to verify the behavior.\nCan't reproduce this in my cluster with artificially injected latency... Adding more debug information in For people who are interested in checking the logs and identify the victim pod, here are the instructions: Download the artifacts: E.g., Find the (density) pod with restart count 0:\nDidn't reproduce the problem either. I think so. GC can't distinguish just-created sandboxes from failed sandboxes. Especially when the time from sandbox creation to ready is longer than GC's minAge. I'm wondering should runtime (e.g. dockershim) filter out those created-but-not-ready sandboxes from sandbox list? or introduce another sandbox state in CRI?\nhas been merged 11 hours ago and scalability suite hasn't flaked since. Will monitor and see if there is any need to reopen.\nOK. But because , there are still potential problems under big pressure.\nFirst, this issue is no longer a release blocker after the fix. Let's open a separate issue to discuss potential improvement, instead of commenting on a closed issue :) Also, FWIW, we used the same logic in the old docker manager, so this is also not a regression.\nAgreed. Opened for this."}
{"_id":"q-en-kubernetes-695d65dcca8e2dfbf0a831cfd440326e17e0bfbd602f29a058bd66c19c3ed1ee","text":"<!-- Thanks for filing an issue! Before hitting the button, please answer these questions.--Is this a request for help? (If yes, you should use our troubleshooting guide and community support channels, see ): What keywords did you search in Kubernetes issues before filing this one? (If you have found any duplicates, you should instead reply there.): Is this a BUG REPORT or FEATURE REQUEST? (choose one): <!-- If this is a BUG REPORT, please: Fill in as much of the template below as you can. If you leave out information, we can't help you as well. If this is a FEATURE REQUEST, please: Describe in detail the feature/behavior/change you'd like to see. In both cases, be ready for followup questions, and please respond in a timely manner. If we can't reproduce a bug or think a feature already exists, we might close your issue. If we're wrong, PLEASE feel free to reopen it and explain why. --BUG REPORT Kubernetes version (use ): Client Version: version.Info{Major:\"1\", Minor:\"6+\", GitVersion:\"v1.6.0-beta.3\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2017-03-11T04:20:22Z\", GoVersion:\"go1.7.5\", Compiler:\"gc\", Platform:\"darwin/amd64\"} Server Version: version.Info{Major:\"1\", Minor:\"5\", GitVersion:\"v1.5.4\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2017-03-07T23:34:32Z\", GoVersion:\"go1.7.4\", Compiler:\"gc\", Platform:\"linux/amd64\"} Environment: Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others: What happened: With kubectl version 1.6.0-beta.3, the output of is not aligned: What you expected to happen: With kubectl version 1.5.4, the output is as follows: How to reproduce it (as minimally and precisely as possible): Run using kubectl version 1.6.0-beta.3. Anything else we need to know:\nwill take a look\nany progress on this?\nhi, tony, sorry, I haven't look at this problem yet, maybe tomorrow, just really busy recent day, you can take it if you already find the problem, thanks\nHappened to see the same issue on my laptop, I can help fix this issue if you do not have time, thx\nthanks, I've been really busy this week, go ahead\nfor this ticket, want to discuss one more thing with you: at present, for each of the update event, command with option will print out a line showing the status change of this event. But this behavior is different with what we saw from the output of linux's command, which I think is not a very good user experience. For this kind of command, we shall reset all the old output and re-print all the pod's information in the standard output, this will make this command more like a real command. do we have any plans on this right now? welcome your ideas on this.\nI think both output format has it's advantage, for current output format, user can copy paste the whole information like checking logs, for linux's command style output which user can stare on only one screen(if it's not too many pods or sth else).\nyep, we do have plan to do this, and I really like a more linux style command, you can refer for more detail. Anyway let's just get this fixed at first."}
{"_id":"q-en-kubernetes-eec911951834a8b68cfeb68af0ebaad86bbebfbac3c64caf805351aa1deaa763","text":"This is in 1.6.0-beta.4: Some context: The line number suggests it is here: I think this is because dockershim now uses a protobuf object and Protocol is now an int. And is allowed in go (apparently) - is \"A\". And the comedic error is because Protocol implements If this is the error, there's something fishy going on because all our UDP tests should fail?\nAha ... I think it is only UDP hostports that would be \"broken\" by the issue. I wouldn't be surprised if we don't rely on any, and that we might not have any test coverage. I don't think this is a regression - I think this code has been there since the start of dockershim, but maybe CRI is now the default? I don't know how widely UDP hostports are used. Given how close we are to the planned release date, cc and\nThere is something wrong in dockershim's protocol parsing. Will file an PR to fix this. /assgin\n/assign\nOpened to fix this.\nthanks for filing the issue. Thanks for quickly addressing the bug. We need node e2e tests to capture the kind of regression at the first place."}
{"_id":"q-en-kubernetes-d40be4f3c031760a38b6058a0976a10a321500417ed42ab8e5c74a6784064b8c","text":"From https://k8s- I've tried to reproduce locally in an endless loop but no failures so far. cc\nInvestigating."}
{"_id":"q-en-kubernetes-9d8298ea29df6af00e52aa7e0daf8d5a57bb31b717f4ee1d5c3cf564f9bbe147","text":"Is this a BUG REPORT or FEATURE REQUEST? (choose one): BUG Kubernetes version (use ): k8s1.3.5. still exist in 1.6 What happened: when exec threadSafeMap.Index func, the returned items contain nil object. That is to say, index is in threadSafeMap but the corresponding has already been deleted. What you expected to happen: when object is deleted from threadSafeMap, all its index must be deleted. How to reproduce it (as minimally and precisely as possible): Suppose we init threadSafeMap with two indexers: indexFunc1 and indexFunc2. Obviously, indexFunc may return error if the object has something wrong or some other reasons. when we add an object, will first add the object to , then compute the index and add the index into threadSafe.indices. when we delete the object, will first delete the index from indices, then delete it from Suppose IndexFunc1(object) return error and indexFunc2(object) works well. I think this circumstance is absolutely possible to happen. Also we know that the order to exec indexFunc is not decided. Feature of iterating map. When we add object, suppose first exec indexFunc2, and then exec indexFunc1. So the index of indexFunc2 is to indices. But the index of indexFunc1 is not , because indexFunc1(obj) return err. Then we delete the object, suppose first exec indexFunc1. It will return error and IndexFunc2 has no chance to run. So the corresponding index indexFunc2(obj) can't be removed. But then the object is removed from items. This problem may result in subtle problems. For example, when we exec threadSafeMap.Index( ) func the returned list may include nil object. Anything else we need to know The reason is that index is still is threadSafeMap but the corresponding object has already been removed. We must clear all corresponding indexs when delete object. And the reason why index is not deleted is that when for-range map, the order is not decided.\nWho can help me? Do i miss something important or misunderstand it ?\n, can you share a unit test to show it? I'd like to help on it.\nThank you for your help. I think i have decribed the condition resulting in the bug indetail. Be simple, because the order of exec indexFunc is random when update indices, and some function don't care the error returned, the obj is deleted but corresponding indices is still in it. NOTE: the obj is deleted but corresponding indices is still in it So when we run index func, the returned list may contain nil object. Because when get obj from threadsafemap with indice, not care whether the obj exist in this map. Then nil object .\n. I am sure this problem exists. And i think we can find a better way to deal with it.\n/sig api-machinery threadSafeMap is in client-go\n/kind bug\nOpened a PR with a fix.\nIn my understanding, should be deterministic, could you elaborate more how your IndexFunc implemented? If an object can not be called on this , then the question is, how could it exist in at the first beginning?\nthreadSafeStore should tolerate some bad conditions and deal with it gracefully, instead generating strange and confusing phenomenon. ThreadSafeStore can contain mutiple index funcs. If one of them is not implemented very well, or one of objects we pass to it has something wrong, STRANGE AND CONFUSING phenomenon happens, see what I said above.\nThat's the point I would like to know more, the call on an object should succeed or fail unconditionally(deterministic).\nwechat\nI also think index function must be a pure function. See my comment on the pr"}
{"_id":"q-en-kubernetes-7a75e672d6d43486fd2e9511dc66d807e860b23e95e556b4fe4828502922e2cc","text":"I was confused by the that looks like more as machine readable rather than as human readable. I feel that UX could be better if it will be written in multiple lines and without curly braces/quotes:\nAgreed on this one, we should pretty-print this output.\nI would like to change that tomorrow if you are not working on it.\nI'm not working on it, so feel free to provide a PR. But I wouldn't do anything until we get approval from\ncc don't some of the tests rely on this output?\nThey probably do. There should be a option or something that we can use in test infra.\nplease have a look at: and related issue: The above is approved and had an lgtm labeled, but removed by the bot because of testing flake.\nI think we can close this ?\nI don't see that it was fixed. Does the output of was improved?\nkubectl now has a option which lets you select between YAML or JSON. However, I believe this is reversed. The more human-friendly output is hidden behind an option and the machine-friendly output is the default. I believe it should be the opposite."}
{"_id":"q-en-kubernetes-88466bc637a805045217da94feabbdbd3df0ba52cd0d9904e95d3984c4cf4efa","text":"( noticed this) HardPodAffinitySymmetricWeight should be in the , not in KubeSchedulerConfiguration If/when fixing this, we need to be careful to do it in a backward-compatilble way.\nDo you mean ? It seems scheduler policy config's args can not passed to FitPredicates; it need to update related interface. For the : mark to be deprecated and set scheduler policy config to HardPodAffinitySymmetricWeight if policy config is empty.\nAre you working on this? if no, I can take it.\n, nop; please go ahead.\n/ / / , When reviewing the PR, there's one question on : the default value of is 1, which means the weight (1) will always add to priorities if term matched (), does it our expectation? And the user has to set it to 0 to disable it. IMO, the default value of should be 0.\nThe value 1 (instead of 0) was chosen so that if you say A must run with B, there is an automatic preference that B should run with A. This seems like reasonable behavior.\nThanks very much; reasonable case to me. But it seems we can not disable it: if set to 0, the default funcs will set it to default (1); similar result to un-set. Anyway, no user complain this now :)."}
{"_id":"q-en-kubernetes-4c1b3784abeb1df3f479372d478ac7f62f51c3ca0f676eb96392db2c96534617","text":"<!-- Thanks for filing an issue! Before hitting the button, please answer these questions.--Is this a request for help? (If yes, you should use our troubleshooting guide and community support channels, see ): What keywords did you search in Kubernetes issues before filing this one? (If you have found any duplicates, you should instead reply there.): Is this a BUG REPORT or FEATURE REQUEST? (choose one): BUG REPORT <!-- If this is a BUG REPORT, please: Fill in as much of the template below as you can. If you leave out information, we can't help you as well. If this is a FEATURE REQUEST, please: Describe in detail the feature/behavior/change you'd like to see. In both cases, be ready for followup questions, and please respond in a timely manner. If we can't reproduce a bug or think a feature already exists, we might close your issue. If we're wrong, PLEASE feel free to reopen it and explain why. --Kubernetes version (use ): build--5-v1.7.5-3 Environment: Cloud provider or hardware configuration: Laptop (bare metal) OS (e.g. from /etc/os-release): Fedora 26 Kernel (e.g. ): Linux localhost.localdomain 4.11.0-0.rc6.git0.1.fc26.x8664 SMP Mon Apr 10 19:45:50 UTC 2017 x8664 x8664 x8664 GNU/Linux Install tools: make && hack/local-cluster- -o _output/bin What happened: built from and ran with local-cluster- All commands gave the following error: Running showed the output: What you expected to happen: for commands to work, and Content-type to show How to reproduce it (as minimally and precisely as possible):\nccing\nI see the text plain content type, but not kubectl errors\ni was using the wrong version of kubectl. updated to new and i'm not getting an error any longer.\nText/plain coming back from the server is still an issue\ncc\nI'm not seeing against head and the pastebin is gone. Did you reproduce?\nI did last week. Insecure and secured port behaved differently\nYou can reproduce this using against a Kubernetes cluster built from HEAD\nHi all, I think we could use write Content-type correctly. I have made a PR to fix this."}
{"_id":"q-en-kubernetes-dcf1ddb387b3e0b3cd259f5a93387d799e78e4196d0022f61b297b306e0dbf83","text":"Currently if one wants to view secrets stored in k8s they must follow the process described here: This involves copy&pasting. It seems to me that kubectl could provide a more helpful interface to base64 decrypt the users on behalf of the user saving them the trouble.\nThe kubelet does decode secrets before exposing them to the pod. Also the title doesn't really make sense. Secrets aren't encrypted so there's nothing to decrypt.\n/kind feature\nYes, though this is low priority since we mostly don't want end users to view these once they get into the system.\nWhy can't that be an API permissions issue? I mean, if I really want to see it, I can just launch a pod to tell me.\nI've update the issue title with what I understand as the request. To the best of my understanding this is just a request for us to avoid the need to pipe into base64 when inspecting secrets.\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. Prevent issues from auto-closing with an comment. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten /remove-lifecycle stale\nI would like to try to make this work before it rots. cc Any guidance on a preferred implementation?\nWhat do you think about adding support for functions in templates? This would allow someone to use the existing and a template that uses the function -- or a wide variety of other interesting functions as a side effect. I believe (but have not yet tested) that the following might be sufficient to enable the requested behaviour (plus the change to the dependencies list, of course):\nHi I suppose my thoughts on this issue have changed a little since I wrote it. The situation isn't quite as bad as I described as it is possible to decode a secret in one go without then need to copy and paste. i.e. That said, that is a bit of a mouthful. Also, it might be nice to more easily 'work' with individual secrets and set them as well as get them. Should a user really need to pull down an entire group of secrets and edit a file just to be able to update a single password? For example something like this might be possible: What do people think? PS: Should this have been an issue on [kubectl] instead? PPS: I think some of the ugliness of this comes from configmaps being very similar to secrets except for the base64 encoding. Have there been any thoughts around bringing them into line? If memory serves, it might actually be nicer, due to YAML syntax-shenanigans, if configmaps were base64 encoded too.) [kubectl]:\nHi ... yes, I was thinking that it would be handy to be able to do something like: which is essentially the same as what you did without the overhead of the base64 process, or embed the information in a larger template like or in theory be able to use any of the other Sprig functions if someone found them useful. I think that having the Sprig methods available for generic formatting would just be useful on its own, really. On the topic of a or a , that seems pretty cool. There seem to be quite a lot of options for that I think people might expect to see mirrored in a if it were to exist. Thoughts?\nIt's also come to my attention that works great on MacOS but on Linux it's , so if someone, say, has a Helm chart where the notes tell the user to retrieve the secret data using one or the other of these, it won't work on the other one. would clearly be a simpler experience for the user than Hmm... thinking further: does does the idea conflict with the documentation in ? Perhaps we'd have to do something like: or similar?\n/remove-lifecycle rotten\nSmall tip: you can use which works on both Linux and Mac.\nThank you\nUpdate: I've looked at the huge amount of additional code that would be imported as a result of importing sprig and decided to go with a much smaller change instead, which is simply to implement a template function using . See PR with code and test.\nplease see the linked PR which addresses the original issue in a minimal fashion -- it simply provides a template function. It has passed all the tests and needs review; if it's good enough to solve the original problem I'm happy to squash the commits and can hope to get it merged. As far as solving evolved description of the problem, I'm not sure where to start or if you really want me to go in that direction. Would you consider taking this change and splitting the additional work for being able to get / set the contents of an individual key in a secret into a separate issue? If you have any guidance on what you'd like to see with that I'd be happy to take a look, but it might be out of my current grasp."}
{"_id":"q-en-kubernetes-298d01bdc71933da3b606790f6cbcc0134996e8eb74a2ee54bef994920502ad1","text":"<!-- Thanks for filing an issue! Before hitting the button, please answer these questions.--Is this a request for help? (If yes, you should use our troubleshooting guide and community support channels, see ): What keywords did you search in Kubernetes issues before filing this one? (If you have found any duplicates, you should instead reply there.): Is this a BUG REPORT or FEATURE REQUEST? (choose one): <!-- If this is a BUG REPORT, please: Fill in as much of the template below as you can. If you leave out information, we can't help you as well. If this is a FEATURE REQUEST, please: Describe in detail the feature/behavior/change you'd like to see. In both cases, be ready for followup questions, and please respond in a timely manner. If we can't reproduce a bug or think a feature already exists, we might close your issue. If we're wrong, PLEASE feel free to reopen it and explain why. --Kubernetes version (use ): master Environment: Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others: What happened: Check , it is still using , we should move it to . What you expected to happen: How to reproduce it (as minimally and precisely as possible): Anything else we need to know:\n/assign\nprobably will want to move the init container also to the PodSpec\nI think the init container still depend on annotations , can I specify in ? If so, I want to file another issue to track it\nFYI Init containers have graduated to GA and now appear as a field. The beta annotation value will still be respected and overrides the field value. (,\nSeems like that the doc should be updated. EDIT: NOTE: This feature is beta in Kubernetes 1.5. Once the feature exits beta, Init Containers will be specified in the PodSpec alongside the app containers array.\nThanks I filed an issue here\nWhat about merge this first and then handle init-container in a separate PR?\nSounds good to me, thanks."}
{"_id":"q-en-kubernetes-779e55151b5e50a688dfd48c738601f38ce31e82808856ff9f5b68f750480421","text":"I'm seeing the following in my logs: There was an errors, and we silently skipped it.\nGood catch! Thanks for pointing."}
{"_id":"q-en-kubernetes-7f131da6239da5493e100e7e0db113773d52bc0117230bb023c250a8a0ec7f1d","text":"Is this a BUG REPORT or FEATURE REQUEST? (choose one): BUG OR Do i miss something important? Kubernetes version (use ): k8s 1.3.5. still exists in 1.6.2 What happened: HA of kube-controller-manager and kube-scheduler is achieved by create endpoint and update it periodically. However, endpoint controller deletes endpoints without corresponding services which is implemented in checkLeftoverEndpoints func: That will result in leader transition. What you expected to happen: checkLeftoverEndpoints func shouldn't delete endpoints used by LeaderElection. How to reproduce it (as minimally and precisely as possible): turn on LeaderElection for kube-controller-manager and kube-scheduler. After five minutes, endpoint controller will delete orphan endpoints. Anything else we need to know: we can add some logic in checkLeftoverEndpoints, for example, not add endpoint in kube-system namespace to queue.\nFixed by\nThanks"}
{"_id":"q-en-kubernetes-5ac33f1f8b3ef7796272a787f5cacaa077313bfdf1133b024a8bda9bc74ddec4","text":"Found in I can reproduce locally using\n/assign (via git blame)\ncc\nThanks. I will take a look.\nGot a fix locally. Testing it."}
{"_id":"q-en-kubernetes-7acdc11e66b534bd9b40935c5714b5b662fdc950e2cfd81a3756b323ed4dda59","text":"Is this a BUG REPORT or FEATURE REQUEST? (choose one): bug Kubernetes version (use ): Environment: Cloud provider or hardware configuration: N/A OS (e.g. from /etc/os-release): RHE 7.2 Kernel (e.g. ): What happened: The node resource is empty when start local cluster by . What you expected to happen: There's one node in this local cluster. How to reproduce it (as minimally and precisely as possible):\nJust check the script/code, it seems kubelet will not register until dynamic config is there ( enabled all alpha features by default).\nI met same problem\nWhat do you mean \"dynamic config\" ?\nI think there's a PR that make kubelet can update configuration dynamically, similar to what we had done in scheduler. But did not find PR :(. I'm checking the code.\nShould be this one: /cc /sig node\nthat shouldn't be the case. As long as you can get a client and a nodeName you should be able to register. What specific error are you seeing that suggests it's dynamic config? Also why does local-up-cluster enable all alpha features by default? That seems wrong to me.\nNot sure why, but it really does:\nI think this should be caused by /cc\n, sorry for the disturbing :). Regarding all alpha features, I prefer to enable is by default: local-up-cluster is a kind of test environment for developer, enabling all alpha features maybe helpful to catch up potential feature interactions.\nI would not expect all alpha features on by default in local up cluster\nNeither would I. Developers, by default, shouldn't have to wonder if bugs are their fault or if they come from unstable alpha components. Users, by default, shouldn't have to deal with unstable alpha features either.\n/ , got that; alpha components is unstable, do NOT cover everything for it.\nmet same problem"}
{"_id":"q-en-kubernetes-3fb2673f9405276d46c5aea2eab38aed5fcd9b0373e443dffe78ebfdcafc2637","text":"We have , but there isn't a label in the k/k repo. There is only . Do we need to add a label or rename to ?\nThere are no sig labels on this issue. Please by:<br(1) mentioning a sig: <br(2) specifying the label manually: <br<brNote: method (1) will trigger a notification to the team. You can find the team list .\n/sig ui :smiley:\n/sig contributor-experience\n/cc\nI just it. Fixed?\nYeah. BTW, the page has a bad md format. It also should be fixed.\nFeel free to send a PR to fix :)\nAs Piotr is the owner, I thought let him fix:) If it's Okay, I will do. Thanks Christoph!\nanyone can contribute to the kubernetes/community repo too :) the PR would be reviewed like any other (usually the sig-lead for that sig would review)\nOkay:)\nYep just fix things when you can. That's just faster :)\nFWIW this lead me to open"}
{"_id":"q-en-kubernetes-34b256b676db37865657290319eedf4da1b8e31c1a5d883263d61b5ed853a57a","text":"<!-- Thanks for filing an issue! Before hitting the button, please answer these questions.--Is this a BUG REPORT or FEATURE REQUEST? (choose one): FEATURE REQUEST <!-- If this is a BUG REPORT, please: Fill in as much of the template below as you can. If you leave out information, we can't help you as well. If this is a FEATURE REQUEST, please: Describe in detail the feature/behavior/change you'd like to see. In both cases, be ready for followup questions, and please respond in a timely manner. If we can't reproduce a bug or think a feature already exists, we might close your issue. If we're wrong, PLEASE feel free to reopen it and explain why. --Kubernetes version (use ): Environment: Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others: What happened: the psp admission controller iterates through applicable psps for both the user and the service account. This leads to at least one RBAC DENY message for each pod. What you expected to happen: \"RBAC DENY\" should be a more meaningful message How to reproduce it (as minimally and precisely as possible): enable psp w/rbac\nclosing in favor of"}
{"_id":"q-en-kubernetes-6ff0be7eefdbbc0b6ac86407e7255e47e147dec11f4ff5b19db49c2bce6e676f","text":"5 tests failed, 6 jobs failed, 151 builds failed. Failure stats cover 1 day time range '2 Jun 2017 20:57 UTC' to '3 Jun 2017 20:57 UTC'. <details<summaryRationale for assignments:</summary </details<br\n/unassign I am probably not the right owner for these tests. Leaving this to to triage it.\nThis seems pretty likely a quota issue. Many of them started after the cleanuped the old gce resources. Probably some tests are leaking resources. Though it doesn't occur on other CIs like . I will dig into it a bit more.\nThe root cause is ~2 years old from We are parsing human readable gcloud output, which recently changed. Additionally, when we fail to create the LB we leak the static IP address. It's failing on GKE but not GCE because GKE uses a pre-release version of gcloud as a canary.\n/priority failing-test\n/reopen ci-kubernetes-e2e-gci-gke-slow-release-1.5 and ci-kubernetes-e2e-gke-slow-release-1.5 will continue failing until we cherry pick the fix. Reopen for now.\nForgot to update: ci-kubernetes-e2e-gci-gke-slow starts passing after was merged.\nCan you cherrypick to both 1.5 and 1.6 branch? But this issue is not 1.7 blocking issue anymore.\nSure, working on it.\nSent cherry picks: release-1.6: release-1.5: release-1.4: Cherrypicking the fix into 1.4 as well because we also have gce-slow-1.4 and gci-slow-1.4 CIs running.\nremove the milestone, this is not a 1.7 issue\nSince all issues listed here are addressed separately. Close this one."}
{"_id":"q-en-kubernetes-8417e5f6eb758a5c8cc847b758ffdcecb505fe8a6d8b364608b1d97579f25d03","text":"We should just use GOFLAGS\nThere are no sig labels on this issue. Please by:<br(1) mentioning a sig: <br(2) specifying the label manually: <br<brNote: method (1) will trigger a notification to the team. You can find the team list and label list\n/sig contributor-experience\n/assign"}
{"_id":"q-en-kubernetes-26a9c70011dc98b2696e398193f03214466a9b96b9886638805fb7b17d9dcc4a","text":"There are no sig labels on this issue. Please by:<br(1) mentioning a sig: <br(2) specifying the label manually: <br<brNote: method (1) will trigger a notification to the team. You can find the team list and label list\n/sig azure\nfixed by\n/label official-cve-feed (Related to kubernetes/sig-security)"}
{"_id":"q-en-kubernetes-b04948f9f58296005120291113735f6d1c38b32ccd2f4856f38e2f525ee40a6d","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: BUG REPORT What happened: When using IPv6 addresses, the kubelet passes the address to client-go, which does not correctly format the address for use in API requests. For example: Failed to list v1.Pod: Get https://2001:129::136:6443/api/v1/pods?fieldSelector=spec.nodeName%3Dnode136&resourceVersi on=0: dial tcp: too many colons in address 2001:129::136:6443 What you expected to happen: The URI should have https://[2001:129::136]:6443, so that SplitHostPort() will correctly be able to separate the host and port components. How to reproduce it (as minimally and precisely as possible): Create cluster using IPv6 address (e.g. kubeadm init --apiserver-advertise-address=2001:129::136). When the kubelet starts up, it'll log this error. Anything else we need to know?: This is being handled as part of effort to support IPv6 in Kubernetes. Was found by a co-worker, I'll work on a commit to fix. Environment: Kubernetes version (use ): Using latest, via KubeAdm Cloud provider or hardware configuration*: bare-metal OS (e.g. from /etc/os-release): Ubuntu 16.04 Kernel (e.g. ): 4.4.0-81-generic Install tools: Using KubeAdm Others:\n/area client-libraries /area kubelet /area ipv6\n/assign\nThe point of failure is here: SplitHostPort will accept IPv4 and IPv6 addresses, but the latter must have square brackets around the IPv6 address.\nThere are no sig labels on this issue. Please by:<br(1) mentioning a sig: <bre.g., for API Machinery<br(2) specifying the label manually: <bre.g., for sig/scalability<br<brNote: method (1) will trigger a notification to the team. You can find the team list and label list\n/sig api-machinery because client-go\nShould we open this on the repository?\nI see in the client-go repo read me that it states: Does that confirm that I should do a PR here?\nOne thing that I was wonder, was whether or not the hostname/address provided in the call to getListener, has been validated. If so, and it can only be a hostname, IPv4, or IPv6 address as a string, we could just check if there is a \":\" in the name to determine IPv6. Otherwise, I could use net.ParseIP(), and net.To4() to determine if IPv6. I'll start working with the latter assumption, but please let me know, if the former could be used.\nAlso, I'm making the assumption that we'd want to make client-go be able to handle IPv6 addresses, rather than having each user of client-go (e.g. kubelet) from ensuring that the IP address is of the right form. IOW, have client-go add the needed [] around IPv6 addresses, rather than have kubelet do that. My thought is that there may be cases where we want just the IP address and not the modified IP. Does that sound OK?"}
{"_id":"q-en-kubernetes-9ba2b90e5f5622f8d05046fe0f8eede587bcd980f5c085e586be1054cfd4fc7b","text":"We want kubemark setups to be as similar to real clusters as possible. Since we are launching kube-aggregator in real clusters, we should add a possibility (and probably also enable by default) to have kube-aggregator in front of apiserver(s).\nwas working on being able to use it in e2es.\nWith kube-aggregator integrated into kube-apiserver, is this still relevant?\nThere are a few flags which need to be set up for kubemark to actually be properly testing with the aggregator. I am digging around in that code this week anyway. I will try to get to it and make sure its set up properly.\nThanks for looking. Let me know if I you need something.\n[MILESTONENOTIFIER] Milestone Labels Incomplete Action required: This issue requires label changes. If the required changes are not made within 3 days, the issue will be moved out of the v1.8 milestone. kind: Must specify at most one of ['kind/bug', 'kind/feature', 'kind/cleanup']. priority: Must specify at most one of ['priority/critical-urgent', 'priority/important-soon', 'priority/important-longterm']. <detailsAdditional instructions available <a href=\"\"here</a</details"}
{"_id":"q-en-kubernetes-3d27c8355ff062866383a0c37de9c2cc55380d33007b2b06788398227835c367","text":"It looks like a test configuration failure:\nThere are no sig labels on this issue. Please by:<br(1) mentioning a sig: <bre.g., for API Machinery<br(2) specifying the label manually: <bre.g., for sig/scalability<br<brNote: method (1) will trigger a notification to the team. You can find the team list and label list\n/assign\nlast passing run: https://k8s- first failing run: https://k8s-\nhttps://k8s- /close"}
{"_id":"q-en-kubernetes-d2a7b9b1e60b711cc9721c924351d1e634c168a7c842e6a5fc5c284006c3b651","text":"/kind bug What happened: We constantly log message like at V(2) that don't tell any useful information."}
{"_id":"q-en-kubernetes-394199b6d7a65649fa5b83dce8d6d5481983a166f4878da048a2921030511f37","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --BUG REPORT: What happened: I'm running a simple Kubernetes cluster with master and one node. All basic PODs are deploying successfully (ie. weave-net) but when I try to deploy GlusterFS as a POD - container creation fails with error: This is a result of failing to perform from docker file. The weird part is that if I run this image directly on node through docker it runs smoothly. What you expected to happen: I'd expect to be able to deploy GlusterFS through Kubernetess properly. How to reproduce it (as minimally and precisely as possible): Setup minimal cluster and try to deploy the follwing POD Anything else we need to know?: Environment: Kubernetes version (use ):Cloud provider or hardware configuration**: virtualbox OS (e.g. from /etc/os-release):Kernel (e.g. ):Docker versionInstall tools: vagrant Others:\nThere are no sig labels on this issue. Please by: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. You can find the group list and label list . The in the method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals\n/sig node\nReiterating the mentions to trigger a notification: <detailsIn response to : /sig node Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nOne update - I can deploy this POD with no problems on version of k8s.\nI am hitting the same bug on a freshly installed k8s cluster running on installed via kubeadm plus weave-net with a currently unsupported docker version. For me, downgrading to a supported docker version helped to resolve this issue (v1.12):\n- Thanks for the info! - Well we can confirm that kubernetes is not working properly with Docker 1.29 and 1.3\nReiterating the mentions to trigger a notification: <detailsIn response to : - Thanks for the info! - Well we can confirm that kubernetes is not working properly with Docker 1.29 and 1.3 Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThis is not really a bug since kubernetes doesn't support the docker version you use yet, but feel free to send a patch if you find the problem.\nI was thinking about your comment and I think it is still a bug since downgrading Kubernetes to 1.6 fixes the problem while having the same setup as described above. So something have changed in 1.7. But yeah. Will try to investigate on my own.\nI'm actually getting this same error with 1.7.1 and docker 1.13.1. Anyone have other thoughts on what the issue might be here? I found this so question () but it looks like everything is configured correctly.\nIn kubernetes 1.7, under pid # 1, there is a \"/ pause\" process, and since init takes pid 1, it is no longer available. In earlier versions of kubernetes, pid # 1 was free . PID USER TIME COMMAND 1 root 0:00 /pause 7 root 3:41 /tiller Kubernetes now shares a single PID namespace among all containers in a pod when running with docker = 1.13.1. This means processes can now signal processes in other containers in a pod, but it also means that the pattern will cause the pod to be restarted rather than a single container.\nI can work around the problem using the option in the kubelet, however this not the desired solution of course.\njust FYI, looks like some images have to be run as PID 1.\nIn 1.7 release, we enabled shared pid namespace if the docker version is docker 1.13 and beyond for debug pod feature. Do anyone know why those images, such as GlusterFS have to run as PID 1? we need to figure out how to resolve this conflict with shared pid namespace. If we cannot, we have to disable that setting for 1.8 since we scheduled to support docker 1.13 in 1.8 timeframe. I am currently assigning you this bug to come up a solution.\ncc/\ngluster/gluster-centos is a container image that runs systemd so that it can run ntpd, crond, gssproxy, glusterd & sshd inside a single container. This isn't how Kubernetes is intended to be used and I don't know how much we should go out of our way to support it (this is a question for & seems to be a good write up of the challenges of running systemd inside of docker. It also prevents Kubernetes from doing process management like: reporting state of processes signaling individual processes for graceful restart ingesting and display daemon output (to stdout/err) Ideally we could change this many-processes-in-a-container to be many-containers-in-a-pod, something like this: This doesn't need a privileged container and doesn't run sshd or ntpd, neither of which are needed in Kubernetes. This also doesn't run crond, but that may actually be needed by gluster. I've never used gluster. (I used flags so I didn't have to investigate how to make these processes log to stdout) If you must run systemd in a container, adding this to the container config will get systemd to run: This tells systemd to run in system mode even though it's not PID 1 and disables its chroot detection (which is unrelated to PID 1 but checked when invoked as \"systemd\", I guess).\nafaict, running in a container is well recommended ( and ) at certain scenarios and a common practice. iic, there are enough containers who make use of as its . If shared namespace which introduced in 1.7 breaks this, I believe its a backward compatibility break :(. I would invite here to share his thoughts as well. :)\nThose articles are about how to work around shortcomings in vanilla docker that Kubernetes solves natively: resource sharing via pods, containing via SELinux, reaping orphaned zombies, etc. My earlier question was along the lines of how much effort should go into supporting use cases that are not using Kubernetes as intended. (I'm actually asking, not trying to make a point.) Does Kubernetes commit to continuing support for every container image that's ever run on Kubernetes? For how long? Indefinitely? or can we change it over the course of multiple releases? Anyway, this change is compatible with running an init system in a container. The problem is with images that assume they will only ever have , which has so far been the case with docker due to a technical limitation, but it was never the case with other runtimes such as rkt, hyperv, etc. Brainstorming other options: shared pid namespace and require users of systemd to set an environment variable in their Kubernetes config (or port the multi-process container to a multi-container pod). shared pid namespace and enable it in a future release where we have the same problem. a solution where pid namespace is configurable per-pod, but not supported indefinitely. It would probably be implemented using annotations, and we will have to modify the Container Runtime Interface to support it. Over the course of multiple releases the default behavior would change but could be reverted based on the presence of an annotation. In a future release, users will have to change an annotation in their config to get the old behavior (rather than the environment variable). In a future future release, the annotation will stop working and users will have to set the environment variable. to supporting per-pod shared vs isolated pid namespaces indefinitely and adding it to the Kubernetes API. It doesn't make sense in some runtimes, so we probably can't require all runtimes to implement it, so it'll likely only see use with Docker. Last time we discussed it in SIG Node, and didn't want to support isolated pid namespaces indefinitely.\nYes running systemd in a container is a critical need for us. There are lots of situations where it is and will be used. Biggest use case if for people simply moving current workloads that run in a VM into a container. I have a service that runs fun under systemd in a VM. If I can simply take the init script and move it whole sale into a container, why wouldn't I want to do this. It's simple and does require a great deal of knowledge by the end user. service containers. Not every project can easily be moved into microservices. But they probably can run inside of containers. If we drop support for systemd running inside of containers, then we loose a big group of users. Getting access to journald information inside of the container. Applications that write to syslog and directly to the journal. Currently in container world without systemd, this information is dropped on the floor. if I run systemd and journald, I can actually retrieve and use this information.\nIf the intension is to force all containers inside of a POD to run with the same PID Namespace? I think this is a mistake since it eliminates another really cool use case. I would love to be able to run two containers inside of a pod, and INIT container with escallated privileges and a locked down container, with no privileges. I want to make sure the locked down container can not see the processes of the privileged container. An example of this, might be a container that loads a kernel module that is needed by the locked down container. Another use case would be a buildah container. The INIT container mounts up a COW file system onto a directory on a location that the locked down container can write, The locked down container writes out the data for his image, when the container completes, the INIT container commits the data to the repo. I believe allowing different containers inside of a POD to run with different security and isolated PIDs should not be removed.\nnone of the things you've mentioned are removed by this change, save using docker to hide other processes in the pod, and I'm not sure why that's a requirement. Is anyone currently using the \"privileged sidecar\" pattern you describe or is it hypothetical? It just occurred to me that an interesting compromise here might be to only enable shared pid for multi-container pods. This would enable pod semantics for PID when it would make a difference while drop-in docker monoliths keep the PID 1 behavior they're expecting in docker. I'd have to double check, but I'm pretty sure that behavior could be implemented entirely in the dockershim with no API or CRI changes.\nHypothetical, but I am pushing it for a solution for problems I hear from users fairly frequently, where then need big privileges at start and stop of the pod, but not for the time is is available to the network.\nok, thanks. ftr, I think we should continue to support this pattern as well, and I don't think a shared PID namespace prohibits it. I will bring up this issue at Tuesday's sig node meetup if you'd like to join us. Does the config I mentioned in fix the issue with systemd?\nCopying from a comment made elsewhere:\nUpdate from SIG Node Meetup: Everyone wants to support these use cases, but we're not yet sure the requirements we want to support medium to long term. is going to sync up with and then we'll figure out what we want to support in a future release. For 1.8 the current plan is to leave docker shared pid namespace enabled by default. Cluster administrators can disable it per-node with the --docker-disable-shared-pid flag. I'll see about whether we should add something back to the release notes.\nThis seems somewhat backwards. Wouldn't it make more sense to go back to supporting systemd use cases by default NOW (by making the shared PID namespace behavior optional or something) until a final decision is reached and can be advertised to current users?\nabsolutely ! thank you very much\nI think that is the best solution is to make the PID Namespace sharing optional.\n- optional at what scope - cluster, node, pod. let's sync up before next sig-node and we can revisit if need be before 1.8 goes out. The main question right now is if pid namespace sharing is opt-in or opt-out. We need to weigh pro/cons of both.\nI would expect it at the POD, since pods would have different requirements, not sure why you would want to have this at the cluster or node level.\nyes it works :)\nawesome, thanks! I'd like to better understand the requirements of people objecting to this change ( I think there's a bit of confusion about what has changed, so let's start with some facts: This container image runs on Kubernetes without changes to the image, but it requires (non-obvious) pod configuration. is nothing inherently incompatible between a shared pid namespace and a process manager like systemd. systemd and other init systems work with pid != 1, but may require special invocation. in this issue is specific to Kubernetes. That is, this container image wouldn't work with vanilla docker PID namespace sharing, either. change does not affect securityContext I can think of a few things that are objectionable about the change, but let's state them as requirements. For example, are these requirements for anyone? Kubernetes must run all containers published on docker hub with no additional pod configuration Kubernetes must run all systemd containers, regardless of how they are built, with zero additional pod configuration. Kubernetes must run this specific container image with zero additional configuration Kubernetes must officially support isolated PID namespaces between containers in a single pod across all runtimes (docker, rkt, hyperv, etc). Configuration changes are ok, but the flag to systemd cannot be used because the documentation says its not useful. I want to use systemd as intended. Rather than having Kubernetes work around this container image, should we patch the Dockerfile to make it compatible with even the native docker shared-pid, unrelated to Kubernetes?\nConfiguration changes are ok, but the --system flag to systemd cannot be used because the documentation says its not useful. I want to use systemd as intended. We can ask the systemd guys why they recommend never using the --system flag, but I would guess one of the reasons is that processes showing up that were not created in the ancestry of systemd, could cause confusion. I would also guess that running more then one container with systemd --system would not be supported. Kubernetes must officially support isolated PID namespaces between containers in a single pod. This is required so that you can run different containers with different security constraints in the same pod at the same time. Your last comment makes the seem like it is the default, when actually it is the exception. almost no containers run with shared-pid normally. I guess in a side car container, it makes sense to support this scenario, but forcing this as the ONLY way to run containers is limiting.\nThis is the part I don't understand. Why? What does PID visibility have to do with security constraints? What am I missing? nor do they run with a shared network namespace, but we impose that on docker containers because that's the pod model.\nAn update from the last SIG Node: we'll disable this by default in 1.8 and revisit for 1.9 I'll use to track that.\nif I have a privileged container process in the same pid namespace as a pid in a non priv container, then the container process in the non priv can attack the priv process. Having them in separate pid namespaces gives you better security. Bottom line, you have a potential change where containers that expect to be run as PID 1, whether legitimately or not, could break when you make this change. Kubernetes in my opinion should make it easy to opt out of this behavior.\n/priority critical-urgent\n[MILESTONENOTIFIER] Milestone Labels Complete Issue label settings: sig/node: Issue will be escalated to these SIGs if needed. priority/critical-urgent: Never automatically move out of a release milestone; continually escalate to contributor and SIG through all available channels. kind/bug: Fixes a bug discovered during the current release. <detailsAdditional instructions available <a href=\"\"here</a</details\nUpdate: is pending review and should be to the v1.8 Milestone to resolve this issue."}
{"_id":"q-en-kubernetes-7164ee736ec8b7198f38602535de086e7396a736ea12447e5be8a654710fb76e","text":"Is this a BUG REPORT or FEATURE REQUEST?: /kind cleanup What happened: It seems kubelet did not update inode pressure; and only test codes are using . So prefer to deprecate or remove it directly. (seems no node-misc group :). )"}
{"_id":"q-en-kubernetes-d8388b2cd8a7df24a3afbbffea88f41963538eebab0200d554d2fd0c161e5075","text":"Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: I have a transient local disk mounted under /var/lib/docker in order to take advantage of faster IO on Azure VMs. This means that all that data is lost when a reboot occurs, but the rest of the data outside that mountpoint persists. After the reboot, the existing pods are relaunched, and kubelet is unable to create the log file symlinks because they already exist, but they are pointing to an invalid place, that container no longer exists. Here are the relevant log lines: The old symlinks are there: What you expected to happen: The correct symlinks should overwrite the old ones. How to reproduce it (as minimally and precisely as possible): Erase /var/lib/docker between reboots with existing pods. Anything else we need to know?: No. Environment: Kubernetes version (use ): Server Version: version.Info{Major:\"1\", Minor:\"7\", GitVersion:\"v1.7.0+coreos.0\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2017-06-30T17:46:00Z\", GoVersion:\"go1.8.3\", Compiler:\"gc\", Platform:\"linux/amd64\"} Cloud provider or hardware configuration**: Azure VMs OS (e.g. from /etc/os-release): Container Linux by CoreOS 1409.6.0 (Ladybug) Kernel (e.g. ): Linux node-2-vm 4.11.9-coreos SMP Thu Jul 6 00:54:42 UTC 2017 x86_64 Intel(R) Xeon(R) CPU E5-2673 v3 @ 2.40GHz GenuineIntel GNU/Linux Install tools:\nThere are no sig labels on this issue. Please by: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. You can find the group list and label list . The in the method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals\nI think this falls under\nReiterating the mentions to trigger a notification: <detailsIn response to : I think this falls under Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-30fc46d5f5ed507f5f4f6a5850c6ff6c915dfceb9d7e2e4288862de4a3e9b98b","text":"/kind bug /sig storage\n/assign\nIt also fails on debian GKE. However, they all pass on debian and GCI GCE. Error msg is mounting timed out. This leads me to suspect that there is some difference in networking for GKE vs GCE.\nI do strongly suspect there is some networking issue on GKE. Many other tests are also failing with \"connection timed out\" errors. 1.7 GKE tests are looking healthy, so it seems like something is broken on master. /sig network\nFixed by /close"}
{"_id":"q-en-kubernetes-bfe758b5fa5482a4ce9e81f7cef2426feb82d8e03468cce3160ef8e0ed9e9bda","text":"What happened: Wrong HPA condition - ScalingLimited have one rc with 1 replica(current have 1 pod) have one hpa with spec.minReplicas=1 and spec.maxReplicas=10 HPA condition ScalingLimited is true and described by below, I expect ScalingLimited=false at that time. What you expected to happen: I expect ScalingLimited=false at above time How to reproduce it (as minimally and precisely as possible): kubernetes with DNS api server with --runtime-config=\"api/all=true\" binaries and images, deploy heapster, influxdb and grafana in kubernetes a rc a v2 hpa hpa conditions by I expect ScalingLimited of hpa conditions should be: Addition info:\nThere are no sig labels on this issue. Please by: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. You can find the group list and label list . The in the method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals\n/sig scalability\ncould you please take a look.\n/sig autoscaling\nIt looks, at first glance, like what's happening is that since CPU usage is zero, we're calculating a desired replica count of 0, which is less that the minimum (1), so we're getting . We might want to detect this condition and provide another status message that's clearer. does that make sense on your end? Does the problem persist if CPU usage is non-zero?\nYes, the problem only persist if CPU usage is zero, and not exist when CPU usage is non-zero. I will keep this issue open if you want to improve and let it more clearer while CPU usage is zero. Thank you.\nIf this is still an open issue, I'd love to take a shot at submitting a pr for it. To confirm, the desired behavior is that if CPU usage is zero, and the desired replica count is 0, which is less than the minimum (1), should still be , but the status message should make it clear that it is because the CPU usage is zero?\nI'd be happy to review such a PR :-). Yep! Having zero usage is a bit of a special case, so it's probably worth letting people know explicitly.\nWonderful! Working on it :)"}
{"_id":"q-en-kubernetes-78c928357d06678fb21dafd04c121deafd7665579e2f7914fc7c347586bc211a","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?:BUG REPORT What happened: the err message in addNamespaceRoleBinding is wrong. ref to\nThere are no sig labels on this issue. Please by: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. You can find the group list and label list . The in the method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals\n/sig auth"}
{"_id":"q-en-kubernetes-c0c4fc1d4121d996f09cdcd0bc1a97d845d1ba23ce6e22136fc9102c88ddcc0a","text":"/kind feature What happened: In playing with the nginx ingress controller deployment, I noticed the following pod name: nginx-ingress-controller--c0ck5 I figure we should possibly put something in place to prevent accidental l33t profanity ;-). What you expected to happen: The following pod name: nginx-ingress-controller--r0ck5\nThere are no sig labels on this issue. Please by: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. You can find the group list and label list . The in the method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals\nMy first github (k8s) issue. I've no idea which SIG would cover this topic. /sig product-management\n/sig api-machinery\nPreviously we removed vowels in . We can 3xt3nd the idea with 0, 1 (can stand for ) and 3. would you like to send a PR? If not, I can work on this.\n/assign jamesawebb1\nReassign to me if you do not want to contribute a PR."}
{"_id":"q-en-kubernetes-a66d8eeb0df3cbbbf4d33eacafb5e7d2a93904948251230882d251407b7ef6c3","text":"Starting a large data set (32gb allocated, 4gb resident) in the api server causes about 7.5gb allocations of . The tree for that shows: This appears to be due to openapi aggregation. If there is an easy fix in 1.7 it might be worth it. Otherwise this should be addressed for 1.8.\nThere are no sig labels on this issue. Please by: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. You can find the group list and label list . The in the method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals\nThis is all API server startup - for small clusters it's about 20-60% of startup time\nMost of those are that is the library we used to convert our spec to proto. The key is this line: Do you have any idea how we can get a better space time/complexity here?\nI have a PR for that now.\nFixed in\nEven without aggregation, starting and serving OpenAPI docs was a huge chunk of startup time. The linked change makes it disappear.\nThis should not be related to aggregation at all (as far as I can tell from that tree). Thanks for the fix\nThanks, it's in gnostic master now."}
{"_id":"q-en-kubernetes-430de08fea1dc48de9bac70ba58faceeda34b765d7b2c3b7e45992fccc2d9329","text":"Both https://k8s- and https://k8s- started failing recently. Their 1.7 counterparts are passing. The difference between the last passing and the first failing tests is: Probably something in broke the tests. /sig scheduling /cc /kind bug\nIt looks like there isn't enough cpu on the master to run the cos-nvidia-installer daemonset. Did we add any addons recently?\nMy previous comment is a red herring. It is an issue with . Ill post a fix soon. From the scheduler logs on the master:\nthat's a benign error. The daemonset is not expected to run on the master. But there is no easy way to express that request in our APIs.\nYes, the pod not scheduling on the master is benign. The actual error is that we do not report node condition , so it is left as . The fix is to report that is false.\nThe tests are still failing. /reopen\n/assign\nI will be updating to track progress.\ntests are green again! /close"}
{"_id":"q-en-kubernetes-3871d24647a528f2ff7df25634241fd3cde165662602f0e95a0cba53fe987aee","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: /kind bug /priority failing-test /area test-infra What happened: job is failing. See https://k8s- https://k8s- https://k8s- https://k8s- What you expected to happen: test should pass\nThere are no sig labels on this issue. Please by: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. You can find the group list and label list . The in the method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals\n/assign\nReiterating the mentions to trigger a notification: <detailsIn response to : /assign Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nI'm guessing that something got messed up in the bazel cache when reverting from 0.5.2 to 0.4.5. I'm not sure how best to fix.\nHow about updating everything to bazel 0.5.3 at once? I know you're trying to do a step-by-step approach to avoid breakages, but I suspect it will be cheaper to just deal with a brief outage while we update? (aka update the bazelbuild image to 0.5.3 and then merge your PR with updated and working BUILD rules?)\nthere's at least one other issue I've discovered with the latest go rules - the binary no longer works, due to changes introduced in that might be arguably better (at least everything still builds?), though it's frustrating that it makes things more broken in a different way."}
{"_id":"q-en-kubernetes-942e344e7d10457a3ad353e06b5c75edb64a568429e6b48a85d1142320135dd2","text":"All controllers that use the field have a default value of 10 for the field, except for , which has a default of 2. We should increase it to 10 for consistency on its default value across controllers. /sig apps"}
{"_id":"q-en-kubernetes-ff5be5ccb557f1d3a285ef8b9f8f41b09089fd4befe8fff0c73fd4bc43e7c68a","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: What happened: I was attempting to filter pods with a using client-go selecting by its . I tried multiple variations until I ended up with . When filtering by , list of pod(s) came back empty even if pod with the IP we are filtering by was running. Then I tried reproducing it by calling the Kubernetes API directly which had the same result, list of pods came back empty: What you expected to happen: When passing the IP of the previously retrieved running Pod, I expected to get that Pod returned in the result. How to reproduce it (as minimally and precisely as possible): create a simple , get its IP (For example using - query Kubernetes: Anything else we need to know?: If I leave out the pod IP, I get all the pods back. I also tried using different formats when doing the curl, not one of them return anything. Also it seems like field is . As a side note I retrieved the URL which I used from debugging which calls client-go made. Environment: Kubernetes version (use ):Cloud provider or hardware configuration**: minikube version: v0.20.0 OS (e.g. from /etc/os-release): minikube OS Kernel (e.g. ): minikube kernel: Install tools: minikube Others:\nThere are no sig labels on this issue. Please by: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. You can find the group list and label list . The in the method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals\n/sig api-machinery\nReiterating the mentions to trigger a notification: <detailsIn response to : Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nCurrently the for only support , , , and . I've already submitted a PR in . PTAL.\ncc"}
{"_id":"q-en-kubernetes-345619cb817e392d4d228d93af69688b454fb065915235d32fb0ccf01aaf36c6","text":"Is this a BUG REPORT or FEATURE REQUEST?: /kind cleanup /sig scheduling Descriptions: According to the discussion at , we'll move the node condition filter into a new predicates; and use feature gate to turn-off/on it. /cc\n/assign"}
{"_id":"q-en-kubernetes-6edf6099c1735f40a076054f74d372e76a0d81ab787b91180807cba35b846d66","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: What happened: What you expected to happen: How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration**: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others:\nThere are no sig labels on this issue. Please by: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. You can find the group list and label list . The in the method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals"}
{"_id":"q-en-kubernetes-cc6c0aaf84d61de20e8a0c1405be5f1ae35cc0d47347eacead0929aba9488d5d","text":"There are two issues: doesn't have merge key The validation code regards it as an error if both Initializers.Result and Initializers.Pending are nil. So if multiple initializer controllers sending patches removing themselves from Initializers.Pending at the same time, the last patch that gets applied will get a validation error.\n/sig api-machinery"}
{"_id":"q-en-kubernetes-f6efad6b91c3c3f7311d37a352e2885c03d3d05bf76a7dfad0b041689ecc548f","text":"Seen in https://k8s- Aug 26 09:35:45.074: Couldn't delete ns: \"e2e-tests-containers-9cscj\": an error on the server (\"missing route (service \"sample-api\" not found)\") has prevented the request from succeeding (get ) (&errors.StatusError{ErrStatus:v1.Status{TypeMeta:v1.TypeMeta{Kind:\"\", APIVersion:\"\"}, ListMeta:v1.ListMeta{SelfLink:\"\", ResourceVersion:\"\"}, Status:\"Failure\", Message:\"an error on the server (\"missing route (service \"sample-api\" not found)\") has prevented the request from succeeding (get )\", Reason:\"InternalError\", Details:(*v1.StatusDetails)(0xc4217a4900), Code:500}})\nhttps://k8s-"}
{"_id":"q-en-kubernetes-8907620e2c573189a89594b3f74b23d33c1aec70e97ae75110db1d075a2d466c","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: unused images does not get deleted What you expected to happen: unused images get deleted How to reproduce it (as minimally and precisely as possible): configure kubelet with these args: Anything else we need to know?: when using above config, the node's events history showing: If I change to it works, node's events history showing these: The difference is there is an extra event, which is not seen with cofnig. Environment: Kubernetes version (use ): Client Version: version.Info{Major:\"1\", Minor:\"6\", GitVersion:\"v1.6.4\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2017-05-19T18:44:27Z\", GoVersion:\"go1.7.5\", Compiler:\"gc\", Platform:\"linux/amd64\"} Server Version: version.Info{Major:\"1\", Minor:\"6\", GitVersion:\"v1.6.4\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2017-05-19T18:33:17Z\", GoVersion:\"go1.7.5\", Compiler:\"gc\", Platform:\"linux/amd64\"} OS (e.g. from /etc/os-release): ubuntu 16.04 Kernel (e.g. ): Linux 192-168-1-1 4.4.0-63-generic -Ubuntu SMP Wed Feb 1 17:20:32 UTC 2017 x8664 x8664 x86_64 GNU/Linux Install tools: kubeadm\nThere are no sig labels on this issue. Please by: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. You can find the group list and label list . The in the method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals\nReiterating the mentions to trigger a notification: <detailsIn response to : Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig area/kubelet\nMaybe we are with the same problem. Kubelet Parameters for tests: When we reach a disk consumption of 20% (for debug) - after 2 minutes, the deletion of logs, image docker, and then PODS should begin. But nothing happens. The debug logs kubelet saw the following: Problems are observed in the versions 1.6.X, 1.7.X In version 1.5.7, it looks like this: Can you help us? Or we help you, with Debug? Environment: Kubernetes version (use kubectl version): OS (e.g. from /etc/os-release): ubuntu 16.04.2 LTS Kernel (e.g. uname -a): Linux node2 4.4.0-62-generic -Ubuntu SMP Wed Jan 18 14:10:15 UTC 2017 x8664 x8664 x8664 GNU/Linux Docker: Docker version 1.11.2, build b9f10c9_ PS We tried the rollback option from , but it did not change anything.\nhello! Can you spare a little time to this problem?\nWell, here is what we have found so far. In our setup, eviction.managerImpl.synchronize() frequently runs with a set of signalObservation objects, which contains the same data as in previous runs, even for \"time\" field. Probably this is due to slow cadvisor. Unfortunately, eviction.thresholdsUpdatedStats() considers such data expired and removes it from the list of observed thresholds. This was always the case, but before commit \"time\" field wasn't initialized, and thresholdsUpdatedStats allowed it to persists between iterations of synchronize() by explicit check for After neither check for zero, no check for one time being more recent then another, doesn't work as iterations are frequently performed with the same summary.Node.Fs object. And here we are with blinking \"node conditions - observed: [DiskPressure]\" in logs: one run of synchronize sets it, and next one discards it (runs are delimited with \"synchronize housekeeping\" message): The question is whether this logic in thresholdsUpdatedStats is correct now or not. Should garbage collection work even with slow cadvisor (if this is the reason of slow updates of summary.Node.Fs)? Probably frequency of synchronize should be adjusted, or summary.Node.Fs updates should be synchronized with execution of synchronize()? We can correct the problem by monkey-patching makeSignalObservations and get rid of summary.Node.Fs.Time usage in signalObservation, but is it right?\nThe 2 minute grace period has not been met, so no action is taken. This is the difference between a \"soft\" and \"hard\" threshold: The soft only takes action after the threshold has been met for the specified period of time. If no new stats exist, the correct behavior is not to take action. This could delay action by 10 seconds, but it is better not to be overactive when evicting (we have had this problem in the past). The eviction manager assumes that disk pressure still exists when it does not have new stats. It basically just \"skips\" that synchronize call when that is the case.\ndekhtyarev@ it looks like your node goes in and out of disk pressure. I see messages woven into your logs.\nHmmm, or maybe there is something here. Looks like the timer is reset for soft eviction thresholds.\nNo, On seeing this, I just ran several PODs on the node and did to achieve the conditions of eviction. I used ~22% of the disk, at the threshold of 80%. And disk consumption on the node did not change.\nWe just think that the timer resets, as the timestamp of the received data has not changed and this is taken into account in the eviction manager logic. Look at the field here: wrote about it, I think.\nI see the issue. Ill try and get a bugfix into 1.8.\ncc Soft Eviction thresholds have not been functional since 1.6...\n[MILESTONENOTIFIER] Milestone Labels Complete Issue label settings: sig/node: Issue will be escalated to these SIGs if needed. priority/critical-urgent: Never automatically move out of a release milestone; continually escalate to contributor and SIG through all available channels. kind/bug: Fixes a bug discovered during the current release. <detailsAdditional instructions available <a href=\"\"here</aThe commands available for adding these labels are documented <a href=\"\"here</a</details"}
{"_id":"q-en-kubernetes-1dee39133164d7ceee5ef0ba5dabca86bb349b1e45efeb86c94adad0f91efe93","text":"We should solve this problem in a better way\ncc :))\n/sig contributor-experience release\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. Prevent issues from auto-closing with an comment. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale /lifecycle frozen\n/assign"}
{"_id":"q-en-kubernetes-cb33721bac86ba994333956f885b04e90ecb2ced92e2ab07428e50bba20583b2","text":"https://k8s- https://k8s- cc: /kind bug /priority failing-test\nThanks a lot for reporting this! I think I have a fix PR for this issue. Or should we roll PR back?\nI've kicked off testing of your PR. If it fixes the cross build, then that may be the easier path than reverting. cc: (sig-node leads)\nI also fixed it in since it also touches the imports in that file. Would love to not rebase it yet again for this.\nThat call would be up to or they would have to determine which approach they prefer (fix on it's own in a small PR, or fix as part of a larger PR to prevent rebase).\nSmaller is probably better; that way if the combined one needs to be reverted it doesn't re-break cross as a result. Will rebase if necessary (will depend on queue order.)"}
{"_id":"q-en-kubernetes-4c5d8eaa47b7287d37f05b07935b01b96a7c286b79b1c7db26c662588981eb82","text":"A couple of gce-scale-performance runs failed due to failing to delete the network, which in turn was due to a couple of leaked firewall-rules (I manually deleted the network later to fix it): Ref: https://k8s- cc (assigning to you, feel free to reassign as apt)\nI digged up a bit and seems like the firewall is created by the service-controller as part of , which should've been deleted in but wasn't. Feel free to close the bug if you think it's just a one-off failure.\nTook a quick look, both and runs have these firewall rules initially but doesn't, which indicates something happen in the run. But oddly there is no build log, logs in artifacts are empty as well. Likely that run is badly broken so cluster got terminated forcibly in the middle, and resource are leaked. Will observe this suite a bit longer. /assign /unassign\nI think it wasn't caused by this job. We're running this one () and on alternating days in the project. And they both create clusters (network, etc) with the same name. And I think of gce-scale-correctness leaked the firewall (and hence the network) as it runs some correctness tests around load-balancer. That likely caused this job to fail.\nRef The resource leak was likely caused by test timeout --- GCE operations took too long to finish and the cluster was terminated ahead. Because heathcheck and firewall cleanup in gce controller are coupled with targetpool deletion, deleting cluster while targetpool deletion is pending will leak the resources.\nFyi - this affected gce-scale-performance again (ref: run 35 of https://k8s-): Is there a way to make sure they get deleted even in case of a timeout? I have a dirty hack in my mind - of deleting those as part of kube-down. Maybe you have some better suggestions?\nConsider the frequency this happens, I think it makes sense to implement some cleanup logic within the e2e test. Will give it a try.\nTurned out we already have such cleanup logic, it is just we use an incorrect firewall name and forget to cleanup node health check resources as well. Sent to fix.\nThanks for fixing!\nSure thing, thanks for reporting!\nReopening as it seems like the issue is not fixed yet. For e.g run-16 of https://k8s- failed due to this: This is also affecting our job, both of them being release-blocking jobs. Could you PTAL?\nBy any chance we could make and use different names so that it will be easier to target the leaking point? Also, in that way one leaking job will not affect the other. Or maybe we could also make different runs use different network names? The scalability job is expensive so it might be great if newer runs wouldn't be affected by the leaked resources from previous run. I haven't figured out how exactly and when the resource are leaked, but given that the operation cycle of service controller and e2e test process proceed in an asynchronous manner (plus always fail ESIPP tests ), there might be chances we will leak resource even after we put cleanup logic in e2e test. Will take a deeper look later. Also what you suggested before --- deleting those as part of kube-down --- may also be a viable solution.\nI thought about this option, but it's better for them to share the same name for network and other resources (instance-prefix, etc) because that way we ensure that if one of them leaks some resources, the other will delete them during it's startup. Else we'll run out of quota in the scale-testing project (for e.g wrt routes or VMs) by trying to start another 5k-node cluster while the previous one still exists. I went to pantheon and deleted the network and that went through successfully (though fails). Wonder if the delete command in kube-down can be modified to delete all sub-resources of the network (like firewalls, etc) before deleting it?\nSounds reasonable to me. We may need yet another env to trigger it? As deleting everything correlated to that network sounds a bit dangerous.\nWell, took another look, seems like the actual thing that prevented us from deleting the network was mostly firewall rules, and deleting those sounds safe enough if we are going to delete the network anyway. Will try to make a PR for this.\nThanks for fixing the issue with firewalls. However seems like the latest run of gce-scale-correctness still leaked the network (but doesn't seem due to firewall this time): ref: https://k8s-\nI thought we're deleting the subnets in this case too when we're using custom subnets mode.\nYep, we are supposed to. I think the scale CI hit the same issue as described. I've sent a PR for fixing it\nFor the time-being I manually deleted the subnets and then the network to unblock the next scale job run. But it would be good to have this fixed soon. cc\nClosing this issue as we're now deleting all remaining firewall rules in the network before deleting the network itself in kube-down. For other networking resources in general, we have issue ."}
{"_id":"q-en-kubernetes-b07ae098223982b5794a3258109fbc2341630ece5d9605db5a254ad0905722a6","text":"/kind bug /sig node When the is set to and a container that was granted exclusive CPUs terminates with a normal exit status, the CPU Manager never receives a call. Kuberuntime should be updated to invoke the container manager's hook on detecting normal termination for these containers. This method is already idempotent since it is currently invoked both from kuberuntime's and (the latter of which is invoked by GC.) How to reproduce it (as minimally and precisely as possible): Start singleton cluster with static CPU manager policy enabled. Start a guaranteed pod with a single container that consumes 1 CPU (to get an exclusive allocation) with restartPolicy never that simply exits . Try to start a similar container that won't fit on the node unless the first container's CPU is reaped. Observe the second container fails with insufficient CPUs, and this problem does not resolve even after an extended period of time. Environment: Kubernetes version (use ): v1.8 beta Cloud provider or hardware configuration**: all OS (e.g. from /etc/os-release): all cc\n/assign\nYikes, that's a fun issue. I guess we need to listen for PLEG events?\nThere's a spot in the kuberuntime sync where we think we can make the call.\nwell, i knew i would miss something ;-) good catch.\nAll thanks to workload perf testing\nBy the way, where is of CPU manager called if is set to or ? invoked by GC?\nHi, Besides, I wonder do we have a mechanism to recover things in state of CPU manager when kubelet restarts? Here I have a trial, seems it doesn't work as expected, that is: a few logs before and after the action of addcontainer and removecontainer in policy, to see the content of state: a pod that declares an exclusive CPU, and saw logs: kubelet the pod, to compare the state with that , I think if recovered, the content of them should be the same. but saw logs: Seems that the content in state did not recovered, this way, the CPU manager would behavior far from what we expected if kubelet restart. Did I miss something important?\nis called as a part of internal container lifecycle pre-stop and post-stop hooks (see and ). Regarding Kubelet restarts, it is a known issue and captured in\npriority/important-soon\n[MILESTONENOTIFIER] Milestone Labels Complete Issue label settings: sig/node: Issue will be escalated to these SIGs if needed. priority/important-soon: Escalate to the issue owners and SIG owner; move out of milestone after several unsuccessful escalation attempts. kind/bug: Fixes a bug discovered during the current release. <detailsAdditional instructions available <a href=\"\"here</aThe commands available for adding these labels are documented <a href=\"\"here</a</details\nWell, I understand business in , what I wonder is upper behavor, that is, the behavior of in containerRuntime, When a container with restartStrategy exits, It seems that the runtime try to restart new container directly, without calling things like . So I failed to see the place where remove related hooks called in this case, so here I have a trial: a pod whose container is in restartStrategy of and would exit after a perid exited after a peiod: try to restart it but failed: restart after quite a preiod: So I think the resource is not recycled synchronously, and the remove related hook is called in GC. Am I right? Does it works in the way we expected?\nYes, that's exactly correct. The linked PR with the fix for this issue also helps the situation you described to converge faster."}
{"_id":"q-en-kubernetes-c6ccdba508f3c0b9c62484a3a0cca04ddbf553af79de181ac9ac96008383fbfa","text":"Follow-up of Docker 1.13 introduced a change into the logging mechanism (), that split the log lines longer than 16K into chunks of no more than that. Current default logging integrations (fluentd-gcp and fluentd-es addons) do not support that and this might result in broken ingestion, i.e. when JSON parsing is broken for entries longer than 16K. One possible solution is introducing a fluentd plugin to the configuration to concat such entries in the similar way does it (assuming there is one out there, cc for that) /cc\nPresumably this should be done in the Docker fluentd logging driver... The discussion on seems to be leaning that way...\nIn fluentd logging driver side, journald driver approach is better: I'm not sure buffering partial message in logging driver is good or not... In fluentd side, implementing or is needed.\ndo we need to block 1.8 release? can we please move this out of 1.8 if it's not?\nMoved to 1.9, since it's not specific for 1.8\nIs there any way this can be fixed by running a custom fluentd? Or is there a fluentd docker driver github issue I can follow for updates on when they expect to have this fixed? Txh\n[MILESTONENOTIFIER] Milestone Issue Needs Approval Action required: This issue must have the label applied by a SIG maintainer. <details<summaryIssue Labels</summary- : Issue will be escalated to these SIGs if needed. : Never automatically move out of a release milestone; continually escalate to contributor and SIG through all available channels. : Fixes a bug discovered during the current release. </details<details<summaryHelp</summary<ul<li<a href=\"\"Additional instructions</a</li<li<a href=\"\"Commands for setting labels</a</li</ul</details\nProblem is that there's no fluentd plugin to fix the problem, so running a custom fluentd won't help, there's nothing to add to the configuration to make it work (unless you want to write your own solution) They don't think it's a docker problem and don't plan to fix it. However, it will be fixed in GKE\nThat seems fair from Docker's perspective. I'd rather have stable containers as well :) So you're thinking of switching from the current log drive - at least that's what my Google Container Optimized image in GKE 1.8.1 uses - to the journald driver to circumvent the problem? And then use a plugin like to recombine the partial messages?\nNo, that's not the case. Solution is TBD, but if you export your logs to Stackdriver (that has a limit of 100KB per entry) you won't see the difference\nI think there's still a difference in that the 100KB stackdriver logging limit is larger than the 16KB docker limit at which the partial messaging starts, so without proper recombination of those chunked messages you'll lose the ability to send log lines larger than 16KB into stackdriver. However we're also running our own fluentd that sends messages straight into pub/sub, which has a limit of 10MB per message. And although we're trying to keep the messages much smaller than 16KB it would be nice to be able to send bigger messages without instrumenting our applications to push the messages into pub/sub themselves. I hope you find a nice solution that will make this possible :)\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten /remove-lifecycle stale\n/remove-lifecycle rotten Still relevant, hasn't been fixed\nI've been reading up on this issue, it's really too bad docker would not allow the configurable log-max-size option, this would be an easy temporary workaround to the issue until fully resolved. Regardless, I also wanted to point out a 2nd issue that the log splitting causes which might help increase the priority of this issue. If you are using json logging and have MCD or Structured Arguements attributes in your log messages, what ends up happening unfortunately is the attributes are lost on the split messages. This is a huge drawback right now. In my case I am using spring cloud sleuth which adds span and traceid's to all log messages allowing you to easily filter and correlate all related logging messages. When the messages are split, the traceid's are not part of the message in elasticsearch making filtering by traceid impossible. My application logging is setup as follows: is it still the plan to have fluentd put the logs back together into a single message? can anyone update if any progress is being made? thanks for your help.\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\n/remove-lifecycle rotten Still relevant, hasn't been fixed\nWould using the \"concat\" fluentd plugin be relevant? It seems to have explicit support for our specific use case. Edit: For EFK, Adding fluentd-plugin-concat to fluentd-es image (, ) and adding to the Works well. Would a PR with those changes be acceptable?\ntrying to solve this as well. Having the same issue as where long lines are being truncated and MDC gone.\nPlease see The big problem here is concurrency. Just concatenating all partial log messages as they arrive may be wrong. Let's say your app logs two events, called and here. Then the Docker log messages may be written to the JSON log file (or whatever Docker logging driver is used, e.g. in plain Docker or Docker Swarm) in this order: The fluentd concat plugin cannot decide what to do. What records belong together? The good message is: This can be solved (similar to ) as Docker logs contain all the needed information here: So you just have to extend these methods and the struct by the partial log meta data: Then you have to fix Mind the ID and ordinal of the partial log metadata so that you concatenate the right records in the correct order.\nI just edited my above comment because I am not sure how the case of linebreaks (i.e. those written by the application) in log messages relate to this issue. However, often you can control the logging of your application and let it write JSON to STDOUT (e.g. for Java / Spring Boot applications) so you do not have linebreaks. Back to the original issue: I have serious doubts that will do it under load with a lot of concurrent partial messages. I observed this in a Docker Swarm setup with fluentd and the concat plugin configured analogously.\nCould you please reopen this issue? The concurrency is the problem here: Let your container simultaneously produce two log messages of 16KB (let's say 20 KB). Then fluentd will concatenate all four messages to one large (16+16+4+4 KB = 40 KB) message. The fluentd (or whatever you use) needs to know the ID of the partial log metadata to concatenate only those messages with identical IDs. So this needs to be implemented in the (default) local JSON file log driver in Docker (moby).\nProblem is that ID comes with Docker 18.06 if I'm not wrong, this is why I didn't use it: do we want to require a docker version which is 2 months old?\nis no longer much involved with the project. if you would like to open this back up to alter the implementation, please file a new issue\nIf you do so, please ping me as well.\nI have the concat plugin but still it is not concatenating, Any Idea?\nNo idea why it does not work. For up-to-date information, I have inserted: Just after the kubernetes_metadate injection."}
{"_id":"q-en-kubernetes-e6323017ba1994aaff872b2f4c1494ddea9fef1339b9854d6ec4e4e72dcfd235","text":"The following that have been failing on our 5k-node tests (https://k8s-): [HPA] Horizontal pod autoscaling (scale resource: CPU) [sig-autoscaling] ReplicationController light Should scale from 1 pod to 2 pods [HPA] Horizontal pod autoscaling (scale resource: CPU) [sig-autoscaling] ReplicationController light Should scale from 2 pods to 1 pod This is part of the correctness suite and large cluster tests are release-blocking. cc\n[MILESTONENOTIFIER] Milestone Labels Complete Issue label settings: sig/autoscaling sig/scalability: Issue will be escalated to these SIGs if needed. priority/critical-urgent: Never automatically move out of a release milestone; continually escalate to contributor and SIG through all available channels. kind/bug: Fixes a bug discovered during the current release. <detailsAdditional instructions available <a href=\"\"here</aThe commands available for adding these labels are documented <a href=\"\"here</a</details\nCould you PTAL into this or redirect as apt? It's a release-blocker.\nxRef\nThe tests seem to be passing on the normal correctness job - https://k8s- (except the latest run) but they're failing on large cluster and the reason seems to be timeout waiting for scaling to happen.\nIf i have to guess it is heapster who is failing to provide metrics. HPA itself has no direct dependency on the number of nodes in the cluster. Assigning to .\nis going to take a look.\nFrom the logs it seems like Heapster is unable to get metrics from nodes. It appears to be a separate issue from addon resizer restarts.\nThanks for checking it. I have got the same logs for smaller cluster (10 nodes), so at least we know that the problem is not related to big number of nodes/pods.\n- I'm seeing the test pass on our regular (small cluster) gce job - https://k8s- And seems like it was passing before and then started failing on our 2k-node job - https://k8s-\nthanks for pointing on it. The main difference between runs on small and big cluster - special node that we are creating for heapster container. And my local experiment have shown that without this node everything works, but with it - fails. So I assume that it's more network related. I'll try to figure out what's going on.\nRemoving this as release-blocker and downing the priority as it's a setup issue specific to use of heapster node. Would continue to work on it (thanks!)."}
{"_id":"q-en-kubernetes-ee95ce3e8354412acf6a858173a67c0d8a56bb420b9e5419dcb2656015655438","text":"When creating a service passing the annotation , two IPs are created in the service, a private and a public: Having a private IP is not a desired scenario for service discovery in federated global environments. If we have, for example, a cluster in GKE and join a local OpenStack cluster, this private IP is written in the DNS zone but the GKE cluster won't be able to access it. We could add an annotation to say we don't want the private IP to be registered in service's LoadBalancer Ingress field, like: or And when it's false, only the public IP would be written in the service, therefore in the DNS zone /kind feature\nThere are no sig labels on this issue. Please by: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. You can find the group list and label list . The in the method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals\n/sig openstack /sig federation\n'' sounds better.\nI don't like adding annotations just to make things work. I counter-propose just not publishing the private IP whenever there is also a floating-ip (without a configuration option to control this behaviour)."}
{"_id":"q-en-kubernetes-cb4afa4f14a2fe70a404a2ed5cc8980ed563f9fdc27f42e45bedd1ceb0aafcec","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: FC plugin returns TargetWWNs at GetVolumeName(), but TargetWWNs are not unique identifier if user has multiple volumes from same target wwn. As a result, after introducing attach/detach functions, FC plugin always fails to attach volumes if a pod requests 2+ luns. The GetVolumeName() is called from GetUniqueVolumeNameFromSpec() during Pod creation. FC plugin should return unique identifier in the GetVolumeName(). If user uses WWID to specify FC volume, this problem doesn't happen because the WWID is unique identifier. What you expected to happen: All FC volumes are attached to pod. How to reproduce it (as minimally and precisely as possible): Create two PVs and PVCs which uses same target wwn. Create Pod with these two PVCs Anything else we need to know?: Environment: Kubernetes version (use ):Install tools: local-up- /cc\n/sig storage\nThere are no sig labels on this issue. Please by: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. You can find the group list and label list . The in the method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals"}
{"_id":"q-en-kubernetes-8a3c1681b5334091fa7e540c5d2e126d46c67a71901875b6907a631fd2f1260d","text":"Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: If a pod is already running in the default namespace with the expected name (i.e. ), the PV recycle method just assumes it's a recycler pod. However, if it is not a recycler, it could result in the volume being reused without the data being cleared. What you expected to happen: The recycled volume should not be reused until the data has been verifiably deleted. The easiest approach would be to make it an error to not run the recycler pod successfully, i.e. delete these lines: I think that should be safe, since the recycler will retry. Other errors in that method should also be considered fatal, e.g.:\n/cc\nHow about generating a unique name for each recycler pod so that it doesn't matter what existing pods are named?\nCould someone from your team pick this up?\nAck\nOpenShift runs the recycler pod in a dedicated namespace, not visible to users. Maybe it could be a good idea to do so also in Kubernetes. Anyway, killing the old recycler is probably the best option - it could be forged by user to block recycling and cheat quota. /assign\nAgree with running in a dedicated namespace, but it sounds like this feature is just being deprecated, so it might not be worth the effort. Thanks for picking this up."}
{"_id":"q-en-kubernetes-4c357555c1a42b6b33c3e0e24783c23e32cbaa5fcb7f8b1dad00d7ac8d48a357","text":"Is this a BUG REPORT or FEATURE REQUEST?: What happened: Kubeadm E2E test is currently broken due to the commit I report the same issue on my local cluster. If I revert PR , everything works fine. We believe the current weave-net plugin doesn't work with the CNI version change. Right now all kubeadm related development is blocked on this. Could we consider reverting this change from head until we figure out what's wrong in weave-net/CNI? More context in\n/assign\nPlease reopen this in kubernetes/kubeadm. There is an issue as e2e tests & PR job are failing, but locally the CNI bit seem to work for me, so I think it has something with building the debs to do. Not technically. We just don't have any coverage right now. Not IMO. We just need to forward-fix the issue. Closing this; you can reopen in kubernetes/kubeadm\nkubeadm e2e test still red. Do you have any progress or hint in terms how to fix it forward? Just to clarify, If I install k8s and kubeadm locally using deb, I can repro the issue on my test box. I could believe it's sth wrong with debs. In the meantime, CNI and plug-in like weave-net change is a big deal. We clearly didn't test it thoroughly for this PR. Reverting the change seems to be the clean solution.\nFYI kubernetes/release bumps new CNI releases for deb/rpm packages.\nThanks.\nIs this the fix needed to fix these tests?\nI actually think that will take us in the other direction. During the kubeadm e2e tests, we use the deb packages built by bazel. During our official release process, we use the debs built by anago (in the release repo). Bumping CNI in our bazel build broken our e2e tests. Bumping CNI in the release repo scripts will similarly break our official release builds as well. If I understand the problem correctly, the only way to fix both of them is for the CNI providers like weave-net to be updated to handle the changes of upgrading CNI to 0.6.0.\nI think this issue worth some more discussion. Re-open the case as it might not be affecting kubeadm only\nsorry, i cannot reopen the issue as it's closed by But let's continue to discuss the option here.\nYes, I think so\nThis will be fixed by ; this was due to the CNI binaries being extracted in the wrong place. It is kind of a kubeadm packaging issue. (Reopening just to make the PR close this)\nnice find!"}
{"_id":"q-en-kubernetes-df0ab5c417abf67f8df66a31d098c03dde39aafedc3f95f9ca324114a8dfa76d","text":"/kind bug Right now we do not validate PV updates so admin can update at any time. Kubernetes has no facility to reflect this update - it won't magically re-mount an existing volume to e.g. a different NFS server. IMO the PersistentVolumeSource to be read-only and set to stone once created.\n/assign\nI think it is a good idea to make the VolumeSource read-only but I am afraid ship has sailed in v1 and existing code/tooling that relies on the ability to modify PVs will break as a result of this change.\n- do you have any advice on how we should proceed on fixing this issue?"}
{"_id":"q-en-kubernetes-8245c28b2444b7939bc70c4ea4378fbd59b77f4e3afe2924bb64375c7fc38064","text":"I also hit when running cri-containerd as the container runtime. The reason is that we don't populate stats like in cri-containerd integration now. is adding the containerd support in cadvisor, after that we should have full summary metrics. However, heapster should not panic in this case anyway. Can we cut a heapster patch release and update heapster in the main repo?\n+1. Many new CRI integration would run into this issue.\nIn fact, as far as I am aware, not a single CRI implementation currently implements any of the stats calls.\nI beg to differ: That's the reason why filed the issue.\nYeah does implement the CRI stats. Currently it only support core metrics, which is only part of summary api. We are working on integration with cadvisor, before that some fields in summary api such as will be nil. Current heapster version will crash when is nil, which shouldn't happen. :)\n/cc\nMy bad, thanks for the pointer :slightlysmilingface:. I'll correct my statement: I believe no CRI implementation except for cri-containerd implements the stats calls. The ones I explicitly checked are: dockershim, cri-o, rktlet.\nSomething I'm wondering about though: The intention of the stats calls is in order to be able to get rid of cAdvisor. Making the stats calls depend on cAdvisor seems to cover up the issues it has. Re:\nHere is the reason that we still have to integrate with cadvisor for now.\n/cc\nIf I understand correctly this is meant as a migration mean, so that both are possible for some time, and then we drop cAdvisor once a reasonable migration period has been granted.\n[MILESTONENOTIFIER] Milestone Issue Current <details open<summaryIssue Labels</summary- : Issue will be escalated to these SIGs if needed. : Escalate to the issue owners and SIG owner; move out of milestone after several unsuccessful escalation attempts. : Fixes a bug discovered during the current release. </details<details<summaryHelp</summary<ul<li<a href=\"\"Additional instructions</a</li<li<a href=\"\"Commands for setting labels</a</li</ul</details\nI will take a look at Heapster and try to fix this. Can you advise how to reproduce the issue? This will help a lot with debugging and verifying a fix once it's ready.\nthis is already fixed at head and 1.4.3 version of heapster The problem that is using kubernetes version from head and we have there 1.5.0-beta.0: I see two possible solutions: version of heapster to 1.4.3 at head of kubernetes. Piotr mentioned that 1.5.0 has still a lot of problems, so it maybe makes sense. a new release of heapster (1.5.0-beta.1) and after that bump version at head.\nAnd thanks for taking this issue.\nNot directly related to this issue: a more general discussion is at We'd like to understand which fields are required and optional in heapster, so that we won't crash heapster with partial result.\nWe've merged a work around for now Hopefully it works. But anyway, we should fix this in 1.9.\nwhat needs to be done for 1.9 here?\nwe'll need an assignee as well\nWe just need to update heapster version."}
{"_id":"q-en-kubernetes-2fd3ecc784563ff8bcfd9a663503954fddec082b43ec87a26ac651f3b8d97742","text":"Is this a BUG REPORT or FEATURE REQUEST?: Bug Report /kind bug What happened: improved disk eviction testing to make it more precise. For example, it expects a pod which uses 299 Mb of disk not to be evicted if it sets its limit to 300 Mi. As shown on , this test passes for all OS distributions except CoreOS. I have ruled out an issue with the storage driver and backing filesystem since both COS and CoreOS use overlay backed by extfs. What you expected to happen: I expected the pod using 299Mb not to be evicted when its limit is 300 Mb. I expected not to see this event: How to reproduce it (as minimally and precisely as possible): The container runs this command to use 299Mb: Each iteration writes 1Mb to disk, producing a total of 299 Mb of disk usage. The pod has set to 300Mi. The test has the LocalStorageCapacityIsolation feature gate enabled. Anything else we need to know?: This may be a discrepancy in the (which we use to measure container disk usage) or , which is used by the test to write exact quantities of disk.\n/sig node\ncc can you take a look at this after thanksgiving?\nI bumped the coreos image... going to see if it clears up the flake\nIt appears the root filesystem is being included in the disk space calculation . Do you think the coreos image might be slightly larger than the other images under test?\nOne more code snippet that is interesting:\nthats the container's root filesystem\ncorrect. the feature appears to add and compare to the quota: The reproducible steps do not take into account the rootfs space. If the quota is set to 300Mi, then the under loop count would have to take into account the rootfs space to not exceed the limit. I do not see anywhere in the tests or the reproduce steps this part of the calculation.\nThe refers only to the writable layer of the container, and excludes base (image) layers. With the provided command, the only files that are in the container's rootfs are the ones written by the command, and should amount to 299Mb. Additionally, the container does not write any logs (through stdout), so that value should be 0.\nI ran a quick test on a vm with the current image, and found that both and correctly wrote and measured disk space correctly. I also tested as it inside a container, and verified that we are measuring the container's writable layer correctly. My guess now is that it has something to do with how we calculate disk usage by logging with journald. I am looking into it\nActually, the previous comment was incorrect. When I ran the test, I found that each file was 1028 kb instead of 1024 kb, which resulted in the command writing 300.3 Mb instead of 299 Mb. I will change the test to require slightly less precision. This appears to be because an empty file in coreos takes up 4 kb. Test this by running . On other OS we test on, empty files use 0 kb. This is what accounts for the difference.\nI tracked down an issue to the dd call. The stdout and stderr need a redirect to /dev/null to suppress the following messages: Problem Line: Fix:\nI PRed that fix and changed it to stderr, since the other tests seem to use stderr.\nleaving this open until we get a green run\nI'll continue to test locally to see if I find anything else.\nopened PR with the help of and CoreOS defaults to using the actual size instead of the apprent size."}
{"_id":"q-en-kubernetes-0b527d9f0a79d0b7d63ded7d416d2a117ee49ce3e1d17a33eb91bdbb8eefc1f9","text":"We should have e2e feature for resizing in GCE environment. Also add unit tests for node restrictions.\n/sig storage /assign"}
{"_id":"q-en-kubernetes-176559dd7b5229bf1beb3f47410b45e4e98dd5485fe7d6838b326c8bf6eefb91","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: What happened: kubelet throws the exception when --cloud-provider=cloudstack is included: What you expected to happen: There shouldn't be an exception. How to reproduce it (as minimally and precisely as possible): Add the following line to /etc/systemd/system/kubelet.service.d/10- Environment=\"KUBELETEXTRAARGS=--cloud-provider=cloudstack --cloud-\" and also configure the file. Anything else we need to know?: The condition check in the InstanceType() function in is wrong. It currently reads: but the correct code is (condition negated): Environment: Kubernetes version (use ): 1.8 Cloud provider or hardware configuration: cloudstack OS (e.g. from /etc/os-release): Ubuntu 16.04.3 LTS (Xenial Xerus) Kernel (e.g. ): 4.4.0-98-generic Install tools: kubeadm Others: /sig cloudstack\nI will have a look\n/sig cluster-lifecycle /area cloudprovider"}
{"_id":"q-en-kubernetes-062f12a7dce306a4f420f6a43e628cc39710d6dd0bd449d72acdcfd0a2463e6b","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: What happened: The test has been failing since Nov 22nd. Looks like it is because the test restarts Kubelet and then calls getLocalNode() which checks for node readiness. Not sure why this problem only started to show up recently. What you expected to happen: How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): 1.9 Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): COS Kernel (e.g. ): Install tools: Others:\n/sig node\nHi, Just FYI, . Maybe it is helpful for you."}
{"_id":"q-en-kubernetes-46097c621c9191c02764bca4140f5d877ea56feb42e42cb9e7567825bc1a3d84","text":"We should not attempt resizing of file system if volume is mounted as read-only. /sig storage cc"}
{"_id":"q-en-kubernetes-41e62ce8a5086af37fc663f085e37c9ab72c68bbd1313a56e312392935a45a6a","text":"/kind feature What happened: The DevicePlugins feature was introduced in 1.8 Kubernetes release as an alpha feature. Since then, we have seen good development work on stabling the feature as well as early adopters on using the framework to support device resources in Kubernetes. This issue is to track the work requirements of entering this feature into beta. FYI, here is list of requirements from the OSS Kubernetes feature stage doc: \"\"\" Beta level: Object Versioning: API version name contains beta (e.g. v2beta3) Availability: in official Kubernetes releases, and enabled by default Audience: users interested in providing feedback on features Completeness: all API operations, CLI commands, and UI support should be implemented; end-to-end tests complete; the API has had a thorough API review and is thought to be complete, though use during beta may frequently turn up API issues not thought of during review Upgradeability: the object schema and semantics may change in a later software release; when this happens, an upgrade path will be documented; in some cases, objects will be automatically converted to the new version; in other cases, a manual upgrade may be necessary; a manual upgrade may require downtime for anything relying on the new feature, and may require manual conversion of objects to the new version; when manual conversion is necessary, the project will provide documentation on the process Cluster Reliability: since the feature has e2e tests, enabling the feature via a flag should not create new bugs in unrelated features; because the feature is new, it may have minor bugs Support: the project commits to complete the feature, in some form, in a subsequent Stable version; typically this will happen within 3 months, but sometimes longer; releases should simultaneously support two consecutive versions (e.g. v1beta1 and v1beta2; or v1beta2 and v1) for at least one minor release cycle (typically 3 months) so that users have enough time to upgrade and migrate objects Recommended Use Cases: in short-lived testing clusters; in production clusters as part of a short-lived evaluation of the feature in order to provide feedback \"\"\" I think we met most of the requirements here, with the commit that if we need to make incompatible API change in the future, we will document the upgrade path and if necessary, may support multiple API versions in a couple of releases. Things that I think would be nice to have to enter beta (entering beta may not necessary depends on their completeness) Regarding to the thorough API review requirement, I think we had good discussions during the design proposal but there have been some recent discussions on whether we should use registration vs probe model for the general Kubelet plugin communication. I think we should have a general discussion on this topic with larger audience like people working on CSI plugins. I will create a FR issue to track this discussion soon. We would like to see multiple device plugin implementations. We know some people have started working on device plugin implementations for Solarflare, FPGA, and Infiniband, and probably want to keep closer track on their progresses. We would like to target this FR to 1.10. Please let us know your thoughts or add things that you think are critical to finish. Thanks! What you expected to happen: How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others:\n/sig node\nI'm a fan of probe model. I do feel it unfriendly to plugin developers to let them care about the re-registration after kubelet's restart. And have received complaint from my colleague.\nHi, for cluster reliability, I think we need more test cases, including scale and soak tests, to simulate more production scenarios.\ncan we consolidate this with ?\nSure. Will do this with Vish after KubeCon.\nSorry to bring up this topic a bit late (feature freeze is the 22nd of Jan). Here are the features we (NVIDIA) consider important or blocker before entering beta: [BLOCKER] Per container allocation Add labels during device plugin registration (e.g: gpu-family: tesla) Add annotations to a container for beta and add an option to CRI for GA to support OCI hooks Things that are not explicitly related to the device plugin: We're also wondering if we should tie up\nThanks a lot for your input, For per container allocation, I think we should be able to implement it in 1.10. I think we should be able to support ER quota in 1.10 as well. I think allowing device plugins to export device properties is an important feature, especially when we introduce ResourceClass, but I don't think it is a beta blocker. For \"Add annotations to a container for beta and add an option to CRI for GA to support OCI hooks\", this seems outside the scope of device plugin. There are different ways to implement them (if they have to be implemented) on Kubernetes. We haven't seen strong use cases indicating that they have to go through device plugin.\nSorry, I meant from the device plugin. i.e: the device plugin's returns container annotations for beta Product wise this would allow us to have the nvidia-runtime be supported by CRI-O (which allows hooks through container annotations) as well as docker + nvidia as the default runtime. BTW That would allow a convergence towards a single device plugin in beta. As for GA this would be a different discussion but it would be part of converging towards using any runtime and have them support the nvidia-hooks.\nNot saying it's a beta blocker but it would be useful to have labels on homogeneous nodes about what devices are present. Especially: GPU memory GPU compute capability ... It's also not a big change, it just requires consensus\nAs a heads up, I am going to send a PR soon to change DevicePlugins feature to beta. After introducing the feature in 1.8, we have seen good development work on improving the feature and received many valuable feedbacks from early adopters on how we may extend the framework to support various kinds of resources on Kubernetes. To us, entering beta means: a) We think we have done enough early experiments during alpha phase to get the confidence that this a useful feature that we are willing to carry on to beta and to GA. b) We think the current DevicePlugins is ready to be tried on production. It has good e2e and e2enode test coverage. Some early adopters have tried it on production and didn't report any major breakage. Very likely we will keep seeing new feature requests coming in after beta, as more people try to use DevicePlugins for their specific use cases. Some of these features may even require incompatible API changes. When this happens, the required complexity to support multiple device plugin API versions on Kubelet side should be manageable and mostly local to the devicemanager component. It should also be manageable for a device plugin to support multiple API versions. As an example, in , I extended gke gpu device plugin to support both v1alpha and v1beta1 versions. For reference, here is list of changes we made on device plugin during 1.10: Support resource quota on extended resources (Issue ). Owner: Added PreStartContainer grpc function to allow device plugins to perform per container start operations. Owner: Add annotations to the device plugin API. Owner: Track unhealthy devices in devicemanager, and have node status capacity to reflect raw capacity but allocatable to reflect healthy device capacity. Owner: Stub device plugin based e2enode test Owner: There are a few pending PRs that would be nice to get into 1.10 but we don't consider as beta blocker: Determine and implement a general Kubelet plugin communication establish model and use it in device plugin. Owner: Uses common checkpointing mechanism in device plugin (PR Issue ) Owner: Device plugin stress test. Owner: Please let us know if anyone has objections on entering the feature to beta in 1.10. Thanks!\nHi, for stress test, here is a PR :)\n/cc\nShould we reopen this tracing issue?\nI think this issue can be closed with We can open separate issues to track future device plugin changes."}
{"_id":"q-en-kubernetes-64d35f957f701c61c194e2a4471bc4b01744795f3cff95f58944cfc8a7bc9e03","text":"Internal volume plugin for CSI creates VolumeAttachment with (= the plugin name). It should be CSI driver name instead, as the external attacher must know which CSI driver to call. /sig storage /kind bug please prioritize PTAL. It should be one-line change in\n[MILESTONENOTIFIER] Milestone Issue Current Note: This issue is marked as , and must be updated every 1 day during code freeze. Example update: <details open<summaryIssue Labels</summary- : Issue will be escalated to these SIGs if needed. : Never automatically move issue out of a release milestone; continually escalate to contributor and SIG through all available channels. : Fixes a bug discovered during the current release. </details<details<summaryHelp</summary<ul<li<a href=\"\"Additional instructions</a</li<li<a href=\"\"Commands for setting labels</a</li</ul</details\nACK. In progress"}
{"_id":"q-en-kubernetes-4a027d49eaebe50fefe5b5d512519944b0a4c991a81fe22c30915de4af6e686c","text":"BUG REPORT While configuring the kubelet with: And having the sock file on that location (default location from cri), the following error occurs: Later line: Uses the correct socket. Result: Disruptive in understanding other failures. Containers still start. Environment: Kubernetes version: 1.8.0-5, 1.9.0-beta.2 Cloud provider: gce OS: Ubuntu 16.04 Kernel (e.g. ): Linux kubeworker01 4.10.0-42-generic ~16.04.1-Ubuntu SMP Mon Dec 4 15:57:59 UTC 2017 x8664 x8664 x8664 GNU/Linux Install tools: The hard way Others: Full kubelet config (as of now): Internal log reference: syslog.v6.3\n/area [kubelet] /sig [node]\n/area kubelet /sig node /assign\nThis is because uses , while is using as default. We should make them consistent. Fixed in . PTAL.\nWhat are the consequences? Is it more than cosmetical? Does it affect the usage of CRI-O in 1.9.0 or how does it affect users?\n/assign Is this relevant for CRI-O in 1.9.0?\nIt (at this moment) mostly looks cosmetical, however when I add a cloud provider (--cloud-provider=gce in this case) to the kubelet config, the whole thing breaks consistently on PLEG. I am trying to figure out if there is a relation between that and the\nCan you weigh in on whether this is a release-blocker for 1.9.0?\nTentatively adding milestone so we can track getting a judgement from sig-node.\n-- i think this is fixed in the packaging, let me close with and\nfixed in packaging yeah, from 1.8 we moved to \"\" so you need to start kubelet with that socket path\nDiscussed at sig-node earlier today, and this is not a release blocker. Also like mentioned above, the issue was fixed in the packaging.\nThere is a reference that needs to be fixed for the new socket location. The referenced pr is required to ensure stats summary endpoint understands the right location, and this keeps it consistent with cadvisor after its rebase in 1.9.\n/priority critical-urgent /kind bug /sig node /status approved-for-milestone\n/status in-progress\n[MILESTONENOTIFIER] Milestone Issue Current Note: This issue is marked as , and must be updated every 1 day during code freeze. Example update: <details open<summaryIssue Labels</summary- : Issue will be escalated to these SIGs if needed. : Never automatically move issue out of a release milestone; continually escalate to contributor and SIG through all available channels. : Fixes a bug discovered during the current release. </details<details<summaryHelp</summary<ul<li<a href=\"\"Additional instructions</a</li<li<a href=\"\"Commands for setting labels</a</li</ul</details\nHi, I am hitting the same error in this thread on my Openshift4.2 cluster which has configured CRIO on the kubelet. I would like to hit to get inventory of the containers. Any body has any insights why I am hitting this error when I make curl calls from the k8s node which has configured for the CRIO? [root /]# curl -v --unix-socket http://localhost/info Trying Immediate connect fail for No such file or directory Closing connection 0 curl: (7) Couldn't connect to server [root /]# sshadmin kubectl version Client Version: version.Info{Major:\"1\", Minor:\"17\", GitVersion:\"v1.17.1\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2020-01-14T21:04:32Z\", GoVersion:\"go1.13.5\", Compiler:\"gc\", Platform:\"linux/amd64\"} Server Version: version.Info{Major:\"1\", Minor:\"14+\", GitVersion:\"v1.14.6+\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2020-01-17T17:43:42Z\", GoVersion:\"go1.12.12\", Compiler:\"gc\", Platform:\"linux/amd64\""}
{"_id":"q-en-kubernetes-3ec2193ca46af582bc204df69a3cc09b9825bd6d0d73ca00d8cfa57d35fcbc1a","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: What happened: If a 1.8.x kubernetes master with a single controller-manager is configured with --leader-elect=false, deleting a deployment through the API with v1meta.DeletePropagationForeground doesn't work. The deployment has its deletionTimestamp set and a foregroundDeletion finalizer is , but the pod, rc and deployment are not deleted. What you expected to happen: The pod, rc and deployment should be deleted when the controller-manager uses --leader-elect=false. How to reproduce it (as minimally and precisely as possible): I used kubeadm with calico networking, and modified the controller-manager configuration via sudo sed -i /etc/kubernetes/manifests/kube-controller- -e '/leader-elect/ s/true/false/' and then ran to create and delete deployments. Anything else we need to know?: I suspect the problem is related to this change: It appears to require leader election before the GC controller will finish initialization. Looking at controller-manager logs with leader-elect=false, I saw this log entry from \"GraphBuilder running\" I did not see the \"started %d new monitors, %d currently running\" log entry at the end of the startMonitors() function. That seems consistent with the GC controller waiting for \"<-gb.informersStarted\" Environment: Kubernetes version (use ): tested with 1.8.2 and 1.8.4 Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): vagrant virtualbox with the ubuntu/xenial64 box Kernel (e.g. ): Install tools: kubeadm Others:\ncc Have you seen this in Openshift?\nReiterating the mentions to trigger a notification: <detailsIn response to : Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\ncc\nI think this problem occurs because initialiuzes the stop channel top nil when leader-elect=false. This causes garbage collector graph_builder to exit early. The following release-1.8 patch seems to fix the problem. creates a dummy channel which will never be closed and allows the garbage collector graph builder to work like it does when leader-elect is true. [Edit - just noticed that my initial patch had some unexpected differences]\ndo you really need to remove the following code?\nThat change was not intended. I also noticed that after I saved my change. I redid the patch and edited my previous comment to reflect that.\nAgree with the diagnosis and fix, are you going to open a PR?\nYes, I can do that. Reading your PR guidelines and ours.\nWill I need to go through to get this into the release-1.8 and release-1.9 branches?"}
{"_id":"q-en-kubernetes-1a12d72b130cd6c5f6223ab1a60f8ad327e65a32c0d0812529cea5e39ed0d863","text":"Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: When enable cpumanager by , but not set and , kubelet will get panic error in log What you expected to happen: Should graceful exit with error not panic. How to reproduce it (as minimally and precisely as possible): set and but not set and Anything else we need to know?:\n/sig node\nGreat that it's not panicing anymore, but I'd also love to get the \"requires + to be greater than zero\" hint in this particular case. It looks like even with this doesn't seem to be the case (since it errors out with the non-actionable \"unable to determine reserved CPU resources for static policy\". Should I open another issue for this?"}
{"_id":"q-en-kubernetes-f8813033efe7bf391fa27c7d7b1bb0c79ec38324266a98fda9672af353ff9519","text":"ref: and An e2e test should be , and the default should be updated. /assign cc\n/assign\nThanks I can take this on.\nCould someone point me to where the manifests are for the e2e test suites, I am happy to submit the patches?\nAre there API server e2e tests currently?\nthere are not HA API server e2e tests that I'm aware of\nI opened a PR for the default changes. Are there suggestions on a good E2E test to write if we don't have HA API server tests?\nwhat is needed for this issue to complete? Do you think you'll get that by Monday, which is Code Freeze?\nWe need some E2E testing or add an integration test as defined . It will not be done today.\nso are you planning to move it to 1.11, then? It's about to be automatically kicked out of 1.10 by automation.\nI do not have permission to change the milestone. If you could kindly change it, that would be awesome!\n[MILESTONENOTIFIER] Milestone Removed From Issue Important: This issue was missing the label for more than 7 days. <details<summaryHelp</summary<ul<li<a href=\"\"Additional instructions</a</li<li<a href=\"\"Commands for setting labels</a</li</ul</details\nReopening as this was reverted in"}
{"_id":"q-en-kubernetes-e5dd559d8e93f3b68dda2f7880256b764394bd676902286e26dcc72a914f8f2f","text":"Is this a BUG REPORT or FEATURE REQUEST?: /kind bug for lack of better idea hack/make-rules/test-cmd-, line 1443 is failing, even when the debug output shows the expected text: from the test log: What happened: See for example https://k8s-\nThis should be owned by whoever owns the \"test-cmd runkubectlget_tests\". This doesn't look like a failure in the testing framework or the infrastructure. I believe this is\nthe linked test is actually failing on the next test, not the .missing test... from the test log: this is very odd:\nthe issue appears to be with\npotential fix in\nseems like the same issue for runpluginstests:\nsame fix will hopefully resolve the issue"}
{"_id":"q-en-kubernetes-11e14369c9d5d1c949589f6b04829ea7c9f008438fd5ea63186b9d0b05c8824d","text":"Happy New Year! Now when I run make I get a diff like this in every zz_generated file: Not sure what a smart solution is, but we could merge something like in the meantime."}
{"_id":"q-en-kubernetes-4d460c4f5c8fb1c08c701d631c43381b00a0fa5b123446c3865b8df9432b458e","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: call delete endpoint when service is not found. What you expected to happen: It is no need to do delete in controller. Because in service restStorage, endpoint is deleted when delete service. See How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others:\n/area controller-manager /sig apps\n/kind bug"}
{"_id":"q-en-kubernetes-acdc78e86833cd2d34fb6f120f4e2db93a40df3e8df49e50ce89ac7da15fac91","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: The OpenStack cloud provider currently uses the name of the kubelet node to register with the API server. However the name is not guaranteed to conform to the DNS-1123 subdomain spec. Given a node name of , the following message appears in the kubelet logs: What you expected to happen: The OpenStack Cloud provider uses the instance id instead of the instance name. This will guarantee that the id conforms the DNS-1123 spec without making any assumptions on what the instance name may be. How to reproduce it (as minimally and precisely as possible): In the OpenStack console, rename a kubelet node to and attempt to register the kubelet node to the api server. Anything else we need to know?: Environment: OpenStack Newton Kubernetes version (use ):Cloud provider or hardware configuration: OpenStack OS (e.g. from /etc/os-release): Ubuntu 14.04.5 Kernel (e.g. ): Linux -cab9-44ff-9c99- 4.4.0-103-generic ~14.04.1-Ubuntu SMP Mon Dec 4 19:33:04 UTC 2017 x8664 x8664 x86_64 GNU/Linux Install tools: Others:\n/sig openstack\n/assign I prefer to conform hostname to DNS-1123 subdomain spec. Already submitted a PR on this.\nI think the issue may be more than just the hostname. The hostname on my vm is set to Since the OpenStack cloud provider uses the instance name to register with the api server, I also get error messages like the following: It seems like for OpenStack specifically, it sets it's node name to the name returned from the OpenStack metadata server () which is not restricted to being DNS-1123 compliant.\nThe node name must match ([A-Za-z0-9][-A-Za-z0-9_.]*)?[A-Za-z0-9], not contain '/'.\nI can confirm that Openstack can boot up instances with the name like . But this will break kubelet currently as stated above. For such cloud provider based nodes, the nodeName registered is retrieved from the provider, which should not be allowed when registering nodes for kubelet. Agree. +1. I've an extra check in PR . PTAL. Thanks.\nI think the question here is: Why did you choose to use as the nodename? This is the display name of the server, a property made for humans by humans. Essentially it is a field that everybody should be able to fill in whatever they think their server should be called. This doesn't need to be unique and has no guarantees about format or length or character format. We've had a number of people putting in unicode characters in the display name, for example. Furthermore, it can be changed during runtime of the server, without rebooting. Other possibilities for the nodename are: instance UUID, which is a string in a defined format, guaranteed to be unique and is only separated by dashes. Found in If you would like to have some way to associate your nodename with whatever is shown in the Horizon dashboard as server display name, there is . OpenStack already takes care to sanitize any weird display name. E.g. some weird display name like gets a reasonable hostname: <img width=\"745\" alt=\"screen shot 2018-01-10 at 08 52 49\" src=\"https://user-\"What do you think about changing the source for selecting the nodename?\ncc\nOnce upon a time, the cloud provider's (only!) role was to discover a dynamic list of nodes by regex. In that world, using the nodename made perfect sense because the admin could then configure it to be something like , with the obvious use of regex patterns to group the relevant nodes. Fast forward a few years, and this style of node registration is long gone, and there is no need to make the node name human readable, and in fact numerous reasons to choose other unique identifiers. There is still to removing \"human-friendly\" node names, however. In the end, I didn't see a good alternative that would please everyone (and meet the required technical character set/length constraints). I stuck with because at least it was under the control of the cluster admin, and so it could be obviously contrived to match our imposed k8s node name == openstack instance name constraint (once that constraint was communicated)."}
{"_id":"q-en-kubernetes-0e5caf5cfb38b9d4512381b892139458143e4a15520f4e7e3e7fae3dff54ca68","text":"/kind feature What happened: When calling , there is no way to override default flag values for the different components of the controlplane. This is surprising, as in the config structure (and in controlplane phase codebase itself), there are explicit settings for doing that (APIServerExtraArgs, ControllerManagerExtraArgs, SchedulerExtraArgs). What you expected to happen: Those config fields should be exposed as flags on and How to reproduce it (as minimally and precisely as possible): Try to deploy K8s using Kubeadm with a custom Admission Control setting (I'll follow up on this issue with a PR)\nbtw, I am not sure which sig owns kubeadm.\nit's sig-cluster-lifecycle.\n/sig sig-cluster-lifecycle\n/sig cluster-lifecycle"}
{"_id":"q-en-kubernetes-25b132162dbde8465ee9607da737f5e345abf92e128951426814368f451b343e","text":"Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: Currently, FC plugin returns volume name and empty FCVolumeSource during ConstrutVolumeSpec during . However, in order to recreate newMounter from volumeSpec, volume information are lacking. As a result, filesystem volume's reconstruction always fails. What you expected to happen: Filesystem volume's reconstruction works properly, then a volume is unmounted and detached from kubelet node. How to reproduce it (as minimally and precisely as possible): (1) Create filesystem volume using FC plugin (2) Restart kubelet service which a pod with FC volume is running Anything else we need to know?: cc Environment: Kubernetes version (use ):Install tools: local-cluster-\n/assign\n/sig storage"}
{"_id":"q-en-kubernetes-3df5572cd58f27bab781055f42524a22c1fb7412416cc5753b700b5ada306110","text":"If I run first time it goes fine (after reboot). For subsequent runs I get nodes NotReady. I've tried restarting docker but that won't help. Before christams this used to work all the time. Running Arch Linux if it helps. In I see\n/sig node\nditto for me: Looking for a work around I noticed that HugePages recently moved to beta. I can get pods to start again by disabling the HugePages feature gate .\n/assign\ncc\ndefault Arch kernel doesn't have a hugetlb cgroup controller compiled in(been there done that, I use Arch as well), so you have to prepare your own kernel to run it.\nIf it were a kernel problem I would have expected the kubelet to fail the first time it was started rather than the second time. This isn't confined to Arch, though. I'm having the same problem with the default Debian kernel:\ncan you check if your Kernel was compiled with flag? You can do it by running: I bet we have to add documentation, describing this behavior. On Ubuntu 16.04 machine:\nSure, looks like it was not: Documentation (especially in the error message) would be great, but if the kubelet isn't going to survive a restart, I would prefer that it not start the first time. Better to fail immediately than eventually.\ncan you please check if works (as in fails!) for you?\nNo, this doesn't resolve the issue. Logs from the first start: Logs from the second start:\nDo you have the directory? in that case what i have is insufficient :( If you don't have that directory, then we should probably look for a better spot to put that check? any advice?\nNo, that directory is missing on my system:\nI'm seeing this issue as well.\nI am preparing a fix to not fail when hugetlb mount is not present. Will post a fix shortly.\nI posted a fix here:"}
{"_id":"q-en-kubernetes-0627bb94bb5fca02aad214f0edd6e37d99d91b1147e43754349af8b983c9a9c4","text":"/kind bug What happened: When I want to enable both plugins by passing below flag, as they both need config file, and I find I donot know what to set If I set , then I get an error in apiserver\n/kind bug /sig auth\n/cc\n/cc\nfirst, you can't put comments in json. the error was complaining about your embedded comments. this would let the server start: second, the docs are really, really, not right (opened with fixes). if you want to specify config for multiple plugins, should point to a config that references individual config files for each plugin (which there's a bug in loading... fixed in ): example\n/assign"}
{"_id":"q-en-kubernetes-6c85c01d652d8d0597af90eb9bbbcf8b50d7bdfc0838cc8c10a9a1245f3eecfc","text":"Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: We face But there is no way to generate default config file How to reproduce it (as minimally and precisely as possible): Environment: Kubernetes version 1.9.2\nI'll take a look it /sig scheduling\n/assign\nHi, same issue with controller-manager\nAny updates here? , and don't have this flag. Is it even will be supported?\n/cc\nif not set, component should be use default value; why we'd like to dump all default config?\nBecause the documentation on the config file format is either nonexistent or difficult to find. If the program itself can generate a template, that fulfills (most of) the need for documentation of the format.\n+1 for this, from my expirence, add an option to write the default config is really helpful.\n/assign\nalso in 1.10\nDeprecating all flags, prompting folks to use config files in advance is great, except when there's no documentation nor example anywhere. That means that we can't actually migrate to the new settings, and I can't seem to find clear alpha/beta/stable status for the new config file spec, nor timeline for the removal. It is also pretty confusing that the scheduler expects us to use a config file, but the controller-manager does not?\nThis issue indeed applies to kube-controller-manager. And there should already had deprecation policy.\nAccording to , the flag will be available in . Since I can't find any reference online to an example config file, perhaps one could be provided or linked to here?\nwill you be able to pick my patch and generate a config file? I do not get my devbox handy so can only post here until Monday."}
{"_id":"q-en-kubernetes-f9fbfb6195731f14481a9309c722cc276aa2f7ee5be384a29725f5ddba99c457","text":"Is this a BUG REPORT or FEATURE REQUEST?: Bug Report /kind bug What happened: Applying a file that has no changes compared to current will output something like , which muddles the output when using What you expected to happen: Using the -o flag should suppress the unchanged messages in stdout to prevent output from being muddled How to reproduce it (as minimally and precisely as possible): Given the file Will result in proper yaml being generated as expected Run again, and is outputted to stdout, and is invalid yaml Anything else we need to know?: I do need to check whether this still affects the most recent version kubectl Environment: Kubernetes version (use ): 1.8.6 Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Darwin Install tools: Others:\nconfirmed on latest stable version (1.9.2)\nReiterating the mentions to trigger a notification: <detailsIn response to : confirmed on latest stable version (1.9.2) Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nI have a solution to this, still waiting on my company to look over the agreement"}
{"_id":"q-en-kubernetes-06a119d3bf4d89d1e39a6309861d68a562d1342ee18d916b2c137cdd125cfea1","text":"Currently as authentication and authorization are both standard config for all clusters, the kubeapiserver insecure http handler is unsecure, which means the cluster is vulnerable to the attack. So it should be deprecated, even removed totally.\n/sig api-machinery\nI think removing the insecure option is a good idea. Can you send an email the sig-apimachinery mailing list with a proposed timeline? I think this should be like other GA features and have a one year clock.\nSure, I will send it ASAP.\n:+1:"}
{"_id":"q-en-kubernetes-44131b7e999a0fabe95cfbf7dd2961f18a52dbb01c2d8163f7aea4b4c5a24f6a","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Today, . This seems unnecessary and should be done once for all the nodes. Is this a BUG REPORT or FEATURE REQUEST?: /kind enhancement /sig scheduling cc:\n/assign Thanks, will send fix per slack discussion.\n/cc"}
{"_id":"q-en-kubernetes-adcb0c9ee53392eab2322a311da8362e7fb12160e635a6d4bbb9e85d9b37786a","text":"https://k8s- https://k8s- cc: /priority failing-test\nSorry, but I will fix it ASAP.\n/assign\nPR is already sent. Wating for CI report..."}
{"_id":"q-en-kubernetes-25711d4a9776b8390a4b4965f8bce143c0c1fb481e60af07fcdd810f54da4ffc","text":"/kind bug What happened: kubectl became very slow, taking 30s to list pods etc. After going deep down the rabbit hole of debugging the controller components, I ran kubectl -v=8 which showed: What you expected to happen: Since caching is required for kubectl to respond in reasonable time and the issue isn't obvious, kubectl should at least print warnings in default log level. How to reproduce it (as minimally and precisely as possible): I got into the situation by running kubectl with sudo, but just chown'ing should be enough to reproduce this. Anything else we need to know?: Environment: kubectl v1.8.3\n/sig cli\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale /lifecycle freeze"}
{"_id":"q-en-kubernetes-d00dbe53908bc0696f2fc8ebc2e17a760855fc34f1003d209bc95418df0ac603","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: /kind feature What happened: Fix the following : The basic logic in of using the is continuing reducing the nodes slices through , , and . It needs 4 slices to do this. This is my thought to reuse the slices: Use 2 nodes slices, such as , The reducing process becomes: -Clear -Clear - please give some suggestion, thanks :smile:\n/sig scheduling\nThe challenge is that clearing a slice and reusing it still needs GC clean up, so it may not be effective. The size of slices shrink as we proceed through the function. So, having two slices and over-writing the elements in them can help reduce GC time, but this approach needs some additional checks to ensure we are not using stale elements.\nTo avoid clearing the slices, this is my thought: Declare to indicate the actual length of the 2 node slices, instead of . Appending a node to a node slice will add 1 to . Clearing a node slice will set to 0. WDYT\nYes, I can't think of any better solution either, but remember that you should append only to the slice in the first for loop of the function. The second slice should be initialized with the same number of elements as the slice used in the first for loop. Setting elements of the slices should be done by indexing them (not by append) in the second, third, and forth loop.\nOK, I will submit a PR to implement it."}
{"_id":"q-en-kubernetes-5beaecd7056580581b1feb4818ebe6236446e3d1a4106bb400a9050a86c90768","text":"http://k8s- /priority failing-test /kind bug /status approved-for-milestone /sig cluster-lifecycle the job is on the master-blocking dashboard and the e2e test has been failing, please take a look cc\ntestgrid change suggests\n(author of the change in that range) -- any idea whether might have broken reboot tests?\nI reverted the change as it may lead to this regression. The other change is really small () and unlikely related to this. Still the continues to say otherwise, I think we can reopen this issue.\nThe reboot test is still failing.\nJust opened to achieve the revert.\nIs the test passing now? Can this issue be closed?\nlooks green now"}
{"_id":"q-en-kubernetes-cc12fea149e57801be309e92f4cbb18173f151c09b5d937430a0de1c73622237","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: A Pod is run with local volume mounted in block mode. When that pod is deleted, the tear down process fails because device unmap fails. The device path appears to be empty, causing the unmap to fail. What you expected to happen: The teardown should have succeeded How to reproduce it (as minimally and precisely as possible): Create a pod with a PV for local volume in block mode. Delete the pod. Look at kubelet log files on the minion node to observe the error messages. Anything else we need to know?: Log messages: Feb 18 09:21:43 kubernetes-minion-group-crj1 kubelet[1510]: E0218 09:21:43. 1510 ] Operation for \"\"\" (\"-148c-11e8-b315-\")\" failed. No retries permitted until 2018-02-18 09:21:44. +0000 UTC m=+. (durationBeforeRetry 500ms). Error: \"UnmapVolume.UnmapDevice on global map path failed for volume \"blkdata\" (UniqueName: \"\") pod \"-148c-11e8-b315-\" (UID: \"-148c-11e8-b315-\") : Failed to unmap device from map path. mapPath is empty\" Feb 18 09:21:43 kubernetes-minion-group-crj1 kubelet[1510]: I0218 09:21:43. 1510 ] operationExecutor.UnmountVolume started for volume \"blkdata\" (UniqueName: \"\") pod \"-148c-11e8-b315-\" (UID: \"-148c-11e8-b315-\") Environment: Kubernetes version (use ): 1.10+ Cloud provider or hardware configuration: GCE OS (e.g. from /etc/os-release): \"Container-Optimized OS from Google\" Kernel (e.g. ): Linux 4.4.86+ Install tools: Others:\n/sig storage\nThis error is returned in this . It same that we set the global mount path to only if the volume plugin is attachable(see ). However, local volume plugin is not attachable, so will be empty and the operation will fail. I think we should extract the operation from the block, and execute this operation just before we try to map the device to global map path, then this global map path can be cached in for both attachable and non attachable volume plugins. Any comments? If not, I can help to fix this.\nThank you for the investigation. We've discussed this problem by offline yesterday. I proposed same fix which you mentioned above, then tested the fix works correctly with local volume plugin. I'll push a PR soon.\nThe device path appears to be empty, causing the unmap to fail. Please note, which was stored in is empty. device path isn't related to this problem.\nSorry about that, this is a typo, what I mean is global mount path.\nNo, no. My above comment is just note for the section. Your investigation is correct."}
{"_id":"q-en-kubernetes-a772bf832357874456524b88d38e41e362822bd4563e96cc87333a25fe1f4c86","text":"Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: The deployment controller for a deployment with a single replica failed to create a new ReplicaSet when the old ReplicaSet included an Evicted pod. Typically, the eviction was caused by \"node low on resource: nodefs\". When we manually deleted the evicted pod, the new RS and accompanying pod were created immediately. We have seen this happen at least a dozen times. Here are the audit logs for a recent case: ! What you expected to happen: The deployment controller ignores the evicted pod and creates the updated replicaset/pod. How to reproduce it (as minimally and precisely as possible): To reproduce, you would need a way to trigger a kubelet-driven eviction on a node that is currently running a single-replica Deployment with . Wait for the pod to be evicted and the new pod to be created, confirm that the evicted pod still exists, and then make a trivial change to the deployment spec. The running pod will be deleted, and no new replicaset/pod will be created. This issue does not seem to occur for manual evictions (via ); pods I evict that way disappear from the API server immediately. Anything else we need to know?: Possibly related to and/or Environment: Kubernetes version (use ): v1.8.5-gke.0 Cloud provider or hardware configuration: GKE OS (e.g. from /etc/os-release): Container-Optimized OS\nReiterating the mentions to trigger a notification: <detailsIn response to : Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThanks for the detailed bug report! From Here's what happened: Deployment was updated by the user, so it scaled all old ReplicaSets to 0 old ReplicaSet was scaled to 0 (ReplicaSet's == 0), but still owned the evicted pod, which wouldn't be killed automatically because of the above comment (ReplicaSet's 0) Deployment found that there were still old Pods, and kept waiting for them to be gone ~(wait for old ReplicaSet's to be 0)~ (Edit: Deployment also checks those old Pods by listing them to make sure there are none, not just by checking ReplicaSet's status) before starting the new rollout. We can either modify Deployment controller (to ignore evicted old pods) or ReplicaSet controller (to kill evicted pods ~when it gets scaled down~), if this is not the intended behavior. WDYT?\nI am actually debugging likely the same production issue with Recreate strategy in two clusters right now. (We run on patched 1.9.1.) I have traced it down to broken RS as well. in 1 cluster RS leaves around pod failed matching the selector in the other one those are evicted pods. all replicas are scaled down to 0 and report 0 in status but pods with valid controllerRefs to them are still present. Same is 0 for me but pod is still there I think the correct behaviour is for replicaSet scaled to 0 to delete all of its pod, not to leave them laying around\nxref - this should fix it"}
{"_id":"q-en-kubernetes-5044536d87edb25c665421bf3dc4b536a0586e8213d754dc6188da659a17c1aa","text":"There are no sig labels on this issue. Please add a sig label. <detailsA sig label can be by either: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. See the . The in method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nprobably caused by\nCould be a clash between these two?\ncc\nI still see the issue with local-cluster on the latest master: F0225 17:06:49. ] error: v1alpha1.KubeProxyConfiguration: FeatureGates: ReadMapCB: expect { or n, but found \", error found in byte of ......\nlooks like that's another issue, can you please file a new bug? with more logs? this one was about none of the hyperkube based processes starting up.\nwill do as it is also broke container kube-proxy as well. Same type of error."}
{"_id":"q-en-kubernetes-8af22912c2c635d67e653d6b757023f56ab288d2f9cd64f1b25799d46149083f","text":"Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: The field passed to is never set. In where is a object that is initialized in the method in Notice the field is never set. What you expected to happen: The field should be set to How to reproduce it (as minimally and precisely as possible): Create a CSI volume and set to true. Check the value of in the CSI call. It should be true, but is false. CC Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others:\nI think is only used for mounting pod volume path to container, here(for ), we can set to , just like other volume plugins. I can help to fix this.\nWhat about the case when a user wants to mount a volume for one of containers in a POD as read only, but the actual volume has ReadOnly False? I would like to suggest an alternative solution which passes the value of ReadOnly attribute, taken from POD's spec.\nThank you and for jumping on this! I feel like the other volume plugins have the wrong behavior here. The purpose of the call is to mount the volume. Therefore, the should receive the pod volume mount readonly field. That said, as mentioned , I think the fact that \" is used for mounting pod volume path to container\" might be hiding a bug: that is never set correctly. I haven't had a chance to validate this but my reading of the code suggests that the is never set. But because the container mount uses , even though the plugin did not get the correct readonly value, the mount ends up getting mounted with the correctly readonly value. Thoughts?\n[MILESTONENOTIFIER] Milestone Issue: Up-to-date for process Note: This issue is marked as , and must be updated every 1 day during code freeze. Example update: <details open<summaryIssue Labels</summary- : Issue will be escalated to these SIGs if needed. : Never automatically move issue out of a release milestone; continually escalate to contributor and SIG through all available channels. : Fixes a bug discovered during the current release. </details<details<summaryHelp</summary<ul<li<a href=\"\"Additional instructions</a</li<li<a href=\"\"Commands for setting labels</a</li</ul</details\nWe discussed the proposed fixes in the Kubernetes CSI implementation meeting today (see ). We decided it would be better for CSI to be consistent with the other plugins (and still broken), and have CSI fixed, along with all the other plugins, whenever we fix the issue (likely post-1.10). Therefore, I will accept fix and we will close Then we can remove the 1.10 tag from this issue and track it independent of the 1.10 release.\nSpoke with offline. Details here: Based on that conversation, I don't think there is anything else we need to do here."}
{"_id":"q-en-kubernetes-911d9a0f19d514fa6d7b32c9bf4057b2fd6096e5392c777c94bf520afe08eb3a","text":"/kind bug /sig cli /area kubectl Hello. This is my first interaction with the Kubernetes community, so please forgive me if i have left out any crucial details, or if, perhaps, this is PEBKAC after all! I am keen to have a working feature, as described in and I am running K8s v1.9.3 and realise that this feature is in alpha, but i thought i would ask anyway. We're not on v1.9.4 yet, but the changelog doesn't mention the feature. What you expected to happen: I have some services deployed to Kubernetes with YAML files, that has been working reliably. We always use to deploy things. When a service is already deployed, and i make a minor change in the config file, for example the service port, i expect to see that in the diff. Instead, gives no output and exits with a . How to reproduce it (as minimally and precisely as possible): Consider the file : I make sure that it is correctly running on the cluster: Great, looking good. Now, i change the value to use Docker image tag and save. Now, i run: The stderr/stdout output is empty, and the exit code is . Anything else we need to know?: Environment: Kubernetes version: Client Version: version.Info{Major:\"1\", Minor:\"9\", GitVersion:\"v1.9.3\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2018-02-09T21:50:44Z\", GoVersion:\"go1.9.4\", Compiler:\"gc\", Platform:\"darwin/amd64\"} Server Version: version.Info{Major:\"1\", Minor:\"9\", GitVersion:\"v1.9.3+coreos.0\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2018-02-10T01:42:55Z\", GoVersion:\"go1.9.2\", Compiler:\"gc\", Platform:\"linux/amd64\"}\nI have this same issue - if I change, say, to in my , running returns empty. If I try tweaking the version I'm diffing (which I don't understand) - e.g. - then I see a bunch of differences, but not actually the change I was expecting in the number of replicas.\nI'll take this. Sorry for the inconvenience."}
{"_id":"q-en-kubernetes-19c38c5aef02916223ebf43e4bcfc44bad182797959c49e81f1f90c4bf11f0f0","text":"Anything else we need to know?: I notice there are TODOs in apimachinerypkgapimeta,return bool instead of error func CommonAccessor, ListAccessor, Accessor. If is it necessary to do,I push a PR for it. Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others:\ncc\n/sig api-machinery\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-testing, kubernetes/test-infra and/or . /close"}
{"_id":"q-en-kubernetes-b8e02ae2095e747824abfe9f4ebb69035e8b5b6940da7076a9beb85ef9935c79","text":"IMO it's surprising that does not force delete a resource. only takes effect when . The description of reads this: We should make it more clear what does and even print warnings when is specified without .\nChanging the meaning of force to actually break cluster consistency be default could lead to data loss, which is why we haven’t changed it so far. agree we should print more info and clarfiy"}
{"_id":"q-en-kubernetes-ffc2193f6b59bee935ee96cd51beb06321ddb31968954b6ce1ed9e325ebbe6f9","text":"What happened: Deployed a webhook backend and service as follows: Please note that there is only 8443 port on the service. Register the webhook as follows: Got the following error when creating pods: What you expected to happen: Based on , should be called. Not the 443 port shown in the error message. How to reproduce it (as minimally and precisely as possible): Just deploy the same setup and register webhook. Anything else we need to know?: Environment: Kubernetes version (use ): HEAD Client Version: version.Info{Major:\"1\", Minor:\"10+\", GitVersion:\"v1.10.0-alpha.1.97+-dirty\", GitCommit:\"\", GitTreeState:\"dirty\", BuildDate:\"2018-01-22T20:00:24Z\", GoVersion:\"go1.9.1\", Compiler:\"gc\", Platform:\"linux/amd64\"} Server Version: version.Info{Major:\"1\", Minor:\"11+\", GitVersion:\"v1.11.0-alpha.0.470+\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2018-03-12T19:36:42Z\", GoVersion:\"go1.9.3\", Compiler:\"gc\", Platform:\"linux/amd64\"} Cloud provider or hardware configuration: GCE OS (e.g. from /etc/os-release): COS Kernel (e.g. ): COS\nthanks for the detailed bug report!\nI actually think that the API specification (or documentation) must change. Not all API servers will have access to look up the Service object and see what ports it defines. If we want to support ports other than 443, the port information must be specified in the webhook object, not discovered by direct Service object lookup\n/assign\nLooks like it will always use 443 no matter what ports are available, right? Should our documentation reflect this or should it behave differently? (also just sets 443 in the custom dial wrapper thing )\nAs a first step, I would adjust the documentation. As noted in , if we want to support ports other than 443, that information needs to be to the webhook config object"}
{"_id":"q-en-kubernetes-3b6dbdb44a6f9df4b3de4b65e9501cd8ca92f52be97882b827b8833b7ded20d6","text":"The new test (introduced by ) seems to be quite flaky. Example failures: https://k8s- https://k8s- https://k8s- Should a \"performance\" test even be run in the pre-submit PR runs? /cc\nIt does connect to etcd and does a fair amount of node additions and deletions. Most of the times I've looked at the error, it seems to be at the startup when connecting to etcd. Is there a way to retain this capability so we can continue to check on demand but not necessarily run every time?\nThe error in has already been fixed, if you notice , we now have a lock and it should not be happening now. For the rest, the error is in the startup itself, namely, , which is starting the framework connecting to etcd. Any idea's on why this should be causing flakiness? This was the same pattern used in schedulerperf -\nCan we please disable this test for the moment? It's causing significant flakiness and blocking other PRs from being merged...\nI'd suggest disable the tests first to stop the bleeding, and debug the issue at your own pace. It's probably not a good idea to run performance tests in this environment anyway since the tests are run in parallel (IIRC). Maybe you can set up a separate post-submit test job for these special-purpose tests.\ncan you take a look at PR ? It adds skip for short tests, which should avoid running this for very PR?\nhttps://k8s- shows that ipamperf tests are now skipped during integration testing. All checks have passed and has given /lgtm. Waiting for auto-merge. can you close this issue once verified? thanks!\nthe bot will do that automatically since your PR description says it fixes this issue."}
{"_id":"q-en-kubernetes-6fd62810a12dfe5feb7096a4c3d840f041d0eb4a5796bd0c4f87a69f5cc6bce0","text":"This is continuation of and . What happened: When StorageClass refers to a key that is not accessible to Kubernetes (i.e. AWS API returns AccessDeniedException), AWS plugin still tries to provision a volume under assumption that while Kubernetes can't access the key, the other components may access it and create the volume and attach it to a node later. Related code is - checks only for . Other errors are just logged and provisioning continues. Problem is that when AWS really can't access the key, it will provision some volume (ec2.CreateVolume returns success + valid AWS EBS), but this AWS EBS disappears after few seconds. Therefore Kubernetes does not know that something went wrong and creates a PV for the returned EBS. Kubernetes later fails to run a pod that refers to such PV, because the PV does not exist. What you expected to happen: Kubernetes must not create PV in this case and must report error during provisioning of such volume. Ideas: According to , the only signal that AWS produces is an event in Amazon CloudWatch Events that a key could not be used for encryption. Kubernetes AWS cloud provider is not subscribed to these events and does not get this notification. We would need new permission for controller-manager to get the events and some new code to filter them and report back to user. Kubernetes AWS cloud provider could wait for X seconds and re-check that created volume still exists and did not disappear because of inaccessible key. What would be reliable value of X? Kubernetes volume plugin for AWS could return both a warning \"I failed to check encryption key, created volume may not work\" that would be propagated to user and the suspicious volume at the same time, but the API between volume plugins and Kubernetes allows only one - either error or volume. It could be extended, still the user experience would be suboptimal. Kubernetes could send an error to user during volume attach (!) that a volume does not exist, please check your KMS key permissions. Again, UX is quite bad here. /kind bug /sig aws\nChecking the KMS key feels wrong here, as you could still create an encrypted EBS volume and not have the DescribeKey permission in IAM or via the key policy. If the main source of errors are sent via CloudWatch Events, then best practice would seem to be to subscribe via SNS to receive those events. Possibly fall back to vague \"send an error to user during volume attach (!) that a volume does not exist, please check your KMS key permissions,\" where a SNS topic isn't provided, but otherwise use CloudWatch events to enrich state/messages. Would then also need to change docs to reflect needing to set up CloudWatch Events -SNS topic. Big drawback of going down this route is a fair few downstream cluster bootstrap projects would likely need updating to support it in their CloudFormation / Terraform bits.\nAccording to it takes quite a long time until an event is available in CloudWatch events. what was the average / maximum delay?\nI haven't calculated the metrics accurately but the cloudwatch events arrival vary(and standard deviation was high). Using cloudwatch events was also considered for maintaining instance cache and as well concluded that - they are not realtime (AWS itself documents them as \"near realtime\"). It may be - we need to measure them again and find out if we can use them for the purpose here. Checking the KMS key feels wrong here, as you could still create an encrypted EBS volume and not have the DescribeKey permission in IAM or via the key policy. from jan's analysis above it appears that - if IAM does not have permission to DescribeKey then EBS volume gets created but then it disappears soon. So net result still is - if \"current user\" can't describekey, she can't provision the volume. Is there any documentation that indicates otherwise? cc\nTesting this for real now. User creates volume with no access to underlying CMK, or insufficient permissions does not seem to generate events with sufficient information as to what went wrong, i.e. don't appear to receive events for create volume failure, only deletion. E.g.: This policy is sufficient to allow the creation of volumes - isn't necessary for the volume to be created. It isn't sufficient to attach the volume however, creating the following error, The permission's required to fix this, to delegate access to EC2 itself: Bit of experimentation shows that only is required for attaching - the ones for creating the volume aren't needed at the attachment stage. The docs at is therefore correct for volume creation, but doesn't describe the volume attachment scenario.\nSo, what is our solution then? From the above it seems Kubernetes can't reliably check if a key is usable for encryption of a PV. And it can't get an usable event either.\nPossibly one of the following: change. Live with the pain. \"GenerateDataKey\" and \"Decrypt\" on the generated ciphertext blob prior to volume creation as a preflight check. The advantage being that we can throw an error that this isn't going to work beforehand, but with the disadvantage of doubling up API calls. This is probably fine, and I think if you're creating encrypted volumes at a rate that's likely to run into API call limits or get super expensive, you might have other problems? We then also do the CreateGrant check during volume attachment (as an optional extra). the same check, but only after an error. Problem being that we get much less in the way of guaranteeing what the root cause is, but we don't add to the API calls in the happy path. In any case, is this bug biting sufficiently to fix here, or should we start thinking about the CSI implementation?\nthanks a lot for detailed investigation. I'll try to get some time to implement 2. even before CSI. Users are very innovative in setting up their cluster permissions wrong and the messaging to user is very confusing in this particular issue.\n/assign\n2. does not work. Kubernetes does not need to have GenerateDataKey and Decrypt permissions to create encrypted volume. My AWS account is severely limited (I am using shared company account) and I can't call GenerateDataKey on a key: But I can create encrypted EBS using the same key: The volume gets created successfully. It's not the user who runs Kubernetes who needs permissions for GenerateDataKey, it's AWS itself.\nAnd the volume stayed created and didn't get reaped after a few seconds? I was confused about this part, it didn't seem to follow how the APIs should work. Has anyone ever created a volume with that key, or is it a new one? I'll try what I did previously in a bit across a few regions.\nYes, the volume stayed around for several minutes, then I deleted it manually. The key is our default encryption key, so when I create an encrypted volume with no key, I get this one. No zones, no regions involved.\nI got an idea, the volume is \"creating\" and then it gets \"available\". I could in a loop until newly created volume either gets \"Available\" or deleted. This will abuse the API quota ( is cheap, right?) + it will slow down provisioning a bit - all volumes I tried were Available in a second or two. Will it work reliably enough?\nMight be the difference then, as I don't think the same perms as I described above are needed for the default key as I assume that it gets automatically delegated to the EC2 service. I only tested & documented the explicit, alternate key path above. Your approach should work, but I don't know when EC2 API limits will kick in and how soon users would see that happen. Worth a shot though?\nIt takes ~4 seconds for a volume to reach \"available\", it's quite a long time. This will affect encrypted volumes with explicit keys only.\nFix:"}
{"_id":"q-en-kubernetes-e2ade88e089eb0f7e7abc747a644152ca565e2ec06df09967007e04822990020","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: pull-kubernetes-e2e-gce and pull-kubernetes-kubemark-e2e-gce both start kube-apiserver with admission and but without enabling priorityclasses and podpresets resources. We can see many error logs from\n/sig test /sig api-machinery\n/kind bug\n/sig testing"}
{"_id":"q-en-kubernetes-7db9dfb05fb9aca73ccd057783fe2daf7b1c3d9c614d505aac5a2093b003a26a","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: What you expected to happen: How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: This was introduced by following code change introduced in v1.10.0: candidate fix is here(only works on Linux, not a good solution): This issue is caused by PR: Environment: Kubernetes version (use ): v1.10 Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others: /sig azure /assign\nUpdate: above fix does not work on Windows, I am still investigating a better solution. This issue is caused by updated after , related code is here: while should not be changed from LUN to real device path in azuredd since AttachDisk happens on master, is a LUN(Logical Unit Number) number after AttachDisk process in the beginning happens on agent node: on Linux node, could be ; while on Windows node, would be a disk number, when is recalled, it could not tell whether it's a LUN or a disk number (this is the key point).\nthanks for issuing this problem. Could you submit a PR with your fix so that we can review it? Thanks!\nfinally I realized should not be used in func any more by referencing the gcepd code, PTAL , awsebs code should have same issue, I filed a aws_ebs issue here:"}
{"_id":"q-en-kubernetes-a6de7fe389275de81ca317b7188db31d3decc42fc5f32f62c44c61ec943de872","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . --Is this a BUG REPORT or FEATURE REQUEST?: What happened: When one pod is running and has two Containers. Container 1 exits with success and container 2 is running. gives correct info - the pod status is . However, , the pod status is . What you expected to happen: , the pod status is . How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): 1.9.3 Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others:\n/sig api-machinery\ni can not see your pod named nginx in your above code\n, updated the typo.\n/sig node /remove-sig api-machinery"}
{"_id":"q-en-kubernetes-a79a9edc6124e8116444ff8626f8ab117b9919dc85f36dd39ad35e13811a6ca3","text":"Is this a BUG REPORT or FEATURE REQUEST?: BUG /kind bug What happened: I am trying to get metrics working. According this there should be support after 1.8 kube. However, I am now using kube 1.10.1 and still I cannot see anything in ? Should these be there or not? What you expected to happen: I except that kubelet should expose prometheus metrics to /metrics path as well. Then it is easy to read with prometheus. How to reproduce it (as minimally and precisely as possible): Create new cinder volume, attach it to pod and try to query result from in kubelet Environment: Kubernetes version (use ): 1.10.1 Cloud provider or hardware configuration: openstack OS (e.g. from /etc/os-release): centos 7 Kernel (e.g. ): Install tools: Others:\ncc\n/sig storage\nPrometheus metrics should be available on default port of Kubelet. Can you try ? If you are on a RBAC enabled cluster (such as openshift) you may have to use token or key to authenticate for fetching metrics.\nreturns many metrics but there is not things. There is one pod which have cinder volume attached running in that node. So there should be some metrics right?\nIt appears that Cinder volume plugin does not implement metric interface necessary for getting those stat metrics - This seems like an oversight and is fairly easy to fix. I can help getting the PR merged if you can open it.\n/sig openstack\nKubernetes v1.19.16 there is no kubelet_volume* metrics in //metrics how to fix it???"}
{"_id":"q-en-kubernetes-8a2849c1d5058c6c5be662d50498532aa1beb7f8030f3f5350dc4d4f5f79fed7","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . If this may be security issue, please disclose it privately via --Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: Scheduler's pod informer has a selector that . This is incorrect. Scheduler must watch all the pods of the cluster, including those scheduled by other schedulers in order to update its cache of the state of the cluster. Otherwise, it will miss important events and the amount of available resources, pods on nodes, ... which are needed for scheduler to work correctly, will not be updated properly. /sig scheduling /assign\ncc\nFor updating it cache, shouldn't the default scheduler rely on node info/status? Also its not very clear, what information the default scheduler needs from other pods (not to be scheduled by it) that it makes decision on? Just trying to understand the issue more clearly.\nScheduler watches informer events to update its cache. Node status is not enough to update all the information that scheduler caches. Scheduler must know what pods are placed on what nodes. For example, in order to check inter-pod affinity/anti-affinity it needs to know the placement of all the pods, including those scheduled by other schedulers.\nBut this information is present here: So it seems that your PR will now make sure to get the scheduled pods' information directly from the pod informer.\nSo when I check the history, I see it was here:\nhave we seen any specific issues due to this change (like between default scheduler and DS controller)?\nYes, but scheduler relies on its pod informers to keep this location up to date. That's exactly why it is important for the scheduler to receive updates for all the pods in the cluster. Yes. We missed the problem at the time. This is problematic only in clusters that use multiple schedulers. I initially thought that DS controller could cause issues as well, but DS controller does not set , so it must not cause an issue. We have seen instances of scheduler's missing certain events in cluster and as a result scheduler's cache getting stale, but I doubt if it is related to this particular issue. This only affects clusters with more than one scheduler. If you run the test that I have written in with the head, it will fail. That tests simulate a multi-scheduler scenario."}
{"_id":"q-en-kubernetes-cee9608c4683b3add1d3f275396ffbdb04ad77cd7e9bec13c0104eb21124d069","text":"Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: endpoint of different k8s components expose Prometheus metrics that are not for them. For example, exposes metrics: or : What you expected to happen: Prometheus metrics only for to appear on kubelet's endpoint. How to reproduce it (as minimally and precisely as possible): Access the endpoint of all Kubernetes components Anything else we need to know?: This problem comes from usage of functions in different packages that automatically register Prometheus metrics. Just importing that package or some package that depends on it can cause those metrics to appear in unexpected places. In the list bellow you can find the different packages that have init functions and affected binaries: [ ] () cloud-controller-manager hyperkube kube-apiserver kube-controller-manager kube-proxy kube-scheduler [ ] () cloud-controller-manager hyperkube kube-apiserver kube-controller-manager kube-proxy kube-scheduler [ ] cloud-controller-manager hyperkube kube-apiserver kube-controller-manager [ ] cloud-controller-manager hyperkube kube-apiserver kube-controller-manager kubelet [ ] cloud-controller-manager hyperkube kube-apiserver kube-controller-manager kube-proxy kube-scheduler [ ] cloud-controller-manager hyperkube kube-apiserver kube-controller-manager kube-proxy kube-scheduler kubectl kubelet [ ] cloud-controller-manager hyperkube kube-apiserver kube-controller-manager kube-proxy kube-scheduler kubelet [ ] hyperkube kube-apiserver kube-controller-manager [ ] hyperkube kube-apiserver [ ] cloud-controller-manager hyperkube kube-apiserver kube-controller-manager kubelet [ ] cloud-controller-manager hyperkube kube-apiserver kube-controller-manager kube-proxy kube-scheduler kubelet Environment: Kubernetes version (use ):Cloud provider or hardware configuration: N/A OS (e.g. from /etc/os-release): N/A Kernel (e.g. ): N/A Install tools: N/A Others: N/A /sig instrumentation\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-testing, kubernetes/test-infra and/or . /close\nClosing this issue. <detailsIn response to : Rotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-4ce997675e7f6cf23dfc0d3dca6a918184b502867bb2d4c4c016335abfd16cd5","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . If this may be security issue, please disclose it privately via --Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: When use an existing azure disk(also called ) in pod, if that disk has partitions, the disk will be formatted in the pod mounting. What you expected to happen: How to reproduce it (as minimally and precisely as possible): create a disk with source from an os disk, e.g. that disk is already formatted before pod could access it Anything else we need to know?: Root cause: current diskLooksUnformatted func is using command to check whether disk is formatted or not while common k8s code() is using to detect disk format, could detect multiple partitions so if we use common k8s code(), following error will be returned for mulitple partition disks, which is expected: Environment: Kubernetes version (use ): v1.7 - v1.11 Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others: /sig azure /assign\ncc\niirc we don't support partitions on Azure volumes. There was a specific hang long ago. in all cases i think should be safe to based check now via the underlaying api\nthanks for the info, I am writing this PR, and also testing carefully. And this issue depends on , could you review it first? Thanks."}
{"_id":"q-en-kubernetes-822110c9798618f89f51c94b1e671c31b59d81fa212917ee9e22cd44bf55e005","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . If this may be security issue, please disclose it privately via --Is this a BUG REPORT or FEATURE REQUEST?: What happened: When trying to mount an Azure File mount it fails, it looks as though its putting a new line between the name of the Azure Storage Account and the Azure storage domain. There are no new line characters in the secret that holds the account name. The following is the secret: The following is the declaration within the section of the Kubernetes yaml: The following is the error that's shown in `kubectl describe po/pod_name: What you expected to happen: How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: azure OS (e.g. from /etc/os-release): Ubuntu 16.04.3 LTS Kernel (e.g. ): 4.13.0-1011-azure Install tools: Others:\n/sig storage\n/sig azure\nIt seems to me that does decode to , so it seems to be behaving correctly. Maybe it's a good idea to trim the return char for and fields after ?\nis specified by user, so trim the is not a good idea in the code. Another simple way is to use kubectl create secret to create azure-secret, the accountname and key won't require base64 encoded, so you won't mess up with let me know if you have any question, thanks.\nThanks !\nBumped into this issue today too on our cluster. Can confirm command worked. Is there an easy way to time the /n off the end of the echo for before using base64? Thanks.\nyes it's\nOf course it is! Doh!\ni think would've made better sense to trim the azurestorageaccountname value when receiving it... anyways the storage account name wont have n in them right? I think that just made better developer experience...\nis not allowed in azure storage account name\nso if azure storage account name does not allow in the name.. then why not just trim the name instead of error... Because devs would just and that would already have the n in the resulting base64-encoded string... right?\nsounds reasonable /reopen\nReopened this issue. <detailsIn response to : sounds reasonable /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nawesome. Thanks\nso then this goes in to kube version 1.14.6 in AKS ? or how does this work?\nIt would only go into k8s v1.16.0 since it's not a critical bug, we don't want to do hot fix to the old release.\ngot it. thanks Andy."}
{"_id":"q-en-kubernetes-7acb340cdbde988e3b24c46b1221b8c48d42cc0a66a01400559ba18c896b5d5c","text":"the said: However the are different, it does not allow to set spec and metadata, but other changes are not ignored. And updates to other parts are not validated.\n/cc for some suggestions.\n/assign\nMade a pull request:\nIs the documentation wrong or the implementation? I think there was a longer discussion about the correct behaviour in the PR, but haven't found it yet.\nIt strange that other parts rather than metadata and spec can be modified and they are not validated. So I think the implementation is wrong, at least.\nWe should follow the design doc, I think.\n/sig api-machinery\n/priority important-soon\nI agree the current behaviour is wrong. Added some comments to Especially, worries me whether we really have to right /status behaviour. Not really related to the given issue. But as we are touching the /status behaviour, we should clarify that as well.\n[MILESTONENOTIFIER] Milestone Issue Needs Approval Action required: This issue must have the label applied by a SIG maintainer. If the label is not applied within 3 days, the issue will be moved out of the v1.11 milestone. <details<summaryIssue Labels</summary- : Issue will be escalated to these SIGs if needed. : Escalate to the issue owners and SIG owner; move out of milestone after several unsuccessful escalation attempts. : Fixes a bug discovered during the current release. </details<details<summaryHelp</summary<ul<li<a href=\"\"Additional instructions</a</li<li<a href=\"\"Commands for setting labels</a</li</ul</details"}
{"_id":"q-en-kubernetes-8b5a7da39d65ea13e9207dcb507b6b11096287e69e5ca1a3e4e3a2ff32b15fc6","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . If this may be security issue, please disclose it privately via --Is this a BUG REPORT or FEATURE REQUEST?: What happened: On Windows Server with sandbox support the CNI gets called for all the containers within the pod including the sandbox container itself. What you expected to happen: The CNI should be called only for the sandbox container since they share the same network details. How to reproduce it (as minimally and precisely as possible): kubelet Pod at the parameters that get passed when the CNI is called Anything else we need to know?: This applies to Windows Server with sandbox support. The CNI should be called only for the infra container. Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others:\n/sig windows /sig network /kind bug /assign /cc\nReopen this because of\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nhas been merged now. /close\nClosing this issue. <detailsIn response to : has been merged now. Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nhas this issue been fixed? Is the CNI getting called only for the infra container? From what I see, the CNI still gets called for every single container from within a pod:\nSee comments It is required for windows containers.\nThe way I understand it, it's that the DNS issue will be fixed in Windows RS5 and there will not be required additional calls. The workaround at the moment is to call the CNI multiple times until the DNS issue gets fixed in Windows. As far as I know, that is the only reason why the additional calls to the CNI are required.\nIs this true? Do we still need to invoke cni for each container in RS5? And even with new cni dns capabilities ()?\nthe PR that you linked it's unrelated to this issue. The CNI should be called for the sandbox pod which is the infra container. All the other containers created within the same pod are sharing all network details, excepting DNS. This is an issue in 1709 and 1803 where not everything gets replicated properly to the other containers. As a workaround for this, you need to call AttachEndpoint from HCS on the newly created container network interface only to replicate the DNS as well. You can also see more details about this issue here:\nCan you re-open the issue in the meanwhile since this issue still exists?\nI still think this isn't a issue, but a requirement for windows containers. Even after RS5, we should probably support old windows versions. WDYH\nWhy would the spec of the CNI should be different on Windows? On Linux you know that the sandbox ID that will get passed to the CNI is always the sandbox container. On Windows the sandbox ID won't be only that, it will also be an ID of a container from within the same pod. Does this sound like normal behavior to you? For now I agree, the workaround is to call the CNI for every single container because of the Windows issue with DNS. But this doesn't mean we have to stick with this for future Windows releases which will not require calling the CNI ADD multiple times. Also, this is a hack for now () which needs to be fixed. Please re-open the issue until we get feedback from others as well, thanks.\nOK. Reopened the issue. Let's wait a while for feedbacks from other folks.\nApologize for the delay. Windows doenst have the namespace support native on the platform. we this support as part of RS5 Platform release. We are working on making runtime spec changes and CRI/ContainerD changes to use this model. here is a runtime spec PR that got merged. This is orthogonal to the DNS issue. Once we are done with the CRI/ContainerD changes then if u use CRI/containerD as runtime and if your OS version is RS5 and above then you don't need this second CNI add call. But as Pengfei said, if you are on a downlevel version or using dockershim as Runtime then you will have to make the second CNI call. So in my opinion this code will stay as workaround forever until we remove the support for downlevel OS. Hope this helps. I would recommend closing this issue.\nthanks for the confirmation. In my opinion this should stay open until it gets fixed at least for future Windows releases, because the way it is at the moment, the CNI will still be called multiple times regardless of the Windows version. Why would you recommend closing this issue? Shouldn't this be fixed for futures Windows releases? The other calls will be useless when Windows gets proper namespace support. Also, I recommend watching this presentation about the CNI which had place at the KubeCon from May I don't see why we shouldn't have this open until the issue gets fixed for future releases and we stick with the official SPEC of the CNI.\nand I had talked before about end of life of downlevel OS like 1803 and 1709. I think we should keep this open or open a new Issue to track updating the code so that the CNI is not called multiple times\nThis code would exists as far as we support 1709, 1803. The new workflow to use the windows namespace is getting implemented in containerd/cri. We can move kubernetes to use this new flow only for RS5 and above, if we plan to do it here as well. There isn't a good way to switch between 2 workflows of cni callflow wrt the Windows version hence we are not planning to change here till 1709 & 1803 exists, without adding some ugly code.\nthanks for the explanation\nI totally agree that this has to be fixed for future release and we are actively working on it. I was recommending to close this issue and open a feature to track this request. We should be able to raise a PR in another 2 weeks to support native network namespace support. This code will stay as long as we support down-level OS'es. I am fine either way. I am trying to close any open issues that is not a bug on Windows Networking(Platform).\n/kind feature Although from a kubernetes standpoint, this isn't a new feature. It doesn't need a new KEP and won't change any API. It will just be a future Windows network code cleanup.\n/kind cleanup\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\n- can you clarify a few things so we know what's required to fix this? 's the minimum OS version needed? Windows Server 2019, or a future release? this be possible to fix for dockershim, or just CRI-containerD? Thanks!\nWhat's the minimum OS version needed? Windows Server 2019, or a future release? [Dinesh]: Minimum OS version needed is 2019 Will this be possible to fix for dockershim, or just CRI-containerD? [Dinesh]: yes its possible to fix for Dockershim as well. but we are starting it with ContainerD. ContainerD is already done and available for testing.\nexcellent. thank you Dinesh!\nping - you have logs coming in email for this. Can you help us determine if it's needed for 1.14 vs 1.15?\nIf this issue has been triaged, please comment . If you aren't able to handle this issue, consider unassigning yourself and/or adding the label. I am a bot run by vllry. ‍\nPing\nIf this issue has been triaged, please comment . If you aren't able to handle this issue, consider unassigning yourself and/or adding the label. I am a bot run by vllry. ‍\nIf this issue has been triaged, please comment . If you aren't able to handle this issue, consider unassigning yourself and/or adding the label. I am a bot run by vllry. ‍\nStatus on this?"}
{"_id":"q-en-kubernetes-65a535cfff5b791a96dcd1ff2f9eb5fbcec23fe2286377bc564401eccda9490b","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . If this may be security issue, please disclose it privately via --Is this a BUG REPORT or FEATURE REQUEST?: /kind feature What happened: I am suffering from mysterious timeouts. See for the evidence from my latest investigation. It includes this line from the function in . The relevant line is this: I still do not know the type of the error in question or where it is coming from. It would be helpful if the log message included the type of the error object --- such as with or, if that does not include the type's package name, perhaps something that would really nail the type. What you expected to happen: I would expect more information about an unexpected type of error. How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): 1.10.3 Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others:\nI think this timeout is from resourcequota admission. Did you enable admission plugin? will create a pr to make ResourceQuota return err.\nthanks for the suggestion, I will respond in the original issue.\nI appreciate removing a source of errors of unexpected type, but this issue is about what to do whenever an unexpected type of error does arrive. I suggest formatting it with instead of .\n/reopen That PR did not really fix this issue."}
{"_id":"q-en-kubernetes-e51351c05e6844b765acf34eaa6a0653d5a4e3452baed4722f32f9c5caa3ece5","text":"/kind bug What happened: Wrote a that returned a faulty json patch which specified an our of range index (e.G. when there is only one container). This resulted in the apiserver crashing. Stacktrace: What you expected to happen: No crash /sig-api-machinery\n/sig api-machinery\nThis maybe a bug in . I will follow this question.\nThis bug had been fixed in PR: you can checkout it.\nThanks for your comment but I don't think the linked PR fixed the issue, as it got merged almost a year ago and I am running Kube 1.10.1 which does contain that patch but still panics. Just from quickly looking at the code I the linked PR I believe there are two things wrong: The check if the index is valid only happens when a negative index was given should be\nOnce the PR of is merged I will update kubernetes.\n/kind bug"}
{"_id":"q-en-kubernetes-6f81f24145ee2902ca3b0bab1114e1b9a4f27530220931ad1230ff6ec4e51769","text":"/kind bug /priority critical-urgent What happened: In I'm removing all the reapers, unfortunately when invoking foreground deletion on a DaemonSet the delete operation hangs and controller indefinitely tries to manage a removed DS. What you expected to happen: A foreground deletion properly removes DaemonSet and all its dependents. How to reproduce it (as minimally and precisely as possible): Invoke a delete operation against kube API (without using command). Anything else we need to know?: This behavior makes it so that none of the consumers of the API is not able to remove the DS controller due to present in the DS reaper. A working fix that passed the tests with the aforementioned PR is in but I'm not 100% it's correct. This is an issue that needs fixing before 1.11 goes out. /assign /sig apps /cc\nfyi\nI filed to fix it. As mentioned in the PR, this is because the DaemonSet controller races with garbage collector when it tries to update DaemonSet status of the DaemonSet being deleted. Here's what happens when hangs: /something performs a foreground deletion on a DaemonSet finalizer and DeletionTimestamp are both set history objects (ControllerRevisions) and pods are being deleted by garbage collector; meanwhile, DaemonSet controller tries to update DaemonSet status. Updating DaemonSet status requires constructing DaemonSet history objects, to figure out current revision and which pods do/don't belong to current revision updating DaemonSet status, DaemonSet controller tries to create a DaemonSet history object that matches current DaemonSet spec collector then tries to delete that DaemonSet history object. And repeat. Because we can't make DaemonSet pods be deleted before DaemonSet history objects (DaemonSet history objects don't own DaemonSet pods!), we cannot reliably calculate DaemonSet status without history objects anyways. Therefore, we don't update DaemonSet status for DaemonSet being deleted. Note that the reason why the kubectl delete hack works is because it forces DaemonSet pods to be removed before history objects.\n[MILESTONENOTIFIER] Milestone Issue: Up-to-date for process Note: This issue is marked as , and must be updated every 1 day during code freeze. Example update: <details open<summaryIssue Labels</summary- : Issue will be escalated to these SIGs if needed. : Never automatically move issue out of a release milestone; continually escalate to contributor and SIG through all available channels. : Fixes a bug discovered during the current release. </details<details<summaryHelp</summary<ul<li<a href=\"\"Additional instructions</a</li<li<a href=\"\"Commands for setting labels</a</li</ul</details"}
{"_id":"q-en-kubernetes-3967a737fc7deecce4890fde681f8a63314f2bde5e934b958729f0dc901e9d82","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . If this may be security issue, please disclose it privately via --Is this a BUG REPORT or FEATURE REQUEST?: /kind bug /sig CLI What happened: Running leads to nil pointer dereference: What you expected to happen: is expected to run delete with wait in that case How to reproduce it (as minimally and precisely as possible): (with any valid resource yaml) Anything else we need to know?: Discovered in when accidentally enabled wait for deletion from other commands than and saw failing integration tests in https://k8s- Environment: Kubernetes version (use ): latest master Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others:\nThis bug was introduced in : is set in that requires to be non-nil flag is being set in is normally being set in method which is not invoked in . a result, the invocation of resource finder leads to nil pointer dereference (see stack trace in the description).\n/cc"}
{"_id":"q-en-kubernetes-3be0627dd980896651c4a48dc6454da5515ef75b86a9211af25b89667f724769","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . If this may be security issue, please disclose it privately via --Is this a BUG REPORT or FEATURE REQUEST?: What happened: After joining a new ppc64le node to the k8s cluster, the kube-proxy DaemonSet fails on the new node. What you expected to happen: It was expected that the kube-proxy DS would \"just work\" as it does when adding an amd64 node to the cluster. How to reproduce it (as minimally and precisely as possible): Set up an amd64 based k8s cluster, then join a ppc64le node (or any non-amd64 arch) to the cluster. See the pod fail via 'kubectl -n kube-system get pods' because it's pulling the amd64 kube-proxy docker image instead of the ppc64le image. Anything else we need to know?: The yaml for the DaemonSet specifies the image: <versionIt would be preferable to have a manifest so that the image in the DS could be specified as <versionand could \"just work\" on any platform. See: Environment: Kubernetes version (use ): v1.10.3 Cloud provider or hardware configuration: self - cluster installed via kubeadm OS (e.g. from /etc/os-release): NAME=\"Ubuntu\" VERSION=\"16.04.3 LTS (Xenial Xerus)\" ID=ubuntu IDLIKE=debian PRETTYNAME=\"Ubuntu 16.04.3 LTS\" VERSIONID=\"16.04\" HOMEURL=\"\" SUPPORTURL=\"\" BUGREPORTURL=\"\" VERSIONCODENAME=xenial UBUNTU_CODENAME=xenial Kernel (e.g. ): N/A Install tools: N/A Others: N/A\nif you are running multi-architecture nodes, your daemonsets should specify a node selector to limit their arch-specific image to run on matching nodes\n/close\nI am well aware that a node selector can be used. That is what we are doing now. But it is a hack. The default should \"just work\". Adding a new non-amd64 node should not result in kube-proxy immediately failing on it because the default daemon set is amd64 arch specific. If kube-proxy is meant to \"just work\" clusterwide, a node selector should not be needed. Please use the manifest-tool so that the default \"just works\". Please see the weave-net plugin. They are using the manifest-tool and it \"just works\" on amd64 and ppc64le. Their daemon set does not require/use a nodeSelector.\nThe current kube-proxy daemon set implementation is not properly architected, IMHO. Please re-open this\nReiterating the mentions to trigger a notification: <detailsIn response to : The current kube-proxy daemon set implementation is not properly architected, IMHO. Please re-open this </details\nah, this was a comment about the kube-proxy image, not the daemonset functionality /reopen\nthe kube-proxy addon daemonset is probably the multi-arch kube-proxy image to support that daemonset might be\nLooks like it's coming from here: /assign /sig cluster-lifecycle Lucas, Do we want to avoid the references?\nA dirty workaround for this is to pull and rename to . The kube-proxy is running as daemonset, created during . That means the kube-proxy image arch keeps consistent with your master node. Currently we don't have a multi-arch kube-proxy image yet. We have to explicitly specify the right name, like .\nHi .. Yes, we just created a separate daemonset with a ppc64le node slector that refers to the ppc64le image directly. But it's a bit of a hack. That is why I opened this ticket looking for a multi-arch image. Thanks!\nI don't know that mixed-ISA clusters are supported. cc\nThe requirement is: only use \"\" in DaemonSet yaml file, and the k8s will download the correct image according to the ARCH info. It's more about image registry's feature; and it seems some registries had supported that: .\nYeah, has already supported multi-arch. But actually all the images now are architecture specific, like and . We've not built any multi-arch images right now.\nfor the test images we have the ball rolling here - - so if you can do the same (or similar) for other images, that would be great!\nI have pending PR in releases project for creating manifest for those -\nThe manifest images created in kubernetes/release should work for future released versions, but will they be created during CI testing?\nif you see one of the jobs that pushes images say you can grep for that's basically from the same files that got changed in so i fully expect the images pushed to CI will reflect the changes made in 516.\nplease confirm that i am reading this right ^ :)\nI was more concerned with ci jobs that do not use cross builds, such as https://k8s-\nplease open an issue about pushing manifests in CI as well"}
{"_id":"q-en-kubernetes-ee3d1281983d536929e8e7b68a82e18c9015d5299536de3b202935810516ee73","text":"BUG REPORT: Portworx volume plugin when a PVC is created using a StorageClass with empty parameters /kind bug What happened: Create a StorageClass using empty parameters and then create a PVC using that storage class. The PVC will remain in pending state. On inspecting the logs of the controller manager, we find that the portworx volume plugin is panicing. What you expected to happen: PVC should have been created successfully. How to reproduce it (as minimally and precisely as possible): Create following storage class and PVC Anything else we need to know?: Environment: N/A Kubernetes version (use ): 1.9.7, 1.10.1, 1.8.8 Cloud provider or hardware configuration: N/A OS (e.g. from /etc/os-release): N/A Kernel (e.g. ): N/A Install tools: N/A Others: N/A\n/sig storage"}
{"_id":"q-en-kubernetes-ef43dca034b09d199443c93bac85dd253a24c9ed4915ae7d3c71cfe2d547b0a0","text":"The mount code () is reasonably well tested but brings in the entire Kubernetes source tree when CSI plugins have to depend on it. With a few exceptions, most of the mount code could be refactored to be a minimal set of utilities that can be layered and make it easier to develop CSI plugins. I noticed that quite a few CSI plugins take a dependency on k/k just to get that one package. Suggestions: and split into a minimal dependency library and move it out of k/k (perhaps into staging and have it published out) stop depending on k/k and start depending on that library dependency on so the library is easily usable. In general, there is a lot of boiler plate for CSI, and pkg/util/mount is a significant portion of reuse.\nIf there no one was working on this, I can help do it. :-)\n/assign\nis this something you are still looking at? It's come up in sig-storage, and is on the planning spreadsheet for 1.12. It's something I'd like to work on as well, so want to see where you are at with it. While it didn't get any traction, I will point out that when me and some coworkers wrote several CSI drivers to the 0.1 version of the spec, we saw this problem, and factored this code out into a project that eventually landed at It looks quite a bit different now, because we found the k/k code didn't handle bind-mounts of block devices correctly, which was something new to CSI and K8s at the time.\ndid some work recently in pkg/util/mount to detect bind mounts better. What I would like to see is this handled in at least 3 steps: Mounter interface to separate out purely mounting helper functions and Kubernetes utilities the refactored mount library into an external repo, maybe under kubernetes-sigs mount library from Kubernetes and vendor the external repo\nAwesome. I plan to start working on this now, and i agree with the refactor approach. I do wonder where the resulting library should end up? Would kubernetes/utils be the right place?, rather than kubernetes/kubernetes-sigs? k/utils seems odd to me, though, as one goal is to not have to vendor a bunch of code you don't care about. but, k/utils has the intent of holding many different projects in it. So if you want one thing out of it, you have to vendor all of k/utils. I like seeing things in separate repos. =) Location doesn't hold up the refactoring, however, so I'll work on that!\nCan it just be a standalone package, like kubernetes/utils/mount?\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\n/remove-kind bug /kind cleanup\nHi Now that is merged, I wanted to confirm that the next step here would be to move to be a package in .\nYup!\nlet me know if you need a hand with git surgery\nThe git-foo wasn't too bad. Just slow to filter on the whole k/k repo. =) I've opened the PR at , feel free to take a look. Once that is in, we'll need to be cognizant to not letting any fixes come into k/k/pkg/util/mount, otherwise they might get lost. Once the package goes into , I can open the PR on k/k to make use of it!\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale"}
{"_id":"q-en-kubernetes-4789ac74fa70ad29e6904544c637f03d5eed63b0beb14cb39823301a3e747f34","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . If the matter is security related, please disclose it privately via --Is this a BUG REPORT or FEATURE REQUEST?: What happened: Currently, Calico CNI spec is specified in calico-node- under cluster/addons/calico-policy-controller without MTU specified (i.e., default 1500 is used). Given this is shared across many environments, we need a way to customize its MTU, say 1460 on GCP. What you expected to happen: How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others:\n/sig network\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-testing, kubernetes/test-infra and/or . /close\nClosing this issue. <detailsIn response to : Rotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-7897225e9f8fc61f52ed13d3fcd148a7378c85f696b86a3e58dea640f8cbe5a4","text":"/kind bug /sig network (I'm not sure this is actually a bug but it is definitely a behavior that is different from iptables and some applications rely on this) What happened: From the host network namespace accessing a ClusterIP when using IPVS does not work (it only works when traffic is sent to a local pod) What you expected to happen: To be able to use ClusterIPs from the host network namespace How to reproduce it (as minimally and precisely as possible): Configure kube-proxy in IPVS mode Create a deployment and the associated service Try to access the ClusterIP from the host Anything else we need to know?: This does not work due to how routing is set up when using IPVS If we create a service with ClusterIP , kube-proxy will add this IP to the kube-ivps0 dummy interface: which will add the IP to the local route table: (where 10.0.0.5 is the main interface for the instance) So when accessing the ClusterIP from the host, the source IP will default to the ClusterIP: After IPVS, the destination will be changed to a pod IP but the source will still be ClusterIP. Since ClusterIPs are not routable the answer will be lost. A workaround would be to modify the local route to use the IP of the main interface as source: I tested this and it seems to work fine. Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Ubuntu 18.04 Kernel (e.g. ): 4.15.0-1008-gcp\n/area ipvs\nDon't you have MASQUERADE rule somewhere? My clusters use flannel to create MASQUERADE rule and I guess it could also be done in kube-proxy with --masquerade-all=true option.\nI don't have any specific masquerade configuration: kubelet is run without so it uses the default (both my pod and service CIDRs are in this range) kube-proxy is run without So I have the following masquerading rules: (from kubelet) from kube-proxy for hairpin traffic My instances are also configured to NAT on KUBE-MARK-MASQ mark but in IPVS mode this chain is not referenced in the default configuration. I confirm that running kube-proxy with works but it feels like a workaround: if we do that we also masquerade pod-to-service traffic which seems unnecessary (and we lose the source pod IP)\nIs it possible to have pod ip and service ip configured in same range? Guess I am just familiar with my own cluster setup which use cluster-cidr.... Hope others can help you.\nSorry, I meant both are in but they are different ( and )\ncc I assume this bug has already been fixed?\nI am going to close this issue since it should be fixed by , please re-open if you still have troubles after picking\n/close\nThank you for looking into this I'll check with kube-proxy 1.11 very soon and will let you know\nI just tested and it does not seem to work I updated kube-proxy to but I still have the problem Looking at the code, it looks like the KUBE-MARK-MASQ is not to the service chain because the ipset is empty (which should not be the case because I actually have some services deployed) Did I miss something?\nDid you configure the --cluster-cidr?\ncc any comments?\nNo I did not I just realized this was the source of the problem here: We usually don't configure it because we have no need for it: we don't advertise the service cidr externally we would rather have pods using loadbalancer IPs actually hit the load-balancers (easier to understand and debug if we have similar routing for hosts outside the cluster and pods for the same \"host\")\nYeah, seems you find the root cause. clusterIPSet should not be empty."}
{"_id":"q-en-kubernetes-f5c15042ccd167fdcbf3bf3185eaab44440e5a9d4c76620fcf565867e306be24","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . If the matter is security related, please disclose it privately via --Is this a BUG REPORT or FEATURE REQUEST?: What happened: When I created a crd e.g.: Then I used . I found that the first time the result will be with an error: And as for the second time it will be normal: And it can be stably reproduced every time I created a new CRD. What you expected to happen: kubectl returns the first time I requested the new CRD instead of an error. Anything else we need to know?: I also found that requesting short name e.g. will be the same. It returns an error the first time but becomes normal the second time. Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others: /kind bug\n/area custom-resources\nConfirmed that could show up even the crd has status /assign\n/sig api-machinery Haven't looked into the problem yet but could be related:\nLooks like this happens because cache file in .kube/http-cache/ or .kube/cache/ is not synced with kube-apiserver. delete all files under .kube/http-cache/ and .kube/cache/ will cause a sync and the error will not show up. This looks like a cli issue.\nfyi\nTried to fix in cli:\n/cc\n/assign\nthis seems similar to\nEven though this bug looks small, it might involve larger changes both in fixing and refactoring.. A typical corner case it is /shrug\nCan you put this into words? What's the actual issue?\nEven though this bug looks small, it might involve larger changes both in fixing and refactoring.. A typical corner case it is Can you put this into words? What's the actual issue? SRY, I should have put it clearer.. I just mean that this issue is sort of fundamental and obviously has been existing for a long time.\n/remove-shrug\n/unshrug looks like an important problem we want to solve (?)\n¯()/¯ <detailsIn response to : /unshrug Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-testing, kubernetes/test-infra and/or . /close\nClosing this issue. <detailsIn response to : Rotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/reopen /remove-lifecycle rotten ouch..\nReopened this issue. <detailsIn response to : /reopen /remove-lifecycle rotten Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/cc\nencountered same issue, as suggested the error will not show up.\nany progress?\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-testing, kubernetes/test-infra and/or . /close\nClosing this issue. <detailsIn response to : Rotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nHere is a little more info, and here is how I am able to reproduce it (kubectl version 1.18.0 and cluster version 1.19.0-beta.2): A couple interesting things i noted... If I deleted my and re-ran it, I could not reproduce it. It is only the first time the short name was EVER used that this occurs. I only tried from one client machine, so I'm not sure this is cluster-specific or client-specific. If I used a different short name, like 'fo2' and re-ran the example, then it reproduces the problem. You can also run after creating the CRD and it has the same effect as deleting your . EDIT TO ADD: This doesn't seem to be a timing thing... if I sleep for 10 seconds after creating the CRD, it still has the problem.\nThe root of the problem is that the is run before the discovery. So in this particular case this line continues to b/c read cached data and never reached out to server. Whereas discovery mapper in will get a not found error and will proceed invalidating cache and re-fetching discovery data and return that back. In the subsequent call the shortcut expander is already aware of the new shortcut so it works just fine. We'd need to teach shortcut expander to re-fetch discovery data, but I'm not sure that's desirable, or we could at least warn a user about this situation.\nSee for possible approach.\nAny new progress on this ?\nlists a candidate solution. This bug is causing steadily failing tests. I'll talk to a few people to see if this can get any traction. do you know who can get this moving? It sounds like you have identified the problem and have proposed a solution."}
{"_id":"q-en-kubernetes-dab1fe7c944440f111a515f3fbe09f856e6908d7170d5b78cbd3cab62ea23db8","text":"/kind bug What happened: When I try to create a priorityclass object like: It fails with And when I try create it with yaml file like: It works. What you expected to happen: Create a priorityclass object with kubectl.\n/sig cli\nDo you have --runtime- set in kube-apiserver?\nNo, if set this, it works.\nAgree this is a bug. v1beta1 should be preferred.\n/sig scheduling"}
{"_id":"q-en-kubernetes-36795c0a12a6a354741c37701c68e5088c93bbab2b6679d97ee06c8316ba8f4d","text":"In the function, there is a that always runs and closes the channel when returning. This means that instead of running periodically, the image garbage collection will be terminated immediately when the function returns. The bug was introduced by in 1.11 /cc\n[MILESTONENOTIFIER] Milestone Issue: Up-to-date for process <details open<summaryIssue Labels</summary- : Issue will be escalated to these SIGs if needed. : Never automatically move issue out of a release milestone; continually escalate to contributor and SIG through all available channels. : Fixes a bug discovered during the current release. </details<details<summaryHelp</summary<ul<li<a href=\"\"Additional instructions</a</li<li<a href=\"\"Commands for setting labels</a</li</ul</details\nBTW, when the threshold is set to 100% (the use case that the PR originally wanted to fix), the code'd try sending to a closed channel and triggers a panic.\ni will check it soon, will submit one patch immediately, thanks\ndoes this bug applies to v1.11.2 ? or v.1.11.2 includes the fix?"}
{"_id":"q-en-kubernetes-13e97d3f49a69eb2215c309d2bf03a7b6720dbd9b08ad4d64caba81ba9efd05a","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . If the matter is security related, please disclose it privately via --/kind feature The field essentially has the type , but currently there is no validation for set and map semantics, i.e. unique elements / keys. Having this validation at API level removes logic to validate at the volume plugin level. edit: this check will not be in cases where this makes it difficult to add other fields and make this field optional. /sig api-machinery /sig storage\n/assign\nThanks for catching this. Would you be able to implement this for this release? We want to try to go beta if we can, and we won't be able to strengthen validation once that happens.\nYes will do!"}
{"_id":"q-en-kubernetes-9aaf31d7e18bd17771309c63b674c455f1fe5dc93f9635054771d17f9021cc73","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . If the matter is security related, please disclose it privately via --Is this a BUG REPORT or FEATURE REQUEST?: What happened: unit tests failed when run under What you expected to happen: unit tests pass How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: I've seen this on both docker for mac and linux, with varying uids (501 and 1000). This patch makes the package's tests pass: If this is acceptable, I will open a PR. Please let me know!\nLGTM\nPR:\n/sig cluster-lifecycle\n/assign\n/assign"}
{"_id":"q-en-kubernetes-eb31eee845f017c2a677d10fcbbb832392bb7172bdca09c5df46ed186d7f16c2","text":"/kind bug What happened: From I expected kubectl apply to support deleting a resource and recreating it, if changes cannot be applied. kubectl replace works: What you expected to happen: How to reproduce it (as minimally and precisely as possible): ` Change clusterIP to any IP address. Anything else we need to know?: Environment: Kubernetes version (use ): client v1.11.0, cluster v1.11.1 Cloud provider or hardware configuration: QEMU/KVM OS (e.g. from /etc/os-release): Cluster CentOS Linux release 7.5.1804 (Core), client macOS Kernel (e.g. ): 3.10.0-862.6.3.el7.x86_64 Install tools: kubeadm Others:\n/cc\n/assign"}
{"_id":"q-en-kubernetes-931f84d81282fff3a41b18e7549f7a758826e8e206296ef5e32bda6ceef3f7d4","text":"is a library yet is is instead of release tag. Please consider vendoring by tag whenever possible. Thanks.\nWe vendored it that way because the commit that fixed a bug (which we needed to be fixed and vendored back in) was This was merged on June 13. The release 1.1.4 () was released on July 1. If we want to update it to 1.1.4, I believe it should be fine. Between the commit we have vendored right now and the latest release, there are only three commits: adds tests merges (?) the case-sensitive fix - I'm not really sure what this commit is doing. fixes renaming when naming strategy is set (looks fine). Also, there are also no licensing changes. opinions?\n/sig api-machinery\n+1 to what said, it was a fire drill and 1.11.0 release was in trouble. Though in hind sight, we should have opened a issue to track that. something better to do next time. cc\nI'm indifferent to updating this. I'm okay with moving up if it doesn't break anything, or I'm fine with leaving things as is. We don't require deps to match semver. We can version out any commit in a deps tree, as we rely on our vendored copy.\nIt is always a good idea to match semver and it should be a default policy. Even if it doesn't matter to you, it should, because it is a best practice. Moreover downstream it matters much more as dependencies may be un-vendored in favour of tested re-usable packages.\n/cc\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nSomeone please take care of this or bot will eventually close this valid issue without action...\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\n/lifecycle rotten\n/remove-lifecycle rotten PR open (thanks,\noops! sorry, I meant to remove it. thanks!\nThanks.\nplease do feel free to open issues/prs of this sort early in the release cycle, so we can work through things needed well before the milestones/freeze etc. we are getting some heat about updating dependencies and we kind of settled on some guidelines here: if you think that needs updating, we could definitely debate that as well :) (on a PR against that doc)\nit have nothing to do with release cycle or justification for upgrade if/when dependency has been vendored incorrectly. It is basically a vendoring bug. No debate is necessary although that kind of issues usually meet some resistance and even hostility here... Bugs are filed when they are discovered, regardless of release cycle. It would be wrong to sit on the bug without reporting it in order to synchronise with release cycle."}
{"_id":"q-en-kubernetes-fde7b7109ad395c3188b90c171ac9a2c1273b27839bd3b08a2945411cdabcf05","text":"When we want to rotate certificates for aggregated API servers, we have to restart (kill) pods in order for API server to pick the new certificates. It would be very useful to implement some sort of self-awareness for the aggregated API servers, so they can pick-up new certificates when they're updated, without user interaction. If we have certificate files in pods, that are provided by some volumeSource, that can be persistent storage or a Secret, when file gets changed on the volume, the kubelet updates file in the pod usually in matters of few seconds. However, aggregated API servers are not self-aware of such change and users have to manually restart the pods. This brings additional complexity, especially if you want to automate the process. The first problem is that you need to know what pods you need to restart. But the potentially biggest problem is that you need additional permissions to do so. That would include RBAC permissions for listing pods, listing deployments and for deleting pods. That can make a security issue in some setups. By making aggregated API servers sort of self-aware, this could be avoided. Create a health check somewhere in , that utilizes a file watcher to watch for certificates changes. When certificate is changed, fail a health check. Failing health check will restart the pod without user interaction. This could be the most easiest and simplest solution to the problem. Such feature should/could be optional and guarded by a flag, let's name it . Some delay could be required between restarting pods. We probably don't want to kill all pods in the same time. This is up to be discussed. I would like to work on this issue, so feel free to assign it to me. /kind feature /sig api-machinery /area apiserver /cc\nthis could also be done by a purpose-built container that watched the same input files. as you point out, you likely would not want to restart all apiservers at once, which makes it unclear whether you'd want to build in the restart logic to the individual servers\nThat coordination would certainly work and be a valuable tool for someone to build to help manage rolling generic processes if we never get a pod restart. However, being able to have a \"usually won't all go down together\" option built-in is easy to understand and will resolve the problem until then. It isn't a stretch to say that if your process isn't fulfilling the current intent of the pod, it isn't healthy. This gives us a thing between dynamic config reloading and nothing. A container that can coordinate based on a service would be really cool, but I think that'd be a separate thing.\n/cc\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-testing, kubernetes/test-infra and/or . /close\nClosing this issue. <detailsIn response to : Rotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-be848cb963df255e526f4381997daea36d8be1f015783c2aa194aadbe017cdc5","text":"/kind bug What happened: Kubelet uses \"iptables\" instead of \"ip6tables\" in an ipv6-only cluster. This causes failed traffic for services (and probably a lot of other problems). For instance the chain is created in but is missing in ; When a is assigned the fails; What you expected to happen: All components should use in a ipv6-only cluster. should work in an ipv6-only cluster. How to reproduce it (as minimally and precisely as possible): Start a ipv6-only cluster Check iptables Optional; try a service type: LoadBalancer Anything else we need to know?: is hard-coded in finc in \"\" Environment: Kubernetes version (use ):Cloud provider or hardware configuration: kvm VMs on Ubuntu 16.04 OS (e.g. from /etc/os-release): Own BusyBox system Kernel (e.g. ):Install tools: None Others:\n/sig network\n/area ipv6\nThanks for reporting this problem. I'll take a look.\nis a link to the code with the hard-coded ipv4 protocol.\nThanks, I will send a pr to fix this issue.\n/assign\nI see this issue too. Also there is some related discussion with issue .\nI just pulled the latest from master, and gave a try with IPv6 cluster, but I still see the error, when using kube-proxy + IPVS: E0829 16:51:56. 1 ] Failed to execute iptables-restore: exit status 2 (ip6tables-restore v1.4.21: The protocol family of set KUBE-LOOP-BACK is IPv4, which is not applicable. Were you able to try this with kube-proxy+IPVS?\nSorry, I didn't try ipv6 cluster with ipvs proxier. Those two issue are related, but not exactly the same issue. Seems we already have conclusion in , nobody working on that?"}
{"_id":"q-en-kubernetes-b924de16d6035105b39c73378e11775c26df0501343981d64b9de4b8561a70c1","text":"/kind bug What happened: From doest not find kubeconfig file What you expected to happen: To find kubeconfig file How to reproduce it (as minimally and precisely as possible): from conformance file Anything else we need to know?: Line 185 : Starting kubelet with args But does not exist. Using from line 80 seems to work properly Environment: Kubernetes version (use ):Cloud provider or hardware configuration: hardware configuration OS (e.g. from /etc/os-release): Fedora 28 Server Kernel (e.g. ): Linux 4.17.12-200.fc28.x86_64\nReiterating the mentions to trigger a notification: <detailsIn response to : Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\n/area conformance"}
{"_id":"q-en-kubernetes-3d30f80339d7c67274e12c6190891fb695bbdad2c822e929f8e060cdf1647787","text":"/kind cleanup /sig testing /priority important-longterm This is a followup issue to address the \"standard build globals\" nit on At the very least I think we shouldn't capture our custom env var. Also think it'd be worth documenting the assumed variables in a list, similar to\ncc\n/assign"}
{"_id":"q-en-kubernetes-3a221a338368674d5126199b1778d869ca3bfa48f374eb3d5df1b4610faecc8a","text":"We should consider updating the command to wait on all resources for a given kind. For example: would wait for every pod in the current namespace to be deleted. /sig cli cc\nInterestingly, the error text implies that this is already supported: Yet when you try to use it, it declares Would really like the wait functionality given some much needed TLC. Still no way to determine what conditions are supported for various resources (eg. I can say and it just waits for the timeout like that resource might actually potato at some point).\nI'm looking into adding the --all flag to wait, sounds like a nice feature to have :)"}
{"_id":"q-en-kubernetes-2c40304f4ce54616aa2ff8fdb6b25ff95ba77b07d7a3e022946c73be95c97cd1","text":"/kind bug /sig release when we use , we must make sure xxx is not nil. So before we use it, we should check err value first. Such as:\nI think even if the file is nil , maybe It doesn't matter\nSometimes it does matter, while can't handle nil correctly. Maybe it is just a good practise.\nAlthough File.Close() can handle cases like this, serverConn.Close() will cause panic to happen. Anyway, these are bugs that exist. /assign"}
{"_id":"q-en-kubernetes-06ee7e79127d9dd9bc4352237b728a9dce0d91a9eb409a46162b7f1619cd999a","text":"Is this a BUG REPORT or FEATURE REQUEST?: /kind feature Alpine fixed a RCE vulnerability in apk in 3.8.1 () Update the base image used for addon-manager () to pick it up. Note: I'd send a PR if it depended directly on alpine image, not bashell/alpine.\nThere are no sig labels on this issue. Please add a sig label by either: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. See the . The in method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assing\nPing"}
{"_id":"q-en-kubernetes-5b95339d6b8555e50db00637b848f6bb68cc6ca87ee7be77a0a1c245f05a3972","text":"Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: \"[sig-storage] HostPath [It] should support existing single file subPath\" test failed on environments which don't support external IPs like What you expected to happen: The test can pass successfully. How to reproduce it (as minimally and precisely as possible): just kick: Environment: Kubernetes version (use ): v1.11.1 Cloud provider or hardware configuration: OpenStack OS (e.g. from /etc/os-release): Ubuntu 16.04.4 LTS Kernel (e.g. ): 4.4.0-134-generic Install tools: kubeadm Others:\n/sig storage"}
{"_id":"q-en-kubernetes-7f892a61af889b000f7a848ee9026bda4dff4173ad9a2b2cd61e9941f3c6d6af","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . If the matter is security related, please disclose it privately via --BUG REPORT: /kind bug What happened: Kubernetes distro testing v1.12.0-rc.2 shows in setups with self-hosted control planes (i.e. apiserver, controller-manager, scheduler run as pods) that previously worked on v1.11.3 and prior, the kube-controller-manager now crash loops. Notably, the kube-controller-manager is run with: which has been fine through to v1.11.3. Full . What you expected to happen: Controller manager may not assume it has root. Or should clarify if an additional mount is now required. May have been introduced in cc How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): v1.12.0-rc.2 Cloud provider or hardware configuration: GCE, DigitalOcean, AWS OS (e.g. from /etc/os-release): Container Linux 1855.4.0 Kernel (e.g. ): 4.14.67-coreos Install tools: Typhoon\nSeems like distros like Typhoon, Tectonic, and Tectonic/Openshift may work around this explicitly passing some other path to kube-controller-manager's or creating an emptyDir mount for . Need to understand the expectations around these certs (multiple pods, restarts, etc.) A release note seems needed to cover the new requirements. Appears this was also noticed by juju in\n/sig release\ncc\nWDYT? Can we get folks to write a note to cover this problem in known issues?\nI've an mount to address that crashloop, which seems to work preliminary (). was able to proceed with self-signed cert creation and hit a new bug I've filed separately EDIT: Perhaps a suitable release note for this change is:\nthe cert should not need to be persisted to disk. I don't think adding a writeable directory requirement just for a self-signed cert is a very good idea /assign\nPossible fixes enable by default if no cert is provided, rather than self-signing the self-sign cert in memory only, no disk persistence needed 't enable by default at all 2 would be ideal but is likely too big a change at this point. 3 would be unfortunate. 1 could be a good compromise if the code change is contained.\nIt's up to the SIG if this is a 1.12.0 blocker, or a Known Issue and 1.12.1 material. I'm inclined to think the latter.\nthanks for the clarifications. I would prefer to avoid my current workaround of giving it a writeable location so those all sound fine to me. is there a full story around these controller-manager changes somewhere? I could mint actual certs, mount from a secret, update clients. If there's merit. Like, how big of a deal is it that this is now always enabled vs before? I'll probably read through the various diffs after vacation.\nIf this is going to get some text in release notes Known Issues (seems like it should), somebody needs to draft the text and make a suggestion of it in ASAP.\n/assign\nDo we want to fix this for 1.12.1 (on friday)?\nSelf signing in memory means you get a different cert each time it starts. It's probably better than nothing but it's not how things are supposed to work (it's too inconvenient to manually verify the cert out of band in places users should trust it). OTOH if you want to do that, maybe you'll provide a writable directory in the first place, or even a real cert. So I guess in-memory is probably OK.\nSecond thought: is a terrible default for the cert path. It should have the word \"certs\" in the path somewhere. If folks (inadvisably) run apiserver and controller manager on the same system, they should at least not both try to use certs from the same place, or worse, write them there. I suggest as the default.\nWhat components are now verifying these certs or supposed to be? I'd expect something to be failing since random certs are generated each time the pod restarts, but don't see any impact (implying this isn't yet improving security). Seems like controller-manager just an HTTPS endpoint, but the HTTP one is still what's used right?\nI guess that aligns with the comment\nThe autogenerated certs only have SANs for localhost and 127.0.0.1.\nThe auto generated certs are self signed, right? If so, and callers won't be able to validate the signer, which SANs they have don't really matter.\nYeah, they're self-signed. So looks like the 1.12 controller-manager serves on HTTP/ (always on) and HTTPS/ (unless --secure-port=0). I guess the point of the original change was so some client could now choose to use the HTTPS endpoint with insecureSkipVerify. Main reason I was interested in the SAN is to get some clue about how a client would be expected to access these controller-manager pods (for contrast, apiserver SAN has DNS:kubernetes and an internal record that always resolves to a kube-apiserve pod).\nis this updated for the openstack cloud provider too?\nThis issue wasn't related to any particular cloud provider. The cloud provider / hardware config field is just a typical part of filing an issue in case it had been related.\nI have a writeable (I've tried other writeable directories too.) I'm encountering this: Should I post elsewhere in a new thread, or continue here?"}
{"_id":"q-en-kubernetes-f24559fc73906a9ab29b78f7d758564488c283b5a15c64d8dd22134579e26576","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . If the matter is security related, please disclose it privately via --Is this a BUG REPORT or FEATURE REQUEST?: /sig azure What happened: The internal IP address is got from instance metadata. I suspect the instance metadata may not be stable, e.g. it has undocumented rate limits (refer ). Cross refer What you expected to happen: How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others:"}
{"_id":"q-en-kubernetes-64c2aa792af3c4318a586d5758ca79a49c8660c3dd72f0d474f4ecf972caedf0","text":"Is this a BUG REPORT: Sonobuoy runs the kubernetes e2e's differently then test-infra. It constructs an in-cluster config vs. being handed a kubeconfig from the infra. In 1.12 something has broken this behavior, I'm still digging into the details. What happened: Run tests from within the cluster error occurs - What you expected to happen: Works fine, just like it had in 1.11 How to reproduce it (as minimally and precisely as possible): Run tests from within the cluster given proper RBAC rules. Anything else we need to know?: Discovered when updating sonobuoy for 1.12. /sig testing /kind bug /cc\nOne data point, the in-cluster-client-configuration example seems to work (but was missing RBAC step):\nis there a known time when it last worked?\n11 worked fine Sonobuoy's legacy in-cluster config works as well. The e2e test suite relies on a single utility wrapper that seems to not be tested. I'm still digging through the e2e setup, I haven't touched it since 1.9-ish timeframe.\nneeds a binary search bisect between 1.11 and 1.12. :\nI'm definitely interested in this as VMware runs e2e tests in cluster. We're still handing the tests a kubeconfig from the outside, but the plan is to move to the same model as used by Sonobuoy.\nThe token field in /tmp/kubeconfig-xxx is missing. BearerToken in client-go is apparently empty. Probably caused by this:\nFYI, the following command can trick sonobuoy/e2e 0.12 into running the tests but beware it is a hack: kubectl create clusterrolebinding cluster-system-anonymous --clusterrole=cluster-admin --user=system:anonymous\ni don't have proof, but looking through swathes of stuff, this one stood out - wanna try a local test in your environment with that removed?\nshared a possible solution in the sig-cluster-life chat, can confirm if that was a fix for the same thing or perhaps something else.\ni'll take back my comment about . trying to recreate the scenario i was able to print out the tmp file with the in-cluster config info. 11 - 12 - Looks like the information is missing in 1.12 exactly as mentioned. So i support the darxkies's theory that is causing it.\nThanks I verified that does fix.\n/reopen\nReopening this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nWhat should we do with Sonobuoy now that this is merged and cherry-picked? Do we need to wait for 1.12.1 to drop before we can use that test suite? Will we use that test suite for 1.12.0 clusters?\nClosing, I've been verifying with release-1.12 and with latest patches is works now.\n/close per\nClosing this issue. <detailsIn response to : /close per Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-77952b9d13e7df2e06ce1a1359e03872c5f8a26ecb63ada472455ba9ab5301bc","text":"The cluster/update-storage- are not maintained. It will be replaced by We still run the hack/test-update-storage- Its invocation will be removed in We can remove the script. /sig api-machinery /assign\nDo you mind if I work on this issue ?\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-testing, kubernetes/test-infra and/or . /close\nClosing this issue. <detailsIn response to : Rotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/reopen /remove-lifecycle rotten\nReopened this issue. <detailsIn response to : /reopen /remove-lifecycle rotten Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nare we ready to pull the plug on these shell scripts?\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\n/remove-lifecycle rotten"}
{"_id":"q-en-kubernetes-543b13c9a9f48ca7798c9f7c1ebbfd9520cb8578e5057e3b1e1232012333a197","text":"<!-- This form is for bug reports and feature requests ONLY! If you're looking for help check and the . If the matter is security related, please disclose it privately via --Is this a BUG REPORT or FEATURE REQUEST?: /kind bug What happened: Current logging is not enough for azure disk diagnostics, change some log level from 4 to 2 What you expected to happen: How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others: /sig azure /assign"}
{"_id":"q-en-kubernetes-aedaa0f61500c219af12281f2f4296d114f9871041077d12470e37a1f6808a04","text":"What happened: I do . I see the last 10 lines of output of each pod matching . What you expected to happen: I was expecting to see only the last line of output of each pod. How to reproduce it (as minimally and precisely as possible): The image prints the date every second. Normally we should just see two lines of output; instead we see 20. Anything else we need to know?: This seems to be a regression in Kubernetes 1.12, as this worked correctly from 1.8 to 1.11. (I observed this problem starting with k8s 1.12.0 and it's still present in 1.12.2) Environment: Kubernetes version (use ): 1.12.2 Cloud provider or hardware configuration: AWS OS (e.g. from /etc/os-release): Ubuntu Kernel (e.g. ): 4.4.0 Install tools: kubeadm Others: <!-- DO NOT EDIT BELOW THIS LINE --/kind bug\nIt looks like the issue was introduced by commit , specifically . If I understand that code correctly, it forces the number of lines to (10) even if is set. Since that seems to be in : /sig cli\n/assign\ndidn't that commit just move the code from ?\nNot quite; the old code was: The new code is: I don't know why it was changed, but the new logic reads as \"if the user specified a selector and a number of lines different from -1, then set the number of lines to 10\"."}
{"_id":"q-en-kubernetes-d65e248fc5643bf4731e91ed5f5258f9b82198151a44f01c3d1f0c1b465a91d1","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks!--Related to , there remains a leak in which does not ensure the channel it passes to is closed. This causes goroutines to leak under common scenarios, such as when calling . /cc What happened: Each call to leaks an active goroutine (in this case spawned by via . The goroutine will not end until the passed to is closed - however this requirement can and should be handled internally. Often callers will pass a context with a longer lifespan. What you expected to happen: should ensure that the polling goroutine is ended after its condition is met. This can be accomplished by creating a new channel with a child cancellation relationship to the original, and passing that on instead. This child channel can be closed once returns. How to reproduce it (as minimally and precisely as possible): The minimal reproduction would be to call with a true condition, and not close the passed . However, a more complete reproduction involving a can be seen . Anything else we need to know?: This leak was discovered in practice and caused a steady leak of memory and cpu in a project I work on. Environment: Kubernetes version (use ):Cloud provider or hardware configuration: Google GKE OS (e.g. from /etc/os-release): n/a Kernel (e.g. ): n/a Install tools: n/a Others: n/a <!-- DO NOT EDIT BELOW THIS LINE --/kind bug\n/sig api-machinery\nfyi Oren has 2 options to fix this. He'll be pasting sample code shortly and we can decide which approach we prefer.\nHi everyone! As promised, I've solved this two similar ways, and wanted to get your temperature on them before submitting the PR. I can then create a child context and pass its done channel to : This ensures that poller's goroutine will be cleaned up either when stopCh is closed or when the condition is met, whichever comes first. If there is aversion to using contexts here, we can mimic the subset of functionality from context that we need (safe to cancel from multiple goroutines + and a parent-child cancellation relationship): I lean towards option 1 and contexts - their semantics are familiar and result in less code on our end. Let me know what you think!\nI would prefer using contexts - we actually should in the long term have context variants of all waits, so it’s the right direction to start using them more here\n+1 to option 1\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\n/remove-lifecycle rotten"}
{"_id":"q-en-kubernetes-9983218d70647aa1d928911832354335c7346ebfe0e219e124bec32787f90cb6","text":"<!-- Please only use this template for submitting enhancement requests --What would you like to be : Add prebound pvs tests cases in integration stress tests Why is this needed: See discussion here: <!-- DO NOT EDIT BELOW THIS LINE --/kind feature\n/assign\n/sig storage"}
{"_id":"q-en-kubernetes-a88bbb486ddd61253aa094215a508e9388a721559376988ad853cf15f1e455ce","text":"<!-- Please only use this template for submitting enhancement requests --What would you like to be : Why is this needed: Because of is not recommended to create cronjob (with flags) anymore. But there is no imperative command available for to create cronjob. /sig CLI <!-- DO NOT EDIT BELOW THIS LINE --/kind feature\n/assign\nGitHub didn't allow me to assign the following users: Pingan2017. Note that only and repo collaborators can be assigned. For more information please see <detailsIn response to : /assign Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\ni wanna work on this. give me several days\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale"}
{"_id":"q-en-kubernetes-91e880b9ea081b4613e244dee4be17575c255c61a3bcc065b9f1e62c1e586e99","text":"What happened: On pods with a lot of DNS traffic, many DNS queries to the DNS ClusterIP fail during a rolling update of the DNS pods (in our case, we use coredns). What you expected to happen: Almost all DNS queries should succeed during a rolling update. How to reproduce it (as minimally and precisely as possible): Create a coredns (this would probably be the same with kube-dns) service backed by 2 pods Access this service by creating a lot of DNS queries, for example: Wait for the IPVS connection table to contain a lot of DNS connections (a few minutes) Delete a DNS pod A significant number of queries will timeout. After 5mn (depending in the IPVS timeout configuration) everything will work ok again. Anything else we need to know?: This comes from the current logic in the graceful termination code: The current logic is to delete a RS if the number of active connections is 0. This makes sense for TCP but for UDP the number of active connections is always 0. This is an issue for DNS queries because the RS will be deleted but the IPVS connection will remain until it expires (5mn by default) and if there are a lot of DNS queries, the port will be reused and queries blackholed. In addition, will not always work here because IPVS only removes connections when a packet is processed (in the case of DNS once the client receives an answer, no more packets are sent). Environment: Kubernetes version (use ): kube-proxy OS (e.g. from /etc/os-release): Kernel (e.g. ): /area ipvs /sig network /kind bug /assign\nI think there is also an issue with the way we clear the conntrack for UDP We call in : If I understand correctly this will clear conntrack entries matching and before graceful termination is done. I think we should only remove the conntrack entries when we delete the real server. In addition, we only clear the conntrack entries for <clusterIP, I think we should also clean them for Node IPs (Nodeports) and External IPs. What do you think?\nI think we still need to keep this issue open even though we have merged - I assume they are different issues.\nyes I think the PR is a great improvement but we probably need a few other small fixes (like the conntrack thing I mentioned)\nThe fix from the PR is not perfect yet for DNS traffic when qps is high (but it's definitely an improvement for UDP traffic which didn't support any kind of graceful termination). For example: Will generate 5000qps and ~500 new entries in the IPVS table per second. With a UDP timeout of 30s this leads to entries maximum. If we have 2 pods, when we delete one we have to wait for the ~7500 entries to expire, given the timeout (30s) many ports will be reused (which resets the timeout) and when the coredns pod finishes its lameduck period, we lose queries. 5000 qps against 2 coredns pods is high and unlikely but to decrease the probability of DNS queries failing, we can either: decrease the UDP timeout (in this example if set to 10s, we dont lose a single packet) increase the lameduck period (I tested with UDP timeout 30s and lameduck 180s and it worked great) (of course with more pods the impact is less important: what matters is the number of ports assigned to Terminating pods) I think being able to control the UDP timeout is very important, the default (300s) is very high (see: ) (I forked dnstrace to add the parameter: , I'll try and get it merged in the main repository)\ncan you please help confirm our issue? in our system, our dns query is quite often, like 1 query per 10 seconds. in this case, we found that if we restart coredns, the old IPVS rules are never deleted on some of our nodes, which caused a lot of failed dns queries. is this because the graceful termination mechanism introduced here? In order to delete the RS and remove the ipvs rules, we have to wait the DNS query expire, but before the old dns query expires, new dns query came in, so grace termination will never see and old ipvs rules are never cleaned up, causing dns query issue. if this is the case, I would suggest to introduce a force termination mechanism\n: it depends how you query DNS: if you have a persistent resolution socket, the IPVS backend may never expire but in that case your resolver should timeout and reconnect which will use an active backend. If you use a standard resolver (glibc for instance) each query will get a new connection so you will reach an active coredns unless you query frequently enough so you end up reusing a same source port associated with a removed backend (but below 100 queries/second this is very unlikely). I have ideas to address both use cases: configure more aggressive timeouts (in particular for UDP) and make them configurable (see issue: ) forcefully remove real servers (coredns in your case) that are still tracked by IPVS after X minutes (today we only delete them when they have 0 active and 0 inactive connections which may never happen depending on your workload) may have ideas on this topic too\nDNS resolution through kube-proxy is tricky for several reasons regardless of the proxy mode (for iptables for instance, many people have conntrack issues when doing a lot of queries). In our case, we are experimenting with node local coredns caches (you can also have a look at which is an implementation of this idea)\nisn't 100 queries/second quite common for a site? QPS can be thousands nowadays. we did not do special configuration for our dns so it should be using glibc. I think what you mentioned about reusing the same source port should be our issue here. by checking the output of ipvsadmin, it seems the source port is reused for different udp queries <img width=\"517\" alt=\"\" src=\"https://user-\"do you have any other suggestions on this issue?\nbtw, forcefully remove real servers is what I want here.\nI meant 100 qps for a single client (which can definitely happen though) Yes it is very likely that you have the issue I mentioned Until we implement the solution I mentioned natively, you can try changing the ipvs udp timeout on your instances. The default are: You can change udp timeout to 30s (or even lower depending on your connection rate): (0 here means do not touch tcp/tcpfin)"}
{"_id":"q-en-kubernetes-ef11399db688c17d851f33d1d9637037609227407835b25496b94b26eb4fa2bd","text":"How to reproduce the problem: Set up a new demo cluster with kubeadm 1.13.1. Create default configurationwith Initialize cluster as usual with Change the list in kube-apiserver manifest to , so that the first etcd node is unavailable (\"connection refused\"). The kube-apiserver is then not able to connect to etcd any more. Last message: kube-apiserver does not start. If I upgrade etcd to version 3.3.10, it reports an error Environment: Kubernetes version 1.13.1 kubeadm in Vagrant box I also experience this bug in an environment with a real etcd cluster. <!-- DO NOT EDIT BELOW THIS LINE --/kind bug\n/sig api-machinery\n/remove-sig api-machinery /sig cluster-lifecycle\n/sig api-machinery apologies, just had another look and it's indeed an api-machinery issue. // that can be used to communicate with a logical etcd cluster. For // example, a three-node cluster could be provided like so: // // Endpoints: []string{ // \"\", // \"\", // \"\", // } // // If multiple endpoints are provided, the Client will attempt to // use them all in the event that one or more of them are unusable. // // If Client.Sync is ever called, the Client may cache an alternate // set of endpoints to continue operation. we are passing the server list straight into etcd v3 client which return the error u reported. not sure if it's designed\nThis is an etcdv3 client issue. See\n/cc\n/assign So live updating a static pod manifest is typically not recommended, was this triggered via some other operation or were you editing your static manifests?\nNo pod manifest involved here. Just a group of etcd and a kube-apiserver. The issue appeared when we rebooted the first etcd node.\nI was able to repro this issue with the repro steps provided by I also reproduced this issue with a real etcd cluster. As previously mentioned, the problem is caused by a bug in the etcd v3 client library (or perhaps the grpc library). The vault project is also running into this: The problem seems to be that the etcd library uses the first node’s address as the for TLS. This means that all attempts to connect to any server other than the first will fail with a certificate validation error (i.e. cert has in SANs, but the client is expecting ). An important thing to highlight is that when the first etcd server goes down, it also takes the Kubernetes API servers down, because they fail to connect to the remaining etcd servers. With that said, this all depends on what your etcd server certificates look like: If you follow the to stand up a 3 node etcd cluster, you get a set of certificates that include the first node’s name and IP in the SANs (because all certs are generated on the first etcd node). Thus, you should not run into this issue. If you have used another process to generate certificates for etcd, and the certs do not include the first node’s name and IP in the SANs, you will most likely run into this issue when the first etcd node goes down. To reproduce the issue with a real etcd cluster: a 3 node etcd cluster with TLS enabled. Each certificate should only contain the name/IP of the node that will be serving it. an API server that points to the etcd cluster. the first etcd node. server crashes and fails to come back up Versions: kubeadm version: v1.13.2 kubernetes api server version: v1.13.2 etcd image: API server crash log: I was unable to reproduce this issue with API server v1.12.5 (n.b. this was somewhat of a non-scientific test =tested by updating the image field of the API server static pod produced by kubeadm v1.13.2)\n/assign\n^ FYI.\nthank you for the investigation\nNo problem! I started looking into the etcd client to see if I could spot the issue, but couldn't quite figure it out.\nI will look into this.\nI think this was the issue more on how gRPC balancer does failover with credentials. I've shared a workaround to fix this issue in upstream gRPC And waiting for their feedback. /cc\nWe have 3 master and 3 etcdservers, a workaround is to change the order of etcdservers. master0: master1: master2:\nI just tried using a SAN cert with my etcd nodes having the names of all etcd nodes. The apiserver connects to only the last etcd node in it's list. We run 3 apiservers per cluster and I can confirm the behavior the apiservers in this particular cluster is consistent that they are picking the last, however if I shutdown the last node the connect to the other 2 without issues, once that comes back up they connect to only the last. and there are entries like this: in the etcd journal, where is the apiserver IP. Without the SAN certs the behavior matches what is described by others here. So the SAN does not make the client distribute load across etcd nodes, it just prevents a total outage if the node the api-server talking to goes down.\n- what's the state here, this is still an issue.\nCan we ping gRPC team again? Open to discuss over the conference call.\nWill do!\nJust discussed with gRPC team and We will rework on this in the next few weeks.\nhi we are trying to create etcd cluster but we are facing the following error please find below error Error: client: etcd cluster is unavailable or misconfigured error : dial tcp 192.168.2.139:2379: getsockopt: connection refused please help us to get out of this thanks in advance\nI'm running kubernetes 1.11.9 and etcd 3.2.24 , running on AWS provisioned by KOPS 1.12.0 and im seeing the same error message : Unable to create storage backend: config (&{etcd3 /registry [http://127.0.0.1:4001] true false 1000 0xc42039fa00 <nil5m0s 1m0s}), err (dial tcp 127.0.0.1:4001: connect: connection refused)\n- any update?\nI just came back from trip. Will start working on this from this week! And post updates here.\nMade some progress here\nIf anyone needs an urgent fix, please use this -\nIs Kubernetes 1.13 or 1.14 going to get updated with the fixed grpc lib? Or it's going to 1.15 only?\nWe are planning to backport the fix to etcd 3.3 after etcd 3.4 release. Then, kubernetes can pick up the latest etcd 3.3.\nCan we please wrap up the backport (to etcd 3.3) within the next week or two? Please see timeline for 1.16 ( ) We need to have sufficient soak time for this change in k/k.\nSure. The earliest we can do is August 13, 2019. need gRPC 1.23 to include upstream bug fixes need a day or two to rewrite 3.3 client balancer Today, we do etcd 3.4 release code freeze, which means we will start running functional (failure injection) tests + kubemark to test new etcd client changes including If anything changes, we will post updates here.\nThanks a ton !\nHi, is this the same issue: I have 3 control planes. On every control plane are one etcd-server. The apiserver can only connect to one etcd-server. If I enter more then one etcd server in the kube- I have Warnings in the log: ] grpc: addrConn.createTransport failed to connect to {172.22.75.8:2379 0 <nil}. Err :connection error: desc = \"transport: authentication handshake f ailed: x509: certificate is valid for 172.22.75.8, 127.0.0.1, ::1, 172.22.75.8, not 172.22.75.7\". Reconnecting... The certificates are correct. I have this check with curl -v --cert apiserver-etcd- --key apiserver-etcd- --cacert ...\na quick clarification, is August 13, 2019 the earliest day the fix going to land on 3.3? thx\nYes\nUpdate: is out. Bumping up gRPC in etcd master branch, in addition to Go runtime upgrade !topic/golang-announce/65QixT3tcmg. Once tests look good, we will start working on 3.3 backports.\nhas been released with all the fixes. Please try. Once tests look good in the next few days, I will release . Update: has been released.\ni just kicked off a WIP on the k/k CI -\nSome tests in k8s may break. We need pass with the new etcd client. Please take a look at\nIs there a hotfix for v1.15?\n- see\nThe issue is still there with Kubernetes 1.14.6 and etcd 3.3.15. Any changes to kubernetes libs or code needed to tackle this issue?\n- see\nhmm, if there are changes required in k8s source code, why has this issue got closed? I'll try to build a custom image and test, though. Thanks.\nthis will be resolved in 1.16 in the changes required (upgrading etcd client libraries) were too invasive to backport to patch releases\nthat leaves people on 1.13-1.15 without proper H/A? I think this issue deserves to be fixed in the three supported releases of Kubernetes. The hotfix mentioned here looks simple enough to be , but you are saying the proper fix will require much more. So this is all sorta obscure and confusing to the community, IMO. Not that I am complaining, don't get me wrong, it's just I think everyone would welcome some clarity on this issue. Maybe document it somewhere and provide some workarounds for the people who are still on v1.13.-1.15? Because right now, bring down the first etcd member and oops, api is not working, cluster is not working.\nagree on documenting the issue at the very least, and recommending any possible workarounds any feedback on ^?\nI will work on the documentation.\njust built custom 1.14 image from the release-1.14 branch and patched that file and it's so much better now when I bring down the first etcd node down. Now I am confused, if the fix is so simple, why does it have to wait until 1.16?\nThe fix requires upgrading from v1.7.5 to v1.23.0, which is quite a big change :0\nI am adding \"Known issue\" section to etcd docs here\nsee what said :)\nSome days ago I had opened this issue , but reading this one I think it is related. Can anyone take a look at my issue and see it they are related and if so I can close it.\nI believe I am running into this issue or at least something similar. I see the etcd going up and down and the api server. This cluster was created with . The logs from etcd show the following. The logs from the api server show the following. Is this the same?\nThere are claims here that the bug is solved, but I am seeing evidence of it not being solved in our cluster: Are we absolutely sure the etcd client fix made it onto the release? I am testing v1.6.2.\nThat bug is not fixed yet. The only fix was for IP address-only connections, not those using DNS names like this. We are waiting on for what will probably be Kubernetes version 1.16.3. The workaround I'm using today is to replace my etcd server certificates with ones that use a wildcard SAN for the members in the subdomain, rather than including the given machine's DNS name as a SAN. So far, it works.\nIt was fixed for IP addresses, but not DNS names (DNS name issue is tracked in ). Additionally, part of the fix regressed IPv6 address handling (). See These two issues have been resolved in master, and is open to pick them to 1.16 (targeting 1.16.3)\nAh... thank you and that explains it. Cheers!\nWould you please explain how to change the SAN on the etcd certificates?\nI generate these certificates myself using , so it's a matter of revising the arguments passed for .\nTried with Kubernetes 1.15.3 and with 1.16.2 but its not working with neither. This is not fixed even for IP addresses:\nI have similar observation as with IP addresses: etcd 3.3.15 kubernetes 1.16.2 kube-apiserver log:\nYou should be fine with 1.16.2 and Etcd 3.3.15 now. I managed to get 3 masters running.\nIn my case, apiserver has been repeating warnings about connecting to external etcd cluster with tls, log snippets as follows My environment: HA kubernetes 1.15.5 cluster made by kubeadm Etcd 3.3.10 cluster with three members but i am not sure whether my issue is releated with grpc. any answer will be apprecicated\nCOMPLETE BACKUP AND RESTORE PROCEDURE FOR ETCD NOTE: Check that in the file \"\" there is the port with the address configured like this below : --listen-client-urls=https://127.0.0.1:2379,https://.......2379 1) Backup ETCDCTLAPI=3 etcdctl --endpoints=https://[127.0.0.1]:2379 -- -- -- snapshot save /tmp/snapshot-pre- NOTE: etcdctl is a command normally found on the master 2) Restore ( for restore the parameter of \"--initial-cluster-token\" you can call it whatever you want ) ETCDCTLAPI=3 etcdctl --endpoints=https://[127.0.0.1]:2379 -- --name=master -- -- --data-dir /var/lib/etcd-from-backup --initial-cluster=master=https://127.0.0.1:2380 --initial-cluster-token etcd-cluster-1 --initial-advertise-peer-urls=https://127.0.0.1:2380 snapshot restore /tmp/snapshot-pre- 3) Change the parameters in the following file vi --data-dir=/var/lib/etcd-from-backup ## Update --data-dir to use new target location (put in the previous restore command) --initial-cluster-token=etcd-cluster-1 ## (put in the previous restore command) volumeMounts: mountPath: /var/lib/etcd-from-backup ## changes with the path set in the previous restore command name: etcd-data hostPath: path: /var/lib/etcd-from-backup ## changes with the path set in the previous restore command type: DirectoryOrCreate name: etcd-data 4) Check the kubernetes environment after the changes See if the container process is back on docker ps -a | grep etcd see if the cluster members have been recreated ETCDCTL_API=3 etcdctl member list -- -- -- --endpoints=127.0.0.1:2379 see if pods, deployments and services have been recreated kubectl get pods,svc,deployments\n\"\"\"\" INSTALL KUBERNETES WITH KUBEADM \"\"\"\" !!! CHECK ALL INSTALLATION PREREQUISITES BEFORE INSTALLING kubernetes ---- !!! prerequisites: Check if it is already installed by running these commands on master nodes: kubectl kubeadm prerequisites: check which version of linux you have: --cat /etc/os-release prerequisites: Letting iptables see bridged traffic prerequisites: Check required ports prerequisites: install docker on all nodes if not already installed -- docker installation: sudo -i apt-get update && apt-get install -y  apt-transport-https ca-certificates curl software-properties-common gnupg2 curl -fsSL | apt-key add - add-apt-repository  \"deb [arch=amd64]  $(lsbrelease -cs)  stable\" apt-get update && apt-get install -y docker-ce=5:19.03.113-0ubuntu-(lsbrelease -cs) cat <<EOF { \"exec-opts\": [\"native.cgroupdriver=systemd\"], \"log-driver\": \"json-file\", \"log-opts\": { \"max-size\": \"100m\" }, \"storage-driver\": \"overlay2\" } EOF mkdir -p /etc/systemd/system/docker.service.d systemctl daemon-reload systemctl restart docker per vedere se docker è attivo: systemctl status docker.service ############################end pre-requisites############################################## INSTALL KUBERNETES Return to the manual -- and install the following components: kubeadm: the command to bootstrap the cluster. kubelet: the component that runs on all of the machines in your cluster and does things like starting pods and containers. kubectl: the command line util to talk to your cluster. I report below the steps of the manual, do it on each node: sudo apt-get update && sudo apt-get install -y apt-transport-https curl curl -s | sudo apt-key add - cat <<EOF | sudo tee deb kubernetes-xenial main EOF sudo apt-get update sudo apt-get install -y kubelet kubeadm kubectl sudo apt-mark hold kubelet kubeadm kubectl systemctl daemon-reload systemctl restart kubelet Vedere la versione di kubeadm --kubeadm version -o short Vedere la versione di kubelet --kubelet --version we won't install \"Configure cgroup driver\" as you do when you don't have docker installed Go to the bottom of the link page mentioned above \"What's next\": will take you to the following link -- Give this command only on the master: do it with normal user kubeadm init Once installed, copy the output that appears and create the directories as follows on the master only: do it with normal user mkdir -p $ sudo cp -i $ sudo chown (id -g) $ generate the token on the master --\"kubeadm token create --print-join-command\" and then copy the output to all workers Still on the master, give the following command: kubectl get nodes ## you will see that the nodes are not active since the network for the nodes and pods has not been installed ENABLE THE NETWORK: user root: Enable the following file /proc/sys/net/bridge/bridge-nf-call-iptables to \"1\" for all CNI PLUGINS !!! by running this command below on all nodes: sysctl Install the network on the master with the normal user kubectl apply -f \"$(kubectl version tr -d 'n')\" Check again if the nodes are in the running state watch kubectl get nodes cat | grep -i image ## see the etcd version on master, then download it with the website instructions below wget -q --show-progress --https-only --timestamping \"\" tar -xvf etcd-v3.4.13-linux- sudo mv etcd-v3.4.13-linux-amd64/etcd /usr/local/bin/ cd /usr/local/bin/ chown root:root etcd check that everything is working ETCDCTLAPI=3 etcdctl member list --endpoints=https://127.0.0.1:2379 -- -- --\nInstallation of the METRICS SERVER on a MASTER without minkube copy the link of the zip file 1) wget 2) Enter on directory \"assignment....\" and subdirectory \"metrics-server\" 3) kubectl create -f deploy/1.8+/ 4) wait 5) kubectl top nodes && kubectl top pods\nDeploying flannel network manually Flannel can be to any existing Kubernetes cluster though it's simplest to add flannel before any pods using the pod network have been started. For Kubernetes v1.17+ kubectl apply -f 1) ## 2) Initialize the master using the following command --kubeadm init --pod-network-cidr=192.168.0.0/16 3) Execute the following commands to configure kubectl (also returned by kubeadm init) mkdir -p $ sudo cp -i $ sudo chown (id -g) $ 4) Install the Tigera Calico operator and custom resource definitions kubectl create -f 5) Install Calico by creating the necessary custom resource kubectl create -f Note: Before creating this manifest, read its contents and make sure its settings are correct for your environment. For example, you may need to change the default IP pool CIDR to match your pod network CIDR 6) Confirm that all of the pods are running with the following command. watch kubectl get pods -n calico-system Note: The Tigera operator installs resources in the calico-system namespace. Other install methods may use the kube-system namespace instead 7) Remove the taints on the master so that you can schedule pods on it kubectl taint nodes --all node- 8) Confirm that you now have a node in your cluster with the following command. kubectl get nodes -o wide It should return something like the following. NAME STATUS ROLES AGE VERSION INTERNAL-IP EXTERNAL-IP OS-IMAGE KERNEL-VERSION CONTAINER-RUNTIME <your-hostnameReady master 52m v1.12.2 10.128.0.28 <noneUbuntu 18.04.1 LTS 4.15.0-1023-gcp docker://18.6.1 About installing calicoctl calicoctl allows you to create, read, update, and delete Calico objects from the command line. Calico objects are stored in one of two datastores, either etcd or Kubernetes. The choice of datastore is determined at the time Calico is installed. Typically for Kubernetes installations the Kubernetes datastore is the default. You can run calicoctl on any host with network access to the Calico datastore as either a binary or a container. For step-by-step instructions, refer to the section that corresponds to your desired deployment. Installing calicoctl as a Kubernetes pod Use the YAML that matches your datastore type to deploy the calicoctl container to your nodes. etcd kubectl apply -f Kubernetes API datastore kubectl apply -f Note: You can also view the YAML in a new tab. You can then run commands using kubectl as shown below. kubectl exec -ti -n kube-system calicoctl -- /calicoctl get profiles -o wide An example response follows. NAME TAGS kns.default kns.default We recommend setting an alias as follows. alias calicoctl=\"kubectl exec -i -n kube-system calicoctl -- /calicoctl\" Note: In order to use the calicoctl alias when reading manifests, redirect the file into stdin, for example: calicoctl create -f - <\nwe are using kops. can someone help me on this showing errors when checking systemctl status kubelet Jan 07 00:42:11 ip-172-50-2-100 kubelet[5299]: E0107 00:42:11. 5299 ] node \"ip-172-50-2-100.ec2.internal\" not found Jan 07 00:42:12 ip-172-50-2-100 kubelet[5299]: E0107 00:42:12. 5299 ] node \"ip-172-50-2-100.ec2.internal\" not found Jan 07 00:42:12 ip-172-50-2-100 kubelet[5299]: W0107 00:42:12. 5299 ] Failed to create summary reader for \"\": none of the resources are being tracked."}
{"_id":"q-en-kubernetes-070b1f10e9f042955ccb3f17037c22d0865cbfab9fc47c6e87e376857ce1e5bc","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks!--What happened: I've bootstrapped a cluster with kubeadm on a banana pi, it is up and running. Upon restarting, some process is slow that kube-apiserver does not boot up, exits with last lines: What you expected to happen: Expected the kube-apiserver to start up. How to reproduce it (as minimally and precisely as possible): On a banana pi (A20) bootstrap a cluster with kubeadm, add kube-router cni, and then restart kubelet. Anything else we need to know?: The version below contains a proposed fix, that is why the GitVersion is dirty. Environment: Kubernetes version Client Version: version.Info{Major:\"1\", Minor:\"13\", GitVersion:\"v1.13.1\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2018-12-13T10:39:04Z\", GoVersion:\"go1.11.2\", Compiler:\"gc\", Platform:\"linux/arm\"} Server Version: version.Info{Major:\"1\", Minor:\"13+\", GitVersion:\"v1.13.2-beta.0.1+-dirty\", GitCommit:\"\", GitTreeState:\"dirty\", BuildDate:\"2018-12-18T19:10:37Z\", GoVersion:\"go1.11.2\", Compiler:\"gc\", Platform:\"linux/arm\"} Cloud provider or hardware configuration: Banana pi m1 plus OS (e.g. from /etc/os-release): PRETTYNAME=\"Debian GNU/Linux 9 (stretch)\" NAME=\"Debian GNU/Linux\" VERSIONID=\"9\" VERSION=\"9 (stretch)\" ID=debian HOMEURL=\"\" SUPPORTURL=\"\" BUGREPORTURL=\"\" Kernel (e.g. ): Linux k8s-master 4.9.0-8-armmp SMP Debian 4.9.130-2 (2018-10-27) armv7l GNU/Linux Install tools: kubeadm Others: Seems that in AddSystemPriorityClasses() adds a function which uses wait.Poll. On line 95 it logs 'Retrying...', howewer the return statement tells wait.Poll() to stop polling. <!-- DO NOT EDIT BELOW THIS LINE --/kind bug\n/sig api-machinery\n/cc\nIf we solve it, we want a piece of pie :)"}
{"_id":"q-en-kubernetes-97c93cd4f69f0ca95556290743929c975d2e8a4e675cf0eed6639f87f9f062e7","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks!--What happened: fails with . What you expected to happen: The suggest that should work for multiple resources in a directory when running . should've been applied to all resources under following was omitted in and I wasn't able to find a discussion regarding this omission in other issues. How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): 1.10.7 Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others: <!-- DO NOT EDIT BELOW THIS LINE --/kind bug\n/sig cli\n/sig apps\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale /remove-lifecycle rotten\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-testing, kubernetes/test-infra and/or . /close\nClosing this issue. <detailsIn response to : Rotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/reopen /remove-lifecycle rotten I agree with the assessment that this is a bug the , , and all work with spec files with multiple resources in them. does not. All of the commands would additionally benefit from only operating on supported resource types for rollout (Deployment, DaemonSet) from the spec file, but that's really a sperate enhancement. Client kubectl 1.16.6 Server EKS 1.14.9 How Status Behaves: How Pause Behaves: How Resume Behaves: How History Behaves:\nYou can't reopen an issue/PR unless you authored it or you are a collaborator. <detailsIn response to : /reopen /remove-lifecycle rotten Server EKS 1.14.9 kubectl rollout status -f .ClusterServicecore- -n kube-system error: rollout status is only supported on individual resources and resource collections - 6 resources were found kubectl rollout pause -f .ClusterServicecore- -n kube-system paused error: serviceaccounts \"coredns\" pausing is not supported error: \"system:coredns\" pausing is not supported error: \"system:coredns\" pausing is not supported error: configmaps \"coredns\" pausing is not supported error: services \"kube-dns\" pausing is not supported kubectl rollout resume -f .dc-infra-k8s-infrastructureClusterServicecore- -n kube-system resumed error: serviceaccounts \"coredns\" resuming is not supported error: \"system:coredns\" resuming is not supported error: \"system:coredns\" resuming is not supported error: configmaps \"coredns\" resuming is not supported error: services \"kube-dns\" resuming is not supported Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nCould you reopen this? If not I guess I'll file a new issue.\n/reopen\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nWhat is needed to move this forward?\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale poke poke\nThere appears to be a PR that was closed without merging Strangely no changes are recorded in the PR\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle stale\n/remove-lifecycle stale\nAny updates on this?\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\nThe Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle rotten [1]:\nThe Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Reopen this issue or PR with Mark this issue or PR as fresh with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /close [1]:\nClosing this issue. <detailsIn response to : The Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closedReopen this issue or PR with Mark this issue or PR as fresh with Offer to help out with [Issue Triage][1] Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/reopen /remove-lifecycle rotten\nReopened this issue. <detailsIn response to : /reopen /remove-lifecycle rotten Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-241230dec6808f63cf55f116d1eb748b0b624425daff266ca78de443c3f550fd","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks!--What happened: kubectl create -f ceph- and result is ok now i want to resize pv from 1G to 2G, edit kubectl describe pvc pvc-expand-test What you expected to happen: How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: ceph storageclass yaml: ceph secret: Environment: Kubernetes version (use ): Client Version: version.Info{Major:\"1\", Minor:\"11\", GitVersion:\"v1.11.5\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2018-11-26T14:41:50Z\", GoVersion:\"go1.10.3\", Compiler:\"gc\", Platform:\"linux/amd64\"} Server Version: version.Info{Major:\"1\", Minor:\"11\", GitVersion:\"v1.11.5\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2018-11-26T14:31:35Z\", GoVersion:\"go1.10.3\", Compiler:\"gc\", Platform:\"linux/amd64\"} Cloud provider or hardware configuration:OS (e.g. from /etc/os-release): CentOS Linux release 7.4.1708 (Core) Kernel (e.g. ): Linux m01 3.10.0-693.el7.x8664 SMP Tue Aug 22 21:09:27 UTC 2017 x8664 x8664 x8664 GNU/Linux Install tools: kubeadm Others: ceph version 13.2.2 () mimic (stable) <!-- DO NOT EDIT BELOW THIS LINE --/kind bug\nReiterating the mentions to trigger a notification: <detailsIn response to : Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nWhat's the version of your ceph cluster? k8s uses the command to get rbd image's size, and it expects the size unit in the output is . However, in your environment the unit is :\nceph version 13.2.2 () mimic (stable)\nIt seems that the output format is different between different ceph versions. I will try to fix it.\n/assign\nyour PR has been merged to master branch, and when will Kubernetes v1.11 fix this bug?\nSorry for the late, I've send cherry-pick PRs:\nIt seems this change is not working in some cases: I think this is due to two concurrent conditions: 1- ceph client still complain on stderr about not being able to load config file even if there is no need as a key is supplied and the result is (hence the output on stderr of ). This happens not only in hyperkube. 2- it seems we capture both stdout and stderr from the execution of the command (please note the output in the error from ) and then there is then a problem trying to unmarshall the JSON (to be clear, the unuseful warning is on stderr, the json output on stdout). I am able to have a workaround creating empty files for and so that the ceph client does not output anymore but I feel it would be much better to parse only stdout. A reference on the source code for Thanks edit:typo\ncould you tell where did you put the empty and ceph.keyring files? On the node running kubelet?\nInside the pod of the as the rbd resize is executed by this component.\nI had the same problem, I ceph. conf and ceph. Keyring and solved it, but I think k8s got the default user of CEPH before going back to ceph. conf and ceph. Keyring. The following code validates my conjecture: //num 672 Ultimately, I want k8s to get those two files through configurable CEPH parameters. What do you think?\nas far as I know you do not need to configure those two files if you use a key. That is still an hypothesis as in the version of Ceph we are using this switch seems to be undocumented: Anyway the warning can be safely ignored and I do not see why we should parse when there is no error\nI two files with empty content before, no error; later I a soft connection called and ceph. Keyring (pointing to the ceph configuration file and key ring generated by my cluster), but ceph did not report an error. . I think ceph only needs two files to exist and doesn't care about its contents. Fortunately, the current measures can solve the problem.\nthis matches my experience (see above, I created the empty files too and it gives no error) but another possibility is just not parse which I find \"cleaner\" as there is no error in this case even from the error code, and even what we get is just a warning.\nSame issue as here. The warnings about the missing etc files break the JSON output. Adding empty files in the container fixes resizing.\nShouldn't this be reopened until the bug is fixed? Creating the two files also \"solves\" the issue for us.\n/reopen cc\nReopened this issue. <detailsIn response to : /reopen cc Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nwhich version of ceph cluster is in use in your setup ?\nWe are running a ceph cluster with . kube-controller-manager is running which internally uses . Also just to prevent any confusion: I've used to resize the PVC, not rancher.\nThis also seems related to .\nWe are running\nAre there plans to resolve this issue?\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-testing, kubernetes/test-infra and/or . /close\nClosing this issue. <detailsIn response to : Rotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nAnswering myself - apparently it has been fixed in k8s 1.20:"}
{"_id":"q-en-kubernetes-383cf749ae60682e75f7da4e2ae8f3b081daf11c30e73c6db187e7b8fd18b6a4","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks!--What happened: In our stress testing, we found the deadlock issue in PLEG relist() for healthy checking and Kubelet syncLoop() The goroutine stack trace is as following We did some investigation, and found the following issue. The main loop for event processing in Kubelet will check if there are some runtimeErrors: if yes, wait until runtime back to normal; if no, it will call syncLoopIteration to read and process events. And the runtime healthChecks depends on the But in the relist() impl, it will send the pod lifecycle event to eventChannel which may introduce deadlock when eventChannel is full. And in the above code, the eventChannel will not be consumed if the runtime is not healthy. In some situation, the eventChannel may full, and the relist() will hang in sending new pod lifecycle event. And the runtime status will not be able to recovered. It is related to issues What you expected to happen: No such issue How to reproduce it (as minimally and precisely as possible): It is not easy to reproduce, it happens after few days stress testing (scaling the deployment repeatedly with high load). Anything else we need to know?: Environment: Kubernetes version (use ): 1.11.5 Cloud provider or hardware configuration: Bare metal OS (e.g. from /etc/os-release): CentOS 7.4 Kernel (e.g. ): Linux bjc-qtt-k8s-node-010 3.10.0-693.2.2.el7.x8664 SMP Tue Sep 12 22:26:13 UTC 2017 x8664 x8664 x8664 GNU/Linux Install tools: kubeadm Others: <!-- DO NOT EDIT BELOW THIS LINE --/kind bug\n/sig node\nSure, thank you We had some internal fix to break the deadlock, but I am wondering if there are good fix for such kind of issue.\nWhen pleg do relist, can we invoke to judge runtime is ok, if not, not do relist\nI submit a PR\nThanks for finding the issue! The issue description makes sense to me. I think PLEG is a core component, we should try our best to unblock it. The event channel becoming full seems more like an issue in the event consumer side, which shouldn't block PLEG. One solution is to queue the events in PLEG and send later, but that doesn't fundamentally solve the problem, if we don't set a limit on the event number we queue, it seems like a potential memory leak to me; if we set a limit, we may eventually hit this problem again. Another solution is to add a retry for sending event in PLEG, and set a bit in PLEG in between the retry which tells that PLEG is not blocking, it is trying to send out the event. And in the function of PLEG, it returns healthy if the bit is set. Actually, even if we fail to send the event and discard it, since we have already updated the pod cache, the periodical pod resync should eventually pick it up. This may delay pod sync for up to . But if the channel is full, it means that either 1) the syncloop is blocked for some reason, or 2) the node is under heavy loads. In both cases, even the event is successfully sent, the pod sync will likely be delayed anyhow. We can: 1) Enlarge the event channel size to reduce the chance that the channel becoming full. Currently it is arbitrarily set as . 2) Discard the event in PLEG if the channel is already full. In the 3 options above, I'm fine with either 2 or 3. Option 2 doesn't have the 1min pod sync delay, but a bit ad-hoc; option 3 is more generic, but may delay pod sync. There might be other options. /cc Any thoughts?\nIf PLEG is already blocked by a full event channel, this doesn't help right?\ncc/ This looks like the one production issue you mentioned to me a while back.\nof options enumerated, option 3 was preferred for me.\nThanks for reporting the issue! what was the load on the node that triggered this condition? If the node wasn't function at all, and the events slowly built up over time, bumping the size of the channel won't help. This should work, and doesn't complicate the logic in PLEG more, so it's my preferred option too. We need some unit tests, and ideally, a stress test, as mentioned in the sig-node meeting.\nAgree, giving a proper size for event channel is tricky as the scale also grows with time. We will ship fix as option 3 () for 1.14. Thanks for summarizing all these up as sig meeting! /assign\nIf possible, could you share in how large the stress is? It helps understand the issue and kubelet stress handling more. :)\nSure, our product team ( ) will input the stress test data after removing sensitive information in follow up comments.\nThanks for reply. Following is our testcase to reproduce this issue: The test cluster has 5 node on aliyun ECS(like EC2). Three 4C8G machine as master node, Two 96C384G as worker node. Firstly, We create 2 deployment running on the worker node. One deployment is http server pods. Another is siege pods to produce network stress to http server deployment. Those stress test deployments produce 50 CPU Core load on worker node. Then we create a job frequently create/delete 20 pod every 5 seconds. After about 1 day of stess test, The worker node will become NotReady. Dig the pprof goroutine stacktrace found the channel block in .\nYeah, I can see how frequent container state change can cause this issue. The number is very useful, thanks a lot!\ncloud you please help to share how to dump the goroutines stack trace of kubelet like what you pasted in your comment. thanks a lot.\nThis issue is fixed by\nusing pprof is a answer to my question thank you anyway."}
{"_id":"q-en-kubernetes-1507c5b6fb86cc25e2e0b848a10eb015abb35982e7e95b1f65c735cf0636b7a7","text":"Related: On Windows, kubelet reports only the cumulative for cpu usage. On Linux, kubelet reports both and the instantaneous . Although In the future, most components/monitoring pipelines will switch to using the cumulative metrics, a few features still depend on the instantaneous metrics today. Examples are: , HPA/VPA. Kubelet should report this metrics on Windows to support the existing features. /cc who explained to me what features depend on the cpu metrics. /cc /cc\n- is this something that's clear enough in the CRI layer that you can implement it, or do we need to fix that first?\n/assign\ndo you have any notes or a doc you can point to that better describes what features depend on which metrics? It would be good for us to share and see if anything other than is broken\nThe horizontal and vertical pod autoscalers also depend on the\nDocumentation on metrics API and metrics server (i.e., the aggregator): The metrics-server pod keeps spewing errors like this: Running : I haven't verified whether pod autoscaling works in my cluster, but I assume they won't since metrics server seems to discard the metrics from the windows pods. I believe the implementation is in kubelet (for the docker integration), so this should not need CRI changes.\nI think this is generally an issue for non-Docker CRIs (not necessarily Windows-exclusive). The kubelet source code suggests that these metrics will be gathered from the CRI (if the CRI is not Docker). The struct that kubelet expects returned from the ContainerStats RPC does not include usageNanoCores. Does this mean that, until this is resolved, custom CRIs are unable to satisfactorily report metrics to the 'metrics-server'?\n- are you able to work on this?\nthat is a good point. For CRI runtimes, the kubelet should calculate the average over 10s to produce this metric.\nGood point. We designed CRI around core metrics API, but some CRI runtimes have integrated with cadvisor directly to fill the gap while we wait for core metrics API. I think it's reasonable for kubelet to do this on behalf of all CRI runtimes. I spoke to offline, and this would probably require no more than 2 days of work. If sig-windows (e.g., can pick this up, that'd be great. has agreed to review the changes (thanks!) If not, we should clearly document that this is not supported in Windows GA for 1.14.\nthanks for the update. we will wait to hear back from\nCurrently, is got from docker's stats API on Windows. From docker stats API , only is available, hence is not set on Windows: Do you know any way to get instantaneous for Windows containers?\nI have no idea. any idea?\nI think this is broader than just a Windows issue. Officially supported tools such as the kubernetes metrics-server expect the field, but custom CRIs that conform to the interfaces defined in the package are not given the liberty to return this value.\nsuggested previously to cache the stats temporarily and compute the stats using the diff. David, could you chime in on how this could be done? Yes, this was mentioned in the previous comments too. There are plans to transition those components off this field, but it's WIP. We are looking for a workaround in kubelet for now.\nmy biggest worry is that if we say this impacts the horizonal POD autoscaler, how is that passing in our end2end and conformance tests?\nDo you test HPA? It's not part of conformance, and I doubt this is included in windows testing.\nAs to how this could be done, in and we could just add a in the struct, which we assign the previous CpuUsage metric for each container. Then, when we need to calculate the rate, we pass the previous CpuUsage along with the current one to , and calculate a rate based on the timestamp and values.\nListPodCPUAndMemoryStats() we could just add a map[string]*runtimeapi.CpuUsage in the criStatsProvider struct I've done some tinkering on this myself and I originally thought that it would be as simple as that. It gets much more complicated when you have to work in between the existing cAdvisor-based logic. I assume the intent is to cut it out entirely eventually, but the information passed through the CRI doesn't even come close to being so comprehensive. So far I've found that this approach fixes the problem for the \"/stats/summary\" api, but not necessarily for the other apis. In particular, \"/stats/container\" seems to be completely based on cAdvisor, being unimpacted by the changes we made to those two functions. Some monitoring tools and metrics scrapers rely on the \"/stats/container\" api for their metrics, which means it is insufficient to update one api without doing the other.\nWe don't need to fill in everything in the summary API, just \"usageNanoCores\". The \"/stats/summary\" api is the one used by the metrics server, which is why we care about it in this case. The \"/stats/container\" isn't something we need to implement to support this.\nwe a couple of tickets to validate if HPA for Windows containers works. My guess is that it works for memory and not for CPU metrics because of this issue. If this is complicated to fix, i wonder if waiting for the CRI containerD changes in windows would be the best option for supporting this.\ncontainerd and cri-o bypass this issue by leveraging cadvisor as a temporary solution. I don't think that'd work for windows.\nSorry the for late response. HPA is actually working for Windows containers. e.g. in my cluster (v1.12.5), cpu-based HPA is working as expected: Which version of kubernetes are you referring for HPA issues?\nare you running heapster, or the metrics server to serve HPA? Heapster uses the cumulative metric, but the metrics-server uses the 10 average rate metric.\nmetrics-server is used in my cluster.\nSee .\n/milestone v1.14"}
{"_id":"q-en-kubernetes-935ea67b78d5ac4a8e0a1fe5c9682e71278884227fb73fad39175c2d6e99cb79","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks!--What happened: I configured Kubelet to run as a service on Windows, with a setting to restart on failure (). I noticed that it restarts even when I explicitly stop the service with , even though the documentation states the following (I left it at , the default): What you expected to happen: I expected the Kubelet to only restart when killed unexpectedly, per above docs. Anything else we need to know?: I think the problem might be due to this part of our service control integration, but I'm not really sure: It looks like we send and then , but the official golang example sends and just returns without exiting. (). Is this difference meaningful? Kubernetes version (use ): 1.13.2 Cloud provider or hardware configuration: GCE OS (e.g. from /etc/os-release): Windows /cc /sig node /sig windows\nThe same should apply to kube-proxy as well.\ni believe you implemented this functionality. can you please take a look? cc:"}
{"_id":"q-en-kubernetes-e5983eb5353bfa835212290605fb5b3b707694894ffe6172cb57ad01176f5242","text":"We merged to allow configurable volume limits for Cinder. But proposed adding such a limit to as well, among other things this will enable us to support node specific limits more easily - /sig openstack /sig storage /sig scheduling\n/assign\ncc\n/kind feature\nThis will require updating in-tree Openstack cloudprovider code. are we onboard with that?\nWhat would be the extent of the modification? It must not break current compatibility, and the in-tree provider is scheduled for removal this year. I'm ok with it as long as existing users aren't broken by it.\nit should be a change into which would add a new configurable field. It would be backward compatible but would allow admins to configure different limits on a node if they must.\n/milestone v1.14 We're entering and this looks relevant to the 1.14 release. Please if I am incorrect"}
{"_id":"q-en-kubernetes-86018c36895e88236f8affe1fbcc4289b1f69fa7ada8133fe78f6fcfbb669c1b","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks!--What happened: When running e2e conformance tests against a cluster using a public HTTPS endpoint for the apiserver, you get the following errors: What you expected to happen: The API server endpoint is so the e2e tests should use and connect on port instead of . How to reproduce it (as minimally and precisely as possible): Run e2e conformance tests against a cluster with the apiserver exposed on a public endpoint with a public CA. Anything else we need to know?: When looking at the code here: It looks like the tlsConfig is not correctly derived from the URL. Diving into I see that this uses deep down: And this seem to return unless the kubeconfig has a CA or other explicit TLS config configured: I believe if this check also checked the of the URL, then it would be correct: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g. from /etc/os-release): Kernel (e.g. ): Install tools: Others:\n/sig testing"}
{"_id":"q-en-kubernetes-51dc025e0fc024f74328e1390039c5664d3d280fc91e5c3f8fc5cb56fd3c2ba9","text":"We've noticed a strange behavior of the scheduler in GCE 5K node scalability tests. There were periods of time, lasting about 6s, when scheduler didn't schedule any new pods. It's ruining our 5s pod-startup-latency SLO. See the example below, pod density-latency-pod-2325-dgqrv created on 09:08:59, scheduler didn't schedule it until 09:09:06. An interesting pattern can be found in the around that time. Before 09:09:00 scheduler schedules ~5 pods / s. Then it goes silent for about 6s and at 09:09:06 it outputs about 30 \"bound successfully\" statements and continues scheduling pods with ~5 pods / s pace. We checked api-server logs for that silent period of time and it turned out that scheduler was doing something, it was constantly sending GET and PUT request for a metrics-server and heapster pod it was trying to schedule. ! We dug deeper and found that scheduler was actually trying to schedule these two pods for over 12h, below is the plot of number of calls per minute that scheduler issued to api-server for these two pods: ! As we can see there are long periods of time when scheduler issues over 1500 requests per minute. That seems strangely a lot given that the requests are only for the two pods. During the whole run scheduler issued over 400K GETs and over 200K PUTs for these two pods. ! We figured out that scheduler wasn't able to schedule these two pods because there was not enough resources in the cluster. While it's understandable that scheduler wasn't able to schedule the pods, I'd expect scheduler to not spam the system with so many retries, which, as we can see in this example, can potentially block / slow-down scheduling of other pods. I've increased the scheduler log verbosity level in , we should have more logs from the tomorrow's run. Will post them here once available.\n/sig scheduling\nReiterating the mentions to trigger a notification: <detailsIn response to : /sig scheduling </details\n/assign for triage\nWhat is the version of your cluster? If you are using the HEAD, we have a back-off mechanism to the scheduler to delay rescheduling unschedulable pods. It starts with 1 second delay and doubles it every time the pod is not schedulable up to 10 seconds.\nIt's from head, so apparently something doesn't work...\nYes, we're using head. The test run on Jan 21 and the cluster was built from head on the same day (commit to be more precise). As you can see in the logs it doesn't look like scheduler is using any back-off mechanism. Definitely not the one you described, it issues requests to api server much more often than 1 per second per pod. !\nAn update on more logs - we won't have more logs. We tried increasing scheduler's log level to v4 and it resulted in the log growing by few orders of magnitude and it was clogged with mostly useless stuff, e.g. it logged \"Node gce-scale-cluster-minion-group-XXX is a potential node for preemption.\" over 250 000 000 times. We reverted the log level change in If more scheduler logs are needed to debug this we probably should add some new, more useful, ones.\ncc/\n/assign I can look into this next Tuesday\nFriendly ping, any updates?\nWhoops. Thanks for the ping. Log level v4 should not be needed. Although the log line you called out should probably get pushed into V9... First off, that's some really great debugging. Thanks for the detailed report. If the scheduler is getting into tight busy loops, that could in theory starve the scheduling thread, causing the observed 6 second pause in scheduling. The bug is in  A pod only gets marked for backoff if . And can increment very quickly in a high churn cluster. This would result in a pod being retried in a tight loop. One (unrelated) observation that stood out to me was this: This is being printed by the package and shows a trace with no steps logged to it, which means that we are hitting one of these four error cases in . Most likely, the PVC for those heapster & metrics-server pods are in a bad/unexpected state.\nI think fixes the root problem here. If we eliminate the API server spam, the main scheduling thread hopefully will not get blocked for 6 second periods. Are you able to test with this patch?\nThanks Jonathan! I've reserved a scale-testing slot for Thursday (Feb 21) to test the fix. Will update this issue once I have the results.\nI reproduced the issue (non-schedulable heapster pod due to not enough resources) on a 5K node cluster. Then I run the load test, first on the version without the fix and the on the version with the fix. Below are the graphs of the number of scheduler requests (for the unschedulable pod) per minute. Without fix: ! With fix: ! As we can see on the graphs, the fix definitely helps, without it there were minutes were scheduler was calling api server up to 2K times per minute (2K QPM). With the fix it didn't exceed 100 QPM. But even with the fix the average QPM is ~60 (per single pod) which is 1 query per second per pod on average. is average 60 QPM (1 QPS) reasonable here or should we expect something less from the rate-limiting mechanism?\nBy default the scheduler rate limit is 20 qps for all scheduler operations. Default backoff delay for unschedulable pods is 1 second and goes exponentially higher every time a pod is retried, but does not exceed 10 seconds. This 10 second max backoff time also means that if a pod is not retried for 10 seconds, for example because there are many other pods in front of it in the queue, its backoff time is reset to 1 second. So the next time it is retried and is found unschedulable it will be backed off for 1 second. I am basically trying to say that 1 qps is not unreasonable. We also an optimization () that should lower qps for unschedulable pods further down. This last PR got merged only a few hours ago. With that PR merged, I would expect to see even lower qps as the scheduler will not update pod status every time it retries scheduling a pod.\nThanks Bobby! That makes sense. In that case I believe we can close this issue once merges.\n/milestone v1.14 We're entering and this looks relevant to the 1.14 release. Please if I am incorrect\nThanks for fixing!"}
{"_id":"q-en-kubernetes-f7fa63d0e351875894e6a6c8d3a66cd26b64ee40c5c6f01ca8059a8d5316e1a7","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks!--What happened: test/e2enode/perf tests are referring to adm only images, this need to be replaced with multi-arch fat manifest images. ref:\n/assign /sig node"}
{"_id":"q-en-kubernetes-a0af38dfdb94314d8001228e6fa629f59fe482692018e968c1705d3de507abad","text":"Kubelet and System-reserve are implemented on Windows along with pod/container reserves which can help prevent overprovisioning, but we still need to test OOM behavior. We need a test case that will: kubelet & system reserve pods until node is full something that will take up system ram so that OOM is triggered /sig windows /milestone v1.14\nI have started to look into this test case and have a few questions: What is the expected behavior we are testing for? What does \"run something that will take up system ram\" mean? Does that mean run a pod that uses up memory or run a system service that uses memory? After setting the reserves and as mentioned in we should be able to schedule pods until full and then once we trigger OOM (either by pod or system service) kubelet should continue running becuase of the reserves? Do the two settings need to be tested independently or they should be tested together? I found that seems to be doing something similar but it is in node-e2e. Is sig-windows running node-e2e tests currently?\n/assign\nGitHub didn't allow me to assign the following users: jsturtevant. Note that only and repo collaborators can be assigned and that issues/PRs can only have 10 assignees at the same time. For more information please see <detailsIn response to : /assign Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nAfter researching this further and talking with I found that the does test the allocatable memory but uses CGroups to set reserves to test the reported vs actual. As windows does not have CGroups this test will not port as is. There is and the is not implemented so will not work either. The proposed testing solution would be two parts: kubelet & system reserve and verify allocatable is reporting properly. Schedule pods with known memory consumption values and verify the allocatable is correclty updated.Schedule until full and verify scheduling is not allowed. kubelet & system reserve and verify allocatable is reporting properly. Schedule Schedule until full and verify scheduling is not into machine and run process that consumes memory and verify kubelet and pods remain running. There is an open question on the second test as to whether kubelet and will be have properly. If it does not we will submit test 1 and open an issue to resolve 2.\nI would suggest adding a third case that would test the behavior when excessive memory pressure is exerted on the node from the pods. This is perhaps a more common use case than 2 since the pod workloads are more likely to exhibit greater resource variability than node system processes. It would also carry over as a test for when/if node allocatable enforcement get implemented for windows nodes. Additionally, to complement this work would it be suitable to test windows pod memory limit enforcement? Or is this tested elsewhere?\nAgreed this is a more common case. I am not sure that it is. Is this tested on linux somewhere?\nI am seeing strange numbers reported when performing a test case. I scheduled a few pods to test memory limits and then schedule a pod with no memory limit and a process that will consume all of the available memory 500mb at a time: After the pod consumes all available memory, When I compare Node Allocatable/Node usage/Pod Usage I am not seeing the numbers add up properly: My node is staying online and kubelet is reporting no memory pressure. The total node memory consumption is over the Allocatable memory. What is the expected behavior and what should the test look like?\nI believe node allocatable will always just be capacity minus reserves (system and kubelet) - So in this scenario it looks like you have and/or set? The memory % on the node is relative to the allocatable amount, and still seems to be under the total memory capacity - are you sure that all available memory has been consumed by your oom pod? The node should be tainted with memory-pressure when a hard or soft eviction threshold is met (). What do you have set for and ?\nThanks for the clarification. I currently have these settings: And on the machine I am seeing most memory being used: If I understand correctly, I start to see memory-pressure once I get into the setting above. In that case I should start to see pods get evicted, so for a valid test I could adjust the values to be easier to hit then test to see that pods get evicted properly which would be similar the . I am going to adjust a few of the settings and manually test that this is working then see what it will take to port the PriorityMemoryEviction, which currently uses cgroups which won't work on windows. This combined with testing that Memory Allocatable is being properly set should give us a good test coverage for windows on this scenario.\nI performed manual testing and found that even when memory was used up on the machine the \"Memory Pressure\" notification was not fired. After looking deeper I found is not implemented for windows. The , which uses cgroups, provides the . This means even though is being it isn't used. On Windows, Node allocatable when used with limits and requests does prevent the node from over scheduling. I believe the the test and  (this only shows cpu and I could not find a test that demonstrated memory) demonstrate this works. But this doesn't stop a pod from consuming all the memory on the node and since eviction (hard/soft/node allocatable) is not implemented on windows this means a node can end up resource starved. I think the only valid test to be written for this scenario, as it stands now, is validating that Node Allocatable is being reported properly. Beyond that there is no functionality that can be validated around OOM behavior for the windows node.\nYeah, that sounds to be in agreement with what I've observed on the windows nodes that I run. To prevent resource starvation (as a workaround), would setting mem/cpu requests equal to mem/cpu limits for all windows pods work? I think the request / limit values get passed down into the docker daemon when it starts a container, so is it possible that docker could enforce the memory limit?\nYes, setting limits==requests would help the scheduler not over provision the nodes. I think the best way to approach this particular issue for 1.14 is to make sure docs call out what works/doesn't work and give guidance to on how to prevent resource starvation on nodes. Is there a place in the docs that would make sense for this information? The call outs would be: , , and flags do update Node Allocatable is not implemented and are not implemented Condition is not implemented in Windows There are no OOM actions taken by kubelet As a work around to previous, and should always be equal to prevent memory resource starvation as this allows the scheduler to to use limits and Kubelet running on the windows node does not have memory restrictions. and system-reserve does not set limits on kubelet or processes running the host. This means kubelet or a process on the host could cause memory resource starvation outside the and scheduler\nthank you for the thorough investigation"}
{"_id":"q-en-kubernetes-c275e37ec6def532051cbfafd9b1ce5ce881f1f7d3b02dd8dc5a2079f1d8218d","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks!--What happened: We created a LoadBalancer service with type 'nlb' with two listeners forwarding traffic on two different ports. The service was created correctly along with the target groups and was functionning correctly. Around 20 hours later, no traffic was flowing from clients to any k8s nodes. While troubleshooting we discovered that the k8s node instance security group was lacking the input rules allowing traffic to the node ports for this LB. We deleted the service and recreated it. The service started working normally again, but we found out that the two target groups were missing their k8s tags. The logs contain the following error: What you expected to happen: We expected the NLB target groups to have k8s tags, so that k8s could remember it created them. How to reproduce it (as minimally and precisely as possible): Create a k8s service with NLB: Add a simple deployment that is matched by this service. Anything else we need to know?: Looking at AWS cloudtrail we could see the following error event: So the API request was called with two Resource ARNs, which is confirmed by the code: Since we have two target groups, there are two resources to tag. Except that the AWS API while allowing to send more than one resource ARNs doesn't allow it returning a validation error. This could be solved by doing one per target groups. Note that this can't happen on a Classic LB because there's no target groups. I believe that those missing tags on the target groups are the root cause of the deletion of the security group rules allowing traffic to reach the nodeports, with this (hypothetical and unproved) scenario: 8s creates the target groups, the NLB, the listeners, the sg input rules 8s fails to tag the target groups, but doesn't rollback anything a subsequent convergence loops, k8s notices that there are two k8s-managed Security Group input rules, but those don't match any Target Groups or match Target Groups that are not managed by k8s (because they don't have the owner tags) 8s deletes those rules Unfortunately we don't see the rule deletion in the controller manager logs, but it appears in Cloudtrail: Let me know if you have any questions, Environment: Kubernetes version: 1.10.7 Cloud provider or hardware configuration: AWS OS (e.g. from /etc/os-release): CoreOS 1967.3.0 Kernel (e.g. ): Linux 4.14.88-coreos Install tools: home grown Others:\nReiterating the mentions to trigger a notification: <detailsIn response to : Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig aws /assign\nHi, the root cause of your problem should already be solved by It's cherry-picked into old versions and available as v1.11.7, v1.12.5 and v1.13.3(hasn't released yet)\n, Thanks, you're most certainly right about the root cause of the issue being . It indeed looks exactly like what we've experienced. Still there's the problem of target groups not being tagged."}
{"_id":"q-en-kubernetes-53de30f3668d2bf6e518e4cf30fa599a726d298c8d97549d5fe70c2f806840e5","text":"What happened: What you expected to happen: Human readable error output. Anything else we need to know?: The same is happening when in duplicate key while creating a with binaries. !!! The error output will be all the of the binary itself !!! Environment: Kubernetes version (use ):\n/sig cli\n/cc"}
{"_id":"q-en-kubernetes-661e7682be627df950d21ecc4319690b966f02b280949d4d768b9e03cdf0463c","text":"What happened: We currently unregister the CSI driver if an error happens while registering it. Then, if another error happens, we return only the unregistration error. The problem is that the user will only see the unregistration error if they check the registrar logs. This gives the impression that it never tried to register the driver. What you expected to happen: The user should see the registration error instead. How to reproduce it (as minimally and precisely as possible): Try to install a CSI driver in a cluster where CSINodeInfo and/or CSIDriver CRDs are not created. Check registrar logs /kind bug /sig storage CC\n/assign"}
{"_id":"q-en-kubernetes-7b703b4b6a73f14b289c5d247020e8c0696bad824ee6eb51b840401e012aa9c8","text":"Which jobs are failing: ci-kubernetes-e2e-kops-aws pull-kubernetes-e2e-kops-aws Which test(s) are failing: [sig-storage] In-tree Volumes [Driver: local][LocalVolumeType: block] [Testpattern: Pre-provisioned PV (xfs)] volumes should allow exec of files on the volume [sig-storage] In-tree Volumes [Driver: local][LocalVolumeType: block] [Testpattern: Pre-provisioned PV (xfs)] volumes should be mountable Since when has it been failing: Don't know.. as the job was down for a while due to Testgrid link: Reason for failure: To be found.. Anything else we need to know: This issue is a blocker for turning on the kops-aws presubmit, see - /cc /kind failing-test /sig storage\ncc\nBy looking at the log this line is suspicious: Both testcases share similar error message. And that comes from when doing I guess its some mount client missing from the AMI so that the mount library can't find it.\nLooks like the test needs to skip XFS on environments that do not support XFS\nThis is where the environment is checked to skip xfs:\nThanks Michelle! - Could you comment if we need to skip it in aws tests?\nLooks like this was broken by - shall we revert? cc\nActually looking at the GCE tests for kops it was broken before . I'm wondering if we assumed xfs back in\n/priority critical-urgent this is blocker for kops currently example of failing build\ndo you have idea could break this?\nYes, in , I tests for local volume which use . Because , PR is not blocked by gce e2e test. hi, Quick way to fix is to remove from supported types, I can submit a PR for this.\nAlternatively, we can skip for environments which do not support XFS . What do you think?\nI do not know what is the best solution in long term but in short term we should get kops e2e tests working. So if removing xfs from supported types is fastest solution - i would go with that. After that tests could be fixed how those should be. Could you do that PR and remove from supported types?\nHard-coding the knowledge which clusters or tests support XFS (because that's what it is - a dependency on a kernel feature, right?) in the test source code is just bound to fail eventually. What if the presence of xfs support depends on the \"gci\" version? This line doesn't cover that: I would prefer to mark tests which depend on XFS kernel support with or something like it (no idea how descriptive feature tags need to be) and then enable those test cases only in jobs where the target cluster is known to have XFS. If you extend that check as a short-term solution, then please file an issue about it and link to it next to the check. /cc\nAgreeing with here... Isn't there a way to check if the kernel we test for/ against is supporting it more directly? Building a huge list of whats compatible and what's not sounds not perfect to me.\nLong term we can enhance the check to ssh into a random node and see if the binary exists. Short term let's just remove xfs testing for local volumes.\nIf these two are newly tests, agree to skip (or remove) xfs testing to unblock PRs.\nassuming all these in-tree tests will be less important over the course of the year as CSI moves forward... so should we assume that the XFS tests will never come back . ?\nThe local volume driver plan is to remain in-tree for the medium-term future. I think we should continue to add the environment check for xfs so we can add it back."}
{"_id":"q-en-kubernetes-405578bf01cf308ae5a6d2cd350037ef43dd399ce60b57a2c3f47a45e605c332","text":"<!-- Please only use this template for submitting enhancement requests --What would you like to be : /sig testing /sig windows In the the PR we're adding a test case for pulling a Windows image from dockerhub (. However, there are no known public Windows images on , so the only option would be to upload it ourselves. Our proposal is to have it uploaded to the using Windows images). The proposed image manifest list to be uploaded is built from [1], which is currently hosted on dockerhub [2] and is currently used for testing Kubernetes with Windows nodes [3]. The manifest list contains the images: e2eteam/busybox:1.29-1803 (for Windows Server 1803) e2eteam/busybox:1.29-1809 (for Windows Server 2019) [1] [2] [3] Why is this needed: The Windows images will be used for the Windows equivalent test case for the following test:\n/assign\n/area conformance\nIt is my contention that we really only want one prod repo and that should be folded into the main repo. The authenticated repo should be for testing authenticated pulls and nothing else. To that end, I propose that we start with this. Does the image need to be named or tagged or something to indicate OS? How about * ? Is that too deep? Does it need Windows in the path or should we put it in the tag? Or can we have a multi-os manifest list? Also, is busybox the right image for testing, or can we do something even more minimal (to address the \"what f there's a busybox CVE\" question)?\nYes, (all images to date have used ), so I'd prefer that route over yet another name or tag. I'd also prefer something other than busybox if possible. I originally created because there were so many images in and it was hard to tell which of them were production images, which were test images, and which had been completely abandoned. Also, only a small number of Googlers could push to this repo since it contained all the production images. To address the first issue, I decided that all e2e test images would be sourced from the directory. To address the second issue, I created a new project with a wider ACL (though still currently limited to Googlers). All that said, given the automation we're building for the main GCR repo, it'd probably be fine to fold these images back in.\nThe image promoter means that a single repo should be BETTER managed than individual repos - we will have a PR trail of who put images there and why.\nWe should probably broaden the discussion to non-busybox Windows test images too. (related: ). Do we have a lower bar for test images? If so, having a separate repo (or making it clear that the images are test only) does have clear benefits.\nDo we want a lower bar? I don't think we do.\nNobody wants a lower bar until they want to add/update the image quickly for a new test, or fix a test failure to unblock the release. If we do effectively hold the same bar today, this shouldn't be a problem.\ni agree. also we don't need schema v2. multi-os, multi-arch image here. just (e.g.) a nanoserver image that satisfies a Windows 1803 and Windows 1809 host. a nanoserver image will obviously imply Windows only. for a test called \"should be able to pull image from [WindowsOnly]\": all we need is to upload this image to [x] and i guess we need to decide on if we want this test to be / passing for Windows Containers GA.\nAnd indeed, a manifest list with nanoserver images could be an alternative (considering that the ) (we did try to find a command which works on both OSes, and ultimately found curl, but not all Linux images contain curl, check PR ).\nan official nanoserver image is a more secure download. do we need to build a nanoserver manifest list or are there official images that we can use? i would like to bring a topic here related to pause. why are all the pause images for Windows that i found 1.1 GB or more? we definitely need the absolute minimal pause container uploaded in a location where users can download it and use that location in the guide. AFAIK, 100 MB is as low as we can go for a native Windows container, but that's still much better... i brought this topic in the other PR. we don't need a command that works on both OSes. if is the command to use on Windows containers only, so be it.\nAnd also, my image? It might be out-of-date.\nwe need to start planning how Windows containers can used be easily on Windows nodes. the pause image should be pre-built, small and easy to pull by deployment tools, such as kops, kubeadm. that is much better, but still a bit too much? can we go down to ~100MB? this image size was reported by a user in a k/k issue recently: my assumption is that the user was building the image himself.\ni am ok to have that conversation, but please file a ticket and let's discuss that in sig-windows\neveryone, thank you for your help and responses. trying to bring back this conversation to the original request. what would you like us to do in order for the few folks from google that have permissions to upload the windows images to be used for testing.\nI am happy to help publish this via the alias, which I believe is the correct long-term path. We need an image - preferably not one as prone to CVEs as busybox historically is. Hence a suggestion for \"pause\" or something equivalently useless and tiny (if it is tiny). Then, we need a path. First, we'll be setting precedent for test images, so I need input on that. I propose as the root. Objections from test folks? Second, do we want to squirrel this away as a Windwows-specific binary, or should we not name it in any particular way, so that we can pull the same image for all tests via a multi-os, multi-arch manifest list (which can be done later). E.g. vs How about * ? Is that too deep? Does it need Windows in the path or should we put it in the tag? Michael Michael <:\nSo, I'll sumarize the previous discussions. The proposed images are: The image currently named image tags:. based on: respectively. image is currently used in Windows E2E tests and should already exist on the nodes, so download should be minimal. The image. a manifest list needs to be created containing the tags: The image is used to tag the image, which is used by kubelet, so it should already exist on the nodes. smaller in size (~350MB). the test should create pods with the command . I'm fine with either one, although I think people will tend towards the nanoserver image, and in fairness, I think it will be the easiest one to maintain long-term. AFAIK, Microsoft updates their container images periodically, so, we should do the same for the image. If not, the test using the image will then have to pull the entire image (350 MB is ok, IMO). E.g. vs Well, a manifest list can contain both Linux and Windows images, So, if the busybox image is picked, the Linux busybox image can be included in the manifest as well. Ideally, most images work on both OSes (with some exceptions, like the nanoserver or alpine images), so I don't think there's a need to put windows in the path. And if a new root path is being for new test images, I would like to ask that images mustn't container slashes in their names (e.g.:). Docker doesn't support images with slashes in their names (see ). Even if all the images on eventually support Windows, it would still be a good idea to be able to use dockerhub for testing purposes.\nWe need an image - preferably not one as prone to CVEs as busybox WHOAH. That's a 2GB image. For testing. I am in favor of this just based on the size (I was thinking something like kuberneets' pause, which is 240 kB. Over 1000 times smaller. We have different comprehensions of what \"nano\" means. 350 MB is not OK. It's acceptable if we have no other choice, but honestly we should create our own that is as small as humanly possible. Then, we need a path. First, we'll be setting precedent for test images, so I need input on that. I propose as the root. Objections from test folks? I don't have enough context on e2e future, so I'd like some input from someone who does. Second, do we want to squirrel this away as a Windwows-specific binary, or should we not name it in any particular way, so that we can pull the same image for all tests via a multi-os, multi-arch manifest list (which can be done later). E.g. vs I agree (except for busybox, which is a known vector for contaigion). Having a multi-os, multi-arch \"serve_hostname\" or something equally dumb seems ideal. Docker HUB doesn't. Docker itself does, and GCR does. If you're using dockerhub for testing, you have to change \"\" to \"e2eteam/whatever\", right? So what does it matter if the GCR path is longer? That said, I would be OK to have a single dir if that is what test-team wants. It just seems likely to get messy eventually.\nYes, dockerhub, I forgot the \"hub\" :) And yes, that's what we're doing and it works for the most part. But, if you take a look here [1], basically, the root is , which, as you know, is not accepted by dockerhub. So, the proposal was to not have \"image names\" with slashes in them (e.g.: in this case). [1]\nI see. I think it would be OK to ask that images don't use subdirs, but the e2eRegistry variable could. e.g. // Docker hub e2eRegistry = \"e2eteam\" // GCR e2eRegistry = \"\" ResourceController = Config{e2eRegistry, \"resource-consumer-controller\", \"1.0\"} That seems workable, though it doesn't answer the question of the actual path in GCR :) Claudiu Belu <:\nThat seems workable, though it doesn't answer the question of the Yep, exactly right. :)\nwhat are the next steps on this ticket to get the images to\nI am still waiting for which image we are going to use, and for any objections to \"\" as a root. Michael Michael <:\nWell, I think we are in agreement that we should use the nanoserver images. As for the root repo, I have no objections.\ncan you please explicitly give all the images to upload in an authoritative list? i am hoping we can upload all of these by monday EOD.\nSure. As we've discussed previously, the images are:\nno manifest list image that supports both 1809 and 1803? are we OK for only having 1809 for that \"pull from GCR\" test?\nYeah, I meant those images should be included in a single manifest list (something like or something), like we've previously discussed. As for the second question, AFAIK, the GCE Windows CI is using Windows Server 1803 at the moment, so it would be nice to have the new test running on that too.\nsee above for the two images we want to push. we also need to repush this image resource-consumer/controller:1.0 to avoid the \"/\" as resource-consumer-controller cc:\ni pulled these images on Windows: size of each is roughly ~350MB. and then saved them to TAR and uploaded them here: to go lower than that we might have hack custom images that trim, for instance a bunch of DLLs that we don't need (for e.g. a minimal pause image), however such a custom image will just require yet another unique download if the users are already basing their workload containers on official images. also for such hacks we need someone to verify if the nanoserver LICENSE allows that. i'm pretty sure it doesn't.\nno hacks please. we need to upload the images as they officially come from microsoft. :)\n/milestone v1.14 We're entering and this looks relevant to the 1.14 release. Please if I am incorrect\nThe slash thing is kinda irrelevant to me, it's just a limitation of Docker Hub as pointed out\ni understand that. but can you make forward progress in testing that autoscale ticket without this work?\ni think you have the two images to upload. let us know when this is complete and thank you so much!\nThe tars shouldn't be necessary. The nanoserver images include 2 layers that are type and should not be pushed. here's an example manifest from the bazel repo. you could probably use the existing rulesdocker to push nanoserver with no changes if you're building and pushing images with bazel\nwe can't follow this process due to Windows licensing. i will hide your post so that it is not followed by accident.\njust to clarify - you can reference those in manifest lists, and docker push / pull them without actually pushing the layer contents to a manifest v2 complaint registry. That's the right way to do it. That separates out user-owned files in the upper layers from those in the base layers that are always served from a Microsoft server due to licensing restrictions\nOh ok then. Keeping in mind the fact that the nanoserver image is a manifest itself (for amd64 and arm) and that you cannot reference another manifest list in your manifest list, then the the actual images which should be to the manifest list are: 1803 doesn't have multiarch.\nthe post is still quoted in your response. best to edit that as well? ok, looks like the licensing of these image is slightly more complicated than expected... i hope a tutorial is not required when pushing to GCR. :)\nWhat is the status of this now? I am THOROUGHLY confused\nme too according to here and here and as far as i understood, we only have to push the JSON files that hold layer references, but not the actual layers because this violates the licensing of the nanoserver images.\nI'll post explicit steps using , etc.\na PowerShell window as an administrator Docker-CE 18.09.2 experimental mode for the Docker CLI Here's the copy & paste steps to pull the existing images, push them to the new registry, and create a manifest. Change $destinationRepo to be something like or your private repo name. The pushes will be nearly instantaneous because they're manifest-only pushes. All layer data is skipped. Now, you can inspect the manifest and it will show multiple images with different os.version And last, push it\nHere's the example repo I made on When things are done this way, the original Windows image is still served from Microsoft as required by the license. If I were to add unique layers on top of this image - that content would be pushed to the registry.\nPushed to my staging, will try to get the promoter running ASAP\nhmm, it looks like our image promoter is not quite ready for multi-platform yet. Let me try pushing to\nyeah, you might be bumping into\nSomeone test that pull?\ndoes that mean no Windows images can live on GCR? Am I misreading it?\nI believe GCR supports it, but some of our tooling (which uses google/containerregistry) may not. OTOH I think tool is just using directly, so it might be unaffected.\n!\nAwesome, so that seems to work for now. I guess the next step is to FROM that in your own containers.\nor just use it in your tests. Do we close this for now, then?\nSGTM, leaving the close call to\nThis looks right. I did a pull on a 1809 machine and the hashes match. is this enough to finish that test case, or are more images needed? We definitely need to get the images into a CI for 1.15. See /\nthis should be enough in terms of the test case.\n^ FYI, there are now Windows test images at GCR. thanks !\nAwesomeness, I've made a PR adding the testcase using this new image. This issue will close automatically once that merges. Test works for me. One question remains though. Do we have to have a Windows equivalent for the , so it fails on Windows. If we have to add a Windows equivalent for that too, then the nanoserver manifest list should also be pushed to the registry as well. Anyways. thanks for the help, everyone!\nare you asking if can push the nanoserver manifest to the authenticated-image-pulling registry as well? Tim?\nYep. It is a conformance test, so it would be nice to be able to cover it as well.\nany chance you push the nanoserver manifest to the authenticated-image-pulling registry as well? thanks for all your help! once you do the above, we can lose this issue\nTry \"\"\nthanks it works! !\nplease, close the issue once you check this.\nthanks Tim! you rock!\ncan you please close this issue if we are done?\nThe issue is supposed to close when the PR merges. Closing anyways."}
{"_id":"q-en-kubernetes-b07c64c89884a118fbbe56fe942df2a7c1d31098a995d286838c2be33aa2c12c","text":"<!-- Please only use this template for submitting reports about failing tests in Kubernetes CI jobs --Which jobs are failing: ci-kubernetes-e2e-gci-gce-serial Which test(s) are failing: LocalVolumeType: gce-localssd-scsi-fs Since when has it been failing: 2/20 Testgrid link: Reason for failure: unsure yet /assign\nAfter , there are still two failing tests: For the first one, looks like reconstruction is not correctly implemented for local PVs: to check if it's ok for mounter to be initialized with an empty path the panic in the operation_generator\ncc Unsure why it's failing only for this local pv type (real disk vs loopback block) I have not investigated the 2nd test yet\n/priority critical-urgent /milestone v1.14 We're entering and this looks relevant to the 1.14 release. Please if I am incorrect\n/assign\nhi, checked , after , \"should be mountable\" has been fixed. failing test is \"should unmount if pod is force deleted while kubelet is down [Disruptive][Slow]\" and it fails for all kinds of local volume types\nThis is because local volume cannot be reconstructed for now (localPlugin.NewMounter requires field is set), so is going to . But uses to , and , so it panics here. I'm not sure how to support reconstruction for local volumes correctly yet. But we can fix this panic first, then reconciler will clean up the existing mount points if the volume is no longer needed. IIUC, this will fix this failing test. After this is fixed, if necessary we can revisit how to reconstruct local volumes correctly.\nI see similar failures in and too; are they connected? cc\nyes they should be similar\nneed to keep track of this to see if failures in and also clear up.\nThe tests are still throwing s in (it's the job there though). I've opened to track the failures in 1.14-blocking. They also seem to not have run since the 24th, I'll look into that in the meantime"}
{"_id":"q-en-kubernetes-078231f9f9e8868f802d2e92f1c77dfbb38893fcb91b9b00f83e92fd0556107c","text":"In , we check mount on first level directory in pod volume directories. It is true for many volume plugins, but not for CSI plugin which . cc\n/assign\nPardon my ignorance of kubelet source code, but this looks very much similar to an issue we have been running to many times recently, when developing a CSI plugin. The problem is that when CSI volume publish is called and the mount point is not propagated from the CSI plugin pod (either missing bidirectional flag in yaml or not supported by underlaying k8s cluster implementation), the kubelet nevertheless starts a pod using the volume. The volume is not there, so the app writes to root filesystem in fact, leading to all other kinds of problems - i.e. losing data, filling up root fs. IMHO what kubelet should do is to check that the mounted filesystem is indeed visible after publish call and if not it should fail the volume attach request. If this ticket was fixed, would it solve the issue we are seeing or should I open a new ticket? thanks\nI think we can potentially add that check but we need to make sure first it can work with ephemeral CSI drivers. Ephemeral drivers would have to \"stage\" their directories somewhere, like /var/lib/kubelet/plugins/<csi driver/, and then bind mount it to the publish path. cc\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nHmm... That is a good question. Does csi require a mount to be made or only that the specified directory be created? Will have to go read through the spec...\n/remove-lifecycle stale\nThe only place in the spec I could find was this comment: // For volumes with an access type of mount, the SP SHALL place the // mounted directory at target_path. So I think having Kubelet enforce that check would be ok with the spec, and a good thing to do.\ncc CSI spec says that the target path has to be a mount. Shall we go ahead and enforce it then? Are we going to break drivers that may not be doing this?\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\n/remove-lifecycle rotten\nThe issue mentioned here involves two functions . Now with PR , the issue in CheckVolumeExistenceOperation is solved by checking mount path instead of volume path. The only issue here is for getMountedVolumePathListFromDisk which is used during pod deletion. This function was introduced by PR . Now I am questioning whether we need to check the mount point or just need to check pod dirs exist before deleting the pod. Firstly, for some plugins, there might not be mount point. Secondly, without checking mount point, it means that pod cannot be deleted as long as there are pod volume dirs. In normal case, volume dirs should be cleanedup during volume teardown. For cases that kubelet restarts, we have reconciler try to clean up volume mounts if possible. So we can propose a couple of solutions here mount point check. It solves the problem, the downside is that there might be some rare case that reconciler could not clean up dirs after kubelet restarts. check mount point, and also check the next level \"/mount\" dir. This might still cause problem for some plugins that do not have mounts. cc\nDiscussed offline with For proposal 1., pod will not be deleted from api server if there are still volume dirs exist. So to make sure it works well, we need to have reliable way to clean up dirs. Normally when kubelet restarts, reconciler will either add it into actual state (if reconstruction works) or try to clean it up directly once unless volume is still in desired state. So if reconstruction does not work, and it fails to clean up during reconciler sync, there is no other chance to clean up the dirs. Or if volume was in desired state and removed quickly afterwords, the dir might not get the chance to be cleaned up at all. We are thinking to improve this part, we could use mark volume \"uncertain\" in actual state feature. During reconstruction, we can remove the check for mounted points (since the mount might be gone during node reboots or driver does not have moutns at all). Reconciler could mark those volumes as uncertain (even if it was in desired state) to make sure those volume got cleaned up or remount whenever needed with retries. With this change, I think it is safe to remove the mount point check for pod deletion too.\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale"}
{"_id":"q-en-kubernetes-fc359c75cc34a58c3210bb3c527d0a85f8d779b140e2dcfa4bc6c30ac099a5d6","text":"Which jobs are failing: Tests which exec into containers and alter lifecycles Which test(s) are failing: All new tests introduced for api change which exec into containers New alpha subpathExpr tests [Feature:VolumeSubpathEnvExpansion][NodeAlphaFeature:VolumeSubpathEnvExpansion] Since when has it been failing: 27/February/2019 Testgrid link: https://k8s- Reason for failure: is only set in e2e tests and the binary is not available in node e2e tests Need to restructure and call the API rather than through the framework convenience function Anything else we need to know:\n/sig node /sig storage /assign /cc"}
{"_id":"q-en-kubernetes-e460483476fddc78feaeed229fdeec11e36a5e2de8f3086c3fa32034eef63aaa","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: Running the vSphere cloud provider on a windows node throws the following error: This is because the vm identifier is retrieved with the command - , which assumes that the uuid and serial are the same. On some vmware vm versions, the uuid and serial are not the same. vmware (comment), vmware (comment) What you expected to happen: Kubelet is able to find vm in vCenter from the identifier looked up on the node. or should be used instead to find the vm identifier. How to reproduce it (as minimally and precisely as possible): Issue only occurs on some vm versions, such as 13. Anything else we need to know?: On the linux side, this issue was resolved by switching from looking up the id from to . See Some discussion about this issue here - Environment: Kubernetes version (use ): 14.0-beta.1 Cloud provider or hardware configuration: Vsphere OS (e.g: ): Windows 2019 Kernel (e.g. ): Install tools: Others:\n/sig windows /sig vmware\n/milestone v1.14\nYou must be a member of the GitHub team to set the milestone. If you believe you should be able to issue the /milestone command, please contact your and have them propose you as an additional delegate for this responsibility. <detailsIn response to : /milestone v1.14 Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign"}
{"_id":"q-en-kubernetes-2c9ef8984ebaf899659836b925069334113c910fc47ebe0516a7d659674f044a","text":"xref In 1.16: change the metadata-proxy addon to use change the kube-proxy addon to use change the masq-agent addon to use stop setting , , and in kube-up\nHey I'm trying to start to contribute to kubernetes and this seem to be an easy task. Can I help with this one?"}
{"_id":"q-en-kubernetes-74f759a3150454232b25764fd3379db803b0434665e1f03690071b3af1284be2","text":"have worked since 1.0 as a way to pass basic auth credentials to the API server. In 1.11, these flags were unintentionally removed in Environment: Kubernetes version (use ): kubectl 1.11+ /sig cli /sig auth\n/assign"}
{"_id":"q-en-kubernetes-9ba58a38d1b26cddbe35d0c0365909f33d458f500f4464f9dc0ae31755bbc823","text":"What happened: The rbd plugin use \"rbd info\" error: The messages before and after json result is warning string and no effect to rbd info, but it is effect to parse the output. So the kubernetes will failed: What you expected to happen: Kubernete no failed and errors How to reproduce it (as minimally and precisely as possible): Remove the config file\n/sig storage\nWe are blocked by this issue, can't resize any PVCs\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nThat is another situation that cluster configuation is empty, aslo will return lots of warnning information: Also will trigger expand volume failed. But should we add -c=/dev/null in rbd info command?\nFolks, the pull request that would have fixed this just got closed as it was rotting away. This is an actual problem in production, can we please fix this?\n+1 we are still being plagued by this issue...\n+1 {\"name\":\"kubernetes-dynamic-pvc--147e-422c-9e67-\",\"size\":,\"objects\":2560,\"order\":22,\"objectsize\":,\"blocknameprefix\":\"rbddata.\",\"format\":2,\"features\":[\"layering\"],\"flags\":[]}, invalid character '-' after top-level value\nI solved by exceuting this command\nWow...I was not expecting this to work ... but it did ;) Any insights on why is it working with just the presence of the file? This is very strange... Appreciate any comments..."}
{"_id":"q-en-kubernetes-df940813d484806d91a703a04e4c9e5021c11833ac55de12dcde60eb77fdbe76","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: I recently saw an issue in which the scheduler hit this log line many times: After some digging, it turns out that PodAffinityChecker has a and . The points at the scheduler's cache instance, while the points at a light wrapper around a . In the scheduler, we use the cache to provide a globally consistent view of the cluster, even if that view is slightly out of date. Data read from the core may not have been processed by the scheduler cache yet. Letting data \"skip ahead\" like this can cause stale info to get persisted in the scheduler cache, and other subtle bugs. While the original problem is not likely caused by this issue, I think it needs to be cleaned up to avoid potential bugs down the line. What you expected to happen: Predicates should only read state from the scheduler's cache in order to maintain consistency. How to reproduce it (as minimally and precisely as possible): This was observed in a cluster with high rate of pod creation and node creation/deletion. Anything else we need to know?: Environment: Kubernetes version (use ): 1.12.5 Cloud provider or hardware configuration: GKE with preemptible nodes OS (e.g: ): Kernel (e.g. ): Install tools: Others:\n/sig scheduling\nWe do have this kind of issues, and here is another one:\nI'm not sure the urgency of this fix, but if its not super urgent, I would love to pick it up, if there's an eng more experienced with the scheduler who could give me some guidance :)\nHi I dug into the source code, and find out is used by the scheduler and kubelet. Maybe we could change the internal struct of , but I do not know why it appears in kubelet which seems irrelevant to the scheduler.\nis just a wrapper of corelisters.NodeLister. kubelet can get node info by it.\nI think the point is that the kubernetes scheduler was designed to support single instance at the beginning. Even with the addition of multiple scheduler support, the cache between each scheduler is isolated. So at this stage, we can accept this without a shared state solution. After all, the node information is not actively modified by the kubernetes scheduler.\nSo is it necessary to ensure the consistency by using node lister instead of cache in the scheduler, or we do not need to fix this issue which seems like to caused by inconsistent cache.\nWe've recently hit this issue with the scheduler. It appears that our scheduler somehow gets out of sync with the nodes, and start scheduling pods on tainted nodes. The node's kubelet then (rightfully) refuse to admit the pod. Looking at the scheduler logs, they are flooded with the following : Similar scenario, this is a high churn cluster with thousands of pods scheduled per minute - and nodes being autoscaled. Can we try to get a fix ?\nI'm glad to pick this up if we decide to fix this issue.\nIt would be awesome if you could fix this issue for us, we are currently seeing 100's of pods incorrectly being scheduled onto nodes with taints.\nThe code cited for logging \"Node not found, %v\" is now in :\n/assign\nThanks for the interest on fixing this bug. Sorry I haven't had time to engage on this issue. Feel free to send PRs my way. I'm not 100% certain that this will fix the problem of pods getting scheduled to tainted nodes. How is the taint being applied to the node? Even if the scheduler is internally consistent, there could be a time between node creation and the taint being applied where we send pods to that node."}
{"_id":"q-en-kubernetes-59fd5979bc4c6fa723beda7d679397fe430c94684d86a74380be654288d5fb6a","text":"What happened: I tried to set up bandwidth limiting by booting a cluster with , which by default sets up kubenet. I used the annotations xM xM and then used to test the ingress and egress bandwidths, with no effect. Does anyone know if this feature is still supported in kubenet? And if so what are the steps I need to take to boot a cluster that can use pod annotations to do tc shaping. What you expected to happen: Per traffic shaping should be implemented on Kubenet as of 4/12/2018. Anything else we need to know?: There are a lot of issues regarding traffic shaping not working. None have a concrete answer of how to get it to work on kubenet: Relevant: I believe this is where it was first implemented. Environment: Kubernetes version (use ): 1.11.8 Cloud provider or hardware configuration: AWS CC:\nReiterating the mentions to trigger a notification: <detailsIn response to : Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\ncc\n/triage unresolved I am a bot run by ‍\ncan we put this on sig-network agenda? the discussion item would be: what is the current status of this feature (alpha, beta, GA?) what should we do next with this API?\nI have this to the agenda, will grab some info.\nAFAIK traffic shaping is still supported in kubenet. Does a pod with these annotations come up successfully but has no traffic shaping applied? Or do things fail earlier? If it comes up successfully, could you run ?\nOn GKE we do not officially support this, our default image even don't have tc installed. I tried with ubuntu, the kubenet is failed to parse the tc output. kubenet expect. // filter parent 1: protocol ip pref 1 u32 fh 800::800 order 2048 key ht 800 bkt 0 flowid 1:1 tc return.// filter parent 1: protocol ip pref 1 u32 chain 0 fh 800::800 order 2048 key ht 800 bkt 0 flowid 1:1 notinhw\nIf this issue has been triaged, please comment . Meta: /athenabot mark-triage-reminder I am a bot run by ‍\nDepending on which CNI plugin you use you need to install the bandwidth plugin, see:\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nsince it is in progress /priority critical-urgent /milestone v1.17\n/lifecycle frozen\n/assign\nBug triage for release 1.17 here. Since this issue has been tagged for the milestone, we would like to check it's status. Code freeze is on Nov. 14. Will this issue be resolved before then? If not, we should move it out of the milestone.\nThe fix is still pending in review.\nWill it be merged by this Thursday for v1.17 code freeze?\n/milestone v1.18 (v1.17 code freeze)\nmatching assigned priority on PR\nHello I'm one of the Bug Triage shadows for the 1.18 release. Since this issue is tagged for 1.18, I'd like to check its status. Is this issue intended for milestone 1.18? Friendly reminder: Code freeze is scheduled for March 5th.\nHello Checking in again. Do you think this will make it in for 1.18 release?\nHey Checking in again as this issue is marked as . Are you close to completing this for v1.18 release? Another reminder that code freeze is just around the corner! (March 5th).\nFriendly reminder that code freeze is in 2 days and this issue is marked as\nThe fix got approved by ,should be able to catch the freeze."}
{"_id":"q-en-kubernetes-a7de2980a1c472b0b751243a37311f481487a7c616962eb414e613f79e03aa15","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: I see only amd64 architecture etcd is used for the which will create issues for the other architectures like ppc64le, arm etc. We need to replace this etcd image with the k8s maintained fat manifest image present in\n/sig testing"}
{"_id":"q-en-kubernetes-46ada0f054c26a971685a915238948a98d23847838fa9107921cf22377ffff7b","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: Following the documentation at the , I tried enabling quota for a namespace other than but I was still not able to create a pod with priorityClass in my namespace with following error: Error from server (Forbidden): error when creating \"\": pods \"memory-demo\" is forbidden: pods with system-cluster-critical priorityClass is not permitted in my-namespace namespace What you expected to happen: Expected pod to be created with namespace where I allowed quota. I can see that both ResourceQuota and Priority Plugins are enabled from logs. I0408 17:15:08. ] Loaded 8 mutating admission controller(s) successfully in the following order: NamespaceLifecycle,LimitRanger,ServiceAccount,TaintNodesByCondition,Priority,DefaultTolerationSeconds,DefaultStorageClass,MutatingAdmissionWebhook. I0408 17:15:08. ] Loaded 6 validating admission controller(s) successfully in the following order: LimitRanger,ServiceAccount,Priority,PersistentVolumeClaimResize,ValidatingAdmissionWebhook,ResourceQuota. How to reproduce it (as minimally and precisely as possible): Follow the steps at Anything else we need to know?: I think, the issue is with priorityPlugin where we are rejecting instead of letting resourceQuota plugin to handle the admission when is enabled. Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Others:\n/sig scheduling /cc\nSame happens to me. Correction, it does not allow to use the default permissions outside of kube-system\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nSee also:\n/remove-lifecycle stale\n/assign Per offline discussion with\nSame issue here. We are using namespaces other than kube-system for system pods as it is getting too crowded, therefore we would like to run cluster-node-critical pods in other namespaces.\nWaiting for this to be available myself. No reason my Node Exporter dataset has to run in the namespace instead of my namespace."}
{"_id":"q-en-kubernetes-3fbd6fe4c52deabe3c900459c29c36a04807fdd1453263d11d2244d21f4a4f4a","text":"Public IPs are expensive resources. Given that we're running scalability tests on 5k-node clusters, we are unnecessarily consuming 5k IPs which is a lot. We should switch our tests so that nodes have only private IP addresses. We are going to do that by configuring NAT so that nodes can talk to outside world (e.g. to download image from dockerhub). But that also requires changing some correctness tests which assume public IPs.\nTo summarize recent work: setting up private gce clusters with nat, feature is gated via the env test and dashboard with correctness tests on gce private cluster - https://k8s- 27 tests fail on fixing the tests - should fix majority of them\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale /lifecycle frozen\n- is there anything missing to be done here?\nNo, we can close this one. It was done long time ago. Thanks for catching /close\nClosing this issue. <detailsIn response to : No, we can close this one. It was done long time ago. Thanks for catching Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-e55c6125db876adb50024429a752975ad52abdece13a004554c4c7b387107f66","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: In Azure disk attach/detach function, it uses slice() to store the disk list: And after changed, the original would also be changed, so original azure data disk list is corrupted after attach/detach disk example code as below: What you expected to happen: How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others: /sig azure /assign"}
{"_id":"q-en-kubernetes-aeced5571e452060173b6f51f44cf87df5de4fd407957efdad9199b147e5189c","text":"What happened: We created a Deployment and continuously updated an annotation on it periodically for a week. Eventually, we noticed errors in the controller-manager logs due to the annotation growing too large. Example annotations: controller-manager error: What you expected to happen: that do not change the revision number should not be to the revision history. revision history should be capped at a fixed number of elements, it should not be possible to have unbounded growth for Deployments that are frequently updated. How to reproduce it (as minimally and precisely as possible): a Deployment modify an annotation on it and note that the annotation gets the same number appended every time. Environment: Kubernetes version (use ):\n/sig scalability Not 100% sure where this one fits, I couldn't determine which group owns the Deployment/ReplicaSet controllers, but this is a scalability problem\n/sig apps\nAgree with these, and if the limit is hit, the beginning of the history can be removed.\n/assign"}
{"_id":"q-en-kubernetes-9a119f1570876b2ff6363d763a7f3ffb1db636e92d42d04cbeb12b6cd390d493","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: Docker has a hardcoded 16K limit on log line length, which splits lines longer than that into multiple messages in the log file. in , however, doesn't account for this when is set to and prints the timestamp in front of every entry in the log file instead of every logical line. This leads to random timestamps appearing in the middle of log lines, potentially breaking the applications that parse these logs. What you expected to happen: inserts timestamps once per logical line, so original log data is preserved. Alternatively, allow API to request the logs in a sane format like JSON, which would allow merging the lines back and combining them with the timestamps more intelligently on the application side. How to reproduce it (as minimally and precisely as possible): Create : apiVersion: v1 kind: Pod metadata: name: log-lines spec: containers: name: main image: busybox command: ['sh', '-c', 'while sleep 1; do echo -n ; for i in $(seq 1 ); do echo -n A; done; echo ; done'] Start the pod with Fetch the logs with , or using the API () Observe random timestamps inserted in the middle of the log lines. Environment: Kubernetes version (use ):\nReiterating the mentions to trigger a notification: <detailsIn response to : Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nI will work on it. /sig cli /remove-sig node\nThe logs are being corrupted by the API endpoint in kubelet, so shouldn't it be in scope of SIG node?\nThe log is not corrupted by k8s. The same log could be got using \"docker container logs -t CONTAINERID\" I think you should post your requirement to Docker. Please refer following code Line 50 ~ 52 if config.Timestamps { logLine = append([]byte(msg.Timestamp.Format(jsonmessage.RFC3339NanoFixed)+\" \"), logLine...) }\nkubelet doesn't rely on Docker when reading logs, it reads the files itself (see the code linked in the issue). I don't think Docker would change the forced line splitting (see for example), so this should be fixed in Kubernetes.\nYour are right and kubelet read log file from host such as /var/log/pods/pods/NAMESPACEPOD-NAMExxx/....(JSON format) I will post a PR for it. /remove-sig cli /sig node\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle rotten\nI saw the attempt fixing this but it was scary (echo this comment: ). I suggested a way it may be improved. Also the proposed fix didn't introduce any e2e tests to validate that runtime behavior is not changed and runtimes keep the no-newline-on-splitted-log logic. are you or anybody on this thread are still interested to work on this? Please re-open if so. Meanwhile closing this ahead of bot doing it because the inactivity. /close\nClosing this issue. <detailsIn response to : I saw the attempt fixing this but it was scary (echo this comment: ). I suggested a way it may be improved. Also the proposed fix didn't introduce any e2e tests to validate that runtime behavior is not changed and runtimes keep the no-newline-on-splitted-log logic. Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/reopen\nYou can't reopen an issue/PR unless you authored it or you are a collaborator. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/reopen\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThe Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Reopen this issue or PR with Mark this issue or PR as fresh with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /close [1]:\nClosing this issue. <detailsIn response to : The Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closedReopen this issue or PR with Mark this issue or PR as fresh with Offer to help out with [Issue Triage][1] Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/remove-lifecycle rotten /reopen will fix it.\nReopened this issue. <detailsIn response to : /remove-lifecycle rotten /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/triage accepted\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\nThe Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle rotten [1]:\nThe Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Reopen this issue or PR with Mark this issue or PR as fresh with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /close [1]:\nClosing this issue. <detailsIn response to : The Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closedReopen this issue or PR with Mark this issue or PR as fresh with Offer to help out with [Issue Triage][1] Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/reopen\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nI received a bug report about this issue. The root cause is that the buffer reads up to a new line. If the line is greater than 4096 bytes and '--timestamps' is turrned on the kubelet will write the timestamp and the partial log line. We will need to refactor the ReadLogs function to allow for a partial line read."}
{"_id":"q-en-kubernetes-b7f4673358735529188f1b03c034a17b0033e8e697c9fcb7fdd887d39df32018","text":"We observed in an e2e that if attach is performed on a volume that was detached before from same node, sometimes the mount of the volume (for 2nd pod) could be stuck waiting on invalid devicePath in call. The problem is - when attach for second pod failed with this caused an dangling volume error from AWS cloudprovider(implemented by us). The dangling volume error is supposed to auto-correct any volumes which are attached to a node that it shouldn't be. But in this case - the dangling volume error was triggered by the fact that, volume was not actually attached to the node but because call for 2nd attachment returned stale data (because of AWS eventual consistency). So, dangling volume exception caused volume to be to actual state of world (so as it can be detached later), but this caused volume to be reported to node's Status field. And that caused second pod to assume volume as attached to invalid device path and hence 2nd pod was stuck for 10 minutes waiting for volume to attach on a device path (/dev/xvdbl) which does not exist.\n/sig storage /sig aws cc\nHere is k8s logs that confirm that, did return stale data for volume: See detach succeeds at 19.30.25: and yet: so second DescribeVolume API call is actually returning stale data.\nThere are couple of ways of fixing this: error is used by AWS and Openstack cloudproviders to fix incorrectly attached volumes. We could fix mechanism of dangling volumes, so as when it adds volumes to actual state of the world - it does so without reporting the volume to node's Status. can make a local fix inside AWS so that when we return error, before returning that error we also verify volume attachment by querying the instance. We already have instance information in that call. Verifying volume attachment by and should reduce likelyhood of returning incorrect errors. There is still a chance though that, both API calls would land on a AWS backend which does not yet have updated information.\nexpanding on fix - I am thinking when we add a volume to actual state of the world via DanglingVolume error, then we should add the volume as \"uncertain\". We will update the code to not report \"uncertain\" volumes in node's status and hence in this case volume state will be fixed correctly. If I understand correctly - did not intend to report \"uncertain\" volumes to node's status anyways.\nThis looks like another symptom of . I tried that, it does not always help. Both and may return the same stale data, see IMO, is the right way to go. Unless it break something else, all these AWS corner cases were complicated already and now anything can break.\n/cc"}
{"_id":"q-en-kubernetes-fced6f44684a9037712e6d8dcb8eb686fedae1fb49dee45866c2deaae9c5def3","text":"Example:\n/assign\ncc\nFor all those PD instances, tests are being skipped due to this line: Fixing it by removing that skip as suggested by\n/open until we verify the issue is fixed\n/reopen\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nLooks like diffResources became green on 5/27 /close\nClosing this issue. <detailsIn response to : Looks like diffResources became green on 5/27 Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-4c16480af7ff1ad6fbb1ec54ab8f674a7a003d0df3336908e852158f437832db","text":"/sig api-machinery /kind bug /assign sttts because the controller that maintains the condition runs multiple instances in HA apiserver clusters, the controller must not duel in cases where the validation messages change slightly between releases"}
{"_id":"q-en-kubernetes-d31efbc530ae7084b5d86765399241bd1a139d5f39674537392dc9a4d165abab","text":"What would you like to be : I would like a method of seeing more clear data while tests are running, e.g.: How many tests will be run, have been run, have passed, have failed, etc. Name of the tests passed/failed/skipped Why is this needed: While a run is going we only have the output from stdout to investigate. There is a LOT of noise in those logs which is great for debugging after the fact, but difficult to consume most of the time. you can't tell how many tests have been run or get a sense of how much time is left you cant easily get a list of all tests that have failed so far My proposal Adding a new, custom ginkgo reporter that will post the data to a given URL (passed via a flag) if the URL isn't set, the reporter is a noop and doesn't waste time/resources Use cases: A user runs the tests dumps stdout to a file for debugging later, meanwhile they run a local server that will just echo the status updates it gets. At any point they look at the output they can get good data about what would be run, how far along it is, how successful it has been so far. Using the conformance test image and Sonobuoy, the Sonobuoy worker can act as the server, taking those status updates and reporting them to the aggregator. This would allow Sonobuoy to present that information to users on demand or stream it depending on the context. The work of adding this is pretty minimal and is clearly useful. In the case of Sonobuoy; this has been the most requested feature for a long time showing that people in the community really do want to have a better high-level picture of the progress of these tests. I think emitting the data to an endpoint like this is a flexible implementation that can help any consumer of the tests (not just Sonobuoy).\n/sig testing /area test /area test-framework\nOne notification URL should work fine. Can you give me details on the data model for the status data to be published?\nThat is one, slightly TBD area. The reporter in the test framework has a bunch of data and I'm not sure what we exactly Sonobuoy may want from any given plugin (not just the e2e plugin). I don't want this implementation to be overly Sonobuoy specific or overly e2e specific. Will update when something more clear is pinned down.\nBrought this up in the testing commons meeting and it was agreed that this approach seemed reasonable and that the customReporters seemed like the only way to generate this data. In addition, it was requested that we find a way to also emit this (or some variation) into stdout/logs so that it is easier for a human to navigate those verbose logs.\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale"}
{"_id":"q-en-kubernetes-5218b57270509d317067165fad9442bb61871d38132a1b2739f3aa03a7555910","text":"<!-- Please only use this template for submitting reports about failing tests in Kubernetes CI jobs --Which jobs are failing: ci-kubernetes-e2e-gci-gce-serial Which test(s) are failing: [sig-network] DNS configMap nameserver [IPv4] Forward external name lookup should forward externalname lookup to upstream nameserver [Slow][Serial] Since when has it been failing: Failing since 6/10 at 16:58. . Decent chance that the failure is due to Testgrid link: Reason for failure: /cc /priority critical-urgent /kind failing-test /milestone v1.15 /cc\nThis is blocking the release /cc\nI noticed in the it thinks it's testing 1.16 alpha:\n/assign\nshould fix it. Sorry for the mess-up!"}
{"_id":"q-en-kubernetes-0188767f896fa4ba360906d38e1639a5304d22bac11e53573f06f9147964b0f1","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: run on macOS, all md5 hashes will be stripped from the output and the script will not fail or warn. What you expected to happen: The script should work or detect this issue and fail. Currently it silently ignores not existing on the host. We have some warning logic, but solely based on bash version, if the user upgraded their bash installation but did not install md5sum we don't warn them. How to reproduce it (as minimally and precisely as possible): run the script Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others: /sig testing /priority important-longterm /assign\ncc"}
{"_id":"q-en-kubernetes-c848fc676f3a1ca426b1a80d108f2c0b445722da206d542521feee5e30f39558","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: If a pod's restart policy is \"Never\", and kubelet failed to create the PodSandbox (for whatever reason) once, the pod will be stuck at \"ContainerCreating\" forever. What you expected to happen: Since the user/application containers were never created, kubelet should retry the PodSandbox creation. The fact that the pod remains in a ContainerCreating state while kubelet explicitly does nothing to resolve the situation also confuses the user. How to reproduce it (as minimally and precisely as possible): I don't have an easy way to trigger a podsandbox creation failure off the top of my head, but this should be doable by messing with the CNI binaries/configurations. Anything else we need to know?: Relevant code snippet: Environment: Kubernetes version (use ): Only checked the master branch (1.16), but all supported release branches should have the issue Cloud provider or hardware configuration: not relevant OS (e.g: ): not relevant Kernel (e.g. ): not relevant Install tools: Network plugin and version (if this is a network-related bug): not relevant; though the initial PodSandbox creation failure could be caused by networking issues. Others:\nthanks for posting! I'm looking at the code, and it seems like one fix could be checking if this pod has an associated sandbox - if it does, we want to respect the current behavior. If it doesn't, we want to attempt to create the sandbox. I was thinking something similar to: Thoughts? Is there a simplex fix?\nHi! Thank you for filing this - it matches the behaviour we've been seeing. Am I correct that this is a regression from (), and so only affects Kubernetes =1.13 ?\nYour solution may not work for the one specific failure I checked, where the PodSandbox container was created, but call to the CNI plugin failed (for some reason; could be caused by a race condition). This left us a not-ready PodSandbox with no IP. Probably have to actually determine whether the user containers were created before. Need to check the code closer though.\nThanks. That's probably correct. I saw this in happening in 1.13\nGotcha - thanks! Taking a look at your pr now :)"}
{"_id":"q-en-kubernetes-40786c964398a3eb953cc30e35304f94ad9c19f867b658ee15ff953c1cbb29fe","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: I was trying to write a test in a service I'm working on to catch if an error returns from . When I overrode the result of the fake client with a reactor to return an error I did not get the error back. What you expected to happen: If the reactor is set to return an error, it should come back. How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: This only happens with the fake. Environment: Kubernetes version (use ): 1.14 Cloud provider or hardware configuration: any OS (e.g: ): any Kernel (e.g. ): any Install tools: any Network plugin and version (if this is a network-related bug): n/a Others:\n/sig testing\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\nLet's take a look at this /priority important-longterm /sig api-machinery /remove-sig testing /remove-lifecycle rotten\nIs there any progress?\nI submited a PR to fix this issue. PTAL .\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-testing, kubernetes/test-infra and/or . /close\nClosing this issue. <detailsIn response to : Rotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n^^^ would it be possible to reconsider this PR? I ran into the same issue today.\nthis hasn't been fixed and we recently ran into this issue. Can we /reopen? Happy to bring up to date.\n/reopen still relevant as above\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/remove-lifecycle rotten /triage accepted\n/assign"}
{"_id":"q-en-kubernetes-b3b78c7492a57a794cfd93b08b3e1af496d6cfe0388474966fdc6026178e9914","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: I was using kubeadm to generate cluster ca: The CA generation failed because of 502 bad gateway returned and the response body isn't empty when it's trying to fetch stable version from internet. In this case, the function returns error instead of falling back to local client version: is how kubeadm fetches stable version and is how the error is handled for . Should kubeadm always fall back to client version when there is any internet issue? Otherwise, this sounds defeat the purpose of fallback. /sig cluster-lifecycle What you expected to happen: kubeadm falls back to local client version when there is any internet issue. How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): 1.13.7 Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others:\nI'd like to take this issue. It would be better if you could give some directions. /assign\nrecreate the issue in kubernetes/kubeadm: /close\nClosing this issue. <detailsIn response to : recreate the issue in kubernetes/kubeadm: Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nthe fix will be removing the when failed to and add some unit test if I don't missing anything\nGot it. Thanks."}
{"_id":"q-en-kubernetes-501b5a7c1e536de36a6012ca39d9a25f4076afc73dd3ed2b05c2a3db131c867f","text":"Found this when trying to see which vendored dependencies does not use a proper tag (and is using a random SHA). The dependency is very old and there have been a lot of updates to it: Looks like the However there does not seem to be anyone around to take care of the portworx plugin. Is it time to deprecate it? is anyone using it? /sig storage\ncc ?\n/area code-organization\nHello The portworx volume plugin is definitely in use. The portworx volume plugin doesn't need frequent updates since the plugin acts as a simple client to all provisioning calls. In the same view, the vendored openstorage package doesn't need to be updated too. The server side implementation in Portworx handles all the latest Portworx features.\ncc:\nin that case, can you please update the dependency to a specific tag instead of a random SHA?\n/assign updated the title of the issue as well\nmy request is not for a new release or tag in libopenstorage/openstorage repository. Just to update the SHA we use in our k/k repository to one of the released versions ( ). This is not asking or demanding anything to be done in libopenstorage/openstorage, just that we should update what we use from there (and avoid a random SHA). We are trying to do this for all vendored libraries (one by one … slowly as part of the sig-arch code organization sub project, so it’s not a random request)\n/assign\nGitHub didn't allow me to assign the following users: lpabon. Note that only , repo collaborators and people who have commented on this issue/PR can be assigned. Additionally, issues/PRs can only have 10 assignees at the same time. For more information please see <detailsIn response to : /assign Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nAfter a chat with - apparently this request is coming across the wrong way I am closing this out as it seems to add a over head /close\nClosing this issue. <detailsIn response to : After a chat with - apparently this request is coming across the wrong way /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-5f307f169b135ecbd266f77aba080899ab7b73833d9e7b0c2dd020b7ab0b6680","text":"What would you like to be : The current error message is vague and could be improved. (credit to Why is this needed: Improve contributor experience, make it easier to identify change to be fixed. /cc\n/assign"}
{"_id":"q-en-kubernetes-58099f73b710cc51b23b448af4de10a1c17da2648d684bf9a5995d6c18f2afc1","text":"What happened: Ref , we noticed that there are some unnecessary check around limit in kubelet, specifically at: and is moving this check to outside of the pods loop make it better, but we should be looking at whether this check can exist in an even higher layer, as it doesn't make much sense in either or . cc\n/sig network\n/assign Pending outcome of the discussion here, I would be happy to present a PR\n/triage unresolved Comment when the issue is assessed and confirmed. I am a bot run by vllry. ‍\n/remove-triage unresolved\npoke :-)\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nHow about lifting the check to the beginning of Kubelet#syncLoop (outside the for loop) ? All the events emitted by CheckLimitsForResolvConf are at EventTypeWarning level (not error). The CheckLimitsForResolvConf call in HandlePodUpdates would be removed as well."}
{"_id":"q-en-kubernetes-d5f405b98ac26e68bb52a403345cda18a0b690bbcb8385b5ec06372b302c6e2d","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: Data race in client-go with below sample code: Data race: What you expected to happen: No data race How to reproduce it (as minimally and precisely as possible): Using above given sample code Kubernetes version (use ):Cloud provider or hardware configuration:OS (e.g: ): OSX\n/sig api-machinery\nI'd like to take it. /assign\n/remove-sig api-machinery /sig auth cli\nBug triage for release 1.17 here. Since this issue has been tagged for the milestone, we would like to check it's status. Code freeze is on Nov. 14. Will this issue be resolved before then?\nAnother reminder here that code freeze is in less than two weeks (Nov. 14). We will need to move this issue out of the milestone if isn't going to be resolved before then."}
{"_id":"q-en-kubernetes-2f216f42903306f14320232ee6b2973489c1ab5dd74f1c869c1fdeb0328721fa","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: When I change default hugepage size 2Mi to 1Gi and reboot, kubelet can't update node status What you expected to happen: kubelet can update node info and node status was ready How to reproduce it (as minimally and precisely as possible): change the node condition validate Anything else we need to know?: Environment: Kubernetes version (use ):Cloud provider or hardware configuration:OS (e.g: ):Kernel (e.g. ):Install tools:Network plugin and version (if this is a network-related bug):Others:\n/sig node\nHi, Ref. the error message, it looks like you have allocated hugepages for multiple sizes, . Can you run: to verify? If more than one of the numbers is , that is the reason. This is currently being discussed in , and the reason for implementing the check is here: Imo. this should be supported, and it would only require removing an if statement in order to implement it. I see no reason to limit the node from allocating huge pages with more than one size. Because of how it is implemented, adding support for allowing a pod to request huge pages with multiple sizes isn't trivial, and will require some potentially breaking changes, but that is not relevant to this problem. /cc\nAlso, can you also provide the result of , that will list the supported huge page sizes (both when is set, and when it is not.).\nThanks ! Looks like your kernel only enables huge pages with the default size. My kernel 5.0+ enable all (atleast 2MiB and 1GiB on x8664) by default. When the node booted with defaulthugepagesz=1GB it \"removed support\" the 2MiB one, and becuse kubelet use patch on the node status when connecting to the api server, you ended up with both and , even tho 2Mi wast supported by your machine (the sw, not the hw). Even if the error is removed, this will result in a crash when kubelet tries to write to the 2MiB hugetlb cgroup files. I think using should work, otherwise the only current solution is to remove the node from the cluster, and rejoin afterwords. Ref. it looks like you have sorted out the problem, but I still think this is a valid issue, so don't close it. The correct behaviour for kubelet is to remove unsupported sizes, and only allow use of the supported ones.\n\"Imo. this should be supported, and it would only require removing an if statement in order to implement it. I see no reason to limit the node from allocating huge pages with more than one size.\" Could you fix this one as well? We use 2M pages only for pods but we need 1G pages for system processes. So it was little surprising when k8s doesn't start because two different pages were configured no matter they are really used or not. Thanks!\nI am hitting the same, I used tuned to change the default size from 2Mi to 1Gi, i hit the same problem, after the reboot node is NotReady and it cannot rejoin , giving the error: Node \"test-worker-0\" is invalid: [status.capacity.hugepages-2Mi: Invalid value: resource.Quantity{i:resource.int64Amount{value:, scale:0}, d:resource.infDecAmount{Dec:(*inf.Dec)(nil)}, s:\"20Mi\", Format:\"BinarySI\"}: may not have pre-allocated hugepages for multiple page sizes, status.allocatable.hugepages-2Mi: Invalid value: resource.Quantity{i:resource.int64Amount{value:, scale:0}, My system just reports information for 1Gi, not 2Mi anymore. Removing the node and rebooting fixed it, the node rejoined again reporting the right capacity. But this is not a solution for us, changing sizes should be supported.\nHitting same K8s installed with my ubuntu 18 enabled 2Mi hugepage by default. As I change my master node's hugepage to 1Gi, this happened. I have only one master, so I have to rebuild the k8s cluster.\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nI'm a little confused. when i used tuned to change the hugepagesz from \"defaulthugepagesz=2M hugepagesz=2M hugepages=8192\" to \" defaulthugepagesz=1G hugepagesz=1G hugepages=16\", i hit the same problem, after the reboot node is NotReadybut when i used tuned to change the hugepagesz from \" defaulthugepagesz=1G hugepagesz=1G hugepages=16\" to \"default_hugepagesz=2M hugepagesz=2M hugepages=8192\", after the rebootkubelet status can still keep Ready. why?\nHi This will be fixed in , but it is still not merged. Hoping to get it into next release. In order to solve the issue (and use 1GiB huge pages) you have to delete the node in kubernetes via , and join the node again.\nThanks\nThanks\nI think this is already fixed by can you confirm?\nHi WIth this would not cause a crash, but is still required in order to fix this properly.\nI am facing the same issue. After changing the hugepage setting to . I stopped the kubelet process on the worker node and removed the node from the cluster using command and rebooted it. But after it comes back online and kubelet starts, it struggles to join the cluster . Do I have to run the join command again ?"}
{"_id":"q-en-kubernetes-0b4a54ad7e83bc37bcbe8e54551f4039ecb15e7cae112f42b6a0a0dc4debfa01","text":"/assign /cc /sig instrumentation\nGitHub didn't allow me to assign the following users: serathius. Note that only , repo collaborators and people who have commented on this issue/PR can be assigned. Additionally, issues/PRs can only have 10 assignees at the same time. For more information please see <detailsIn response to : /assign /cc /sig instrumentation Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nHey Looks like this test is passing on testgrid, https://k8s- Which would point to problem being related to specific environment. From your log looks like problem is specific to Macs. I will try to reproduce it to confirm it. My current ideas for solution are: Disabling test on Mac Instead of passing files as arguments, pass them in STDIN WDYT?\nFWIW ARG_MAX can technically vary even on linux ...? STDIN SGTM"}
{"_id":"q-en-kubernetes-f731f964b0030896cdca6269d2746eb82544a41be31752c20e1617edbddc38f1","text":"<!-- Please only use this template for submitting reports about failing tests in Kubernetes CI jobs --Which jobs are failing: ci-kubernetes-e2e-gci-gce-alpha-enabled-default Which test(s) are failing: CSI hostpath-v0 tests Since when has it been failing: 8/8 Testgrid link: https://k8s- Reason for failure: Anything else we need to know: We are removing CSI 0.3 support in 1.18, which is before selflink is going to be removed in 1.20. So we should just disable hostpath-v0 from alpha tests. /assign\n/milestone v1.16"}
{"_id":"q-en-kubernetes-0fc7990245dc881c7fd7cf440155d25f52f1cc5ba936f8a15b96a9dedcef76f2","text":"/help /kind cleanup /area e2e-testing-framework mini guide: make sure you have signed the CNCF clone kubernetes;/kubernetes search for ginkgo.Fail calls under test/e2e/framework e.g. replace them with this that was in this PR: send a PR for kubernetes/kubernetes\n/remove-kind feature /area e2e-test-framework\n/sig testing\nI will take this.\nthank you! Please feel free to ping #testing-commons in the Kubernetes slack if you have any questions :smiley: /cc\n/remove-help\n/assign"}
{"_id":"q-en-kubernetes-6f49586ccc4b3cd6ff6cd0517689dcdfb34f0d75cb920a80784ed53c3230c152","text":"<!-- Please only use this template for submitting reports about failing tests in Kubernetes CI jobs --Which jobs are failing: Which test(s) are failing: Several tests Eg. this run: Since when has it been failing: Unknown Testgrid link: Reason for failure: (happens to all the images). Pods in the e2e test suites use pull policy , , and it looks like the test during DiskPressure, resulting in kubelet removing some of the pre pulled images. Anything else we need to know: The runs after each test. This should also be the case in the test too.\n/assign /sig node\nThese tests are now constantly failing due to missing images. This makes it hard to see if some of the stuff we do now, or have done in the last few weeks, breaks some of the tests. PR should hopefully fix this, but it is not in the 1.16 milestone, but has and . Should we get this merged into master so that we can see if if fixes the problem, or should we wait until code thaw? Any thoughts? /cc /cc /cc\n/priority important-longterm\nThanks for all the work you've put into this ! I think we should add to the milestone\n+1 to - thanks for your hard work :)\nstill has a lot of failing tests. What are the next steps? /reopen\nReopened this issue. <detailsIn response to : still has a lot of failing tests. What are the next steps? Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nYes, it has over the few last weeks become quite a mess... Looks like fixed some of the issues about image eviction problems, but there are still a few left. I think I have a solution, so ill try to submit a ~PR in a few hours~ edit: will do it tomorrow. A lot of new tests started failing 8-28, and that may be related to , but I am not totally sure..\nLooks like fixed the flaking in ! The last 7 runs was executed after that was merged! ! We still have a lot of failing tests, but those are not flaky; they are failing (ref. comment above). Ill create a new issue about all the failing tests where we can track them each together. Closing this issue, as this was only filed for the flaking becuase of the evicted images. :smile: /close cc\nClosing this issue. <detailsIn response to : Looks like fixed the flaking in ! The last 7 runs was executed after that was merged! We still have a lot of failing tests, but those are not flaky; they are failing (ref. comment above). cc Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-9c8f73e236161e2d343bca6a20cc7f94cd88f2bb1e2aa8cb2cfc579ed66ad498","text":"There are multiple cert controllers that use the same base struct, so when they log, you cannot differentiate which is which.\n/sig auth"}
{"_id":"q-en-kubernetes-262c51dc7025f5bf7d7506961e607bb565387492868519142209d147b19a514d","text":"MaxInt const is exported in scheduler api although it is not needed to be exported. It is only used by . discussed that the common practice is to have API constants in /sig scheduling\ncc\nThe problem is that it is part of a v1 API :(\nOk, it turns out they are not actually considered part of the API, and so we can remove them:\nShould I open a PR to address removing those const values from ?\nYes, we can remove those in a followup PR to the one handling sized values.\nYes, this was the plan. /assign\nremember to remove as well :)"}
{"_id":"q-en-kubernetes-fa633cd4e7e8412128ec0cd6f5463a08b109011171459fc55ab90ca312dfb00d","text":"Which jobs are failing: ci-kubernetes-node-kubelet-orphans Which test(s) are failing: [sig-storage] EmptyDir volumes pod should support shared volumes between containers Since when has it been failing: Testgrid link: Reason for failure:\n/sig testing\n/assign\nThank for comment\nTests are now passing, /close\nClosing this issue. <detailsIn response to : Tests are now passing, Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-29c9eaa4e4392e47704a842958fd45dc6394841924bcd2227a3dae24a3bad60c","text":"I tried to use a local volume block device in minikube. Its trying to call losetup with the -j option and its failing as the losetup in minikube does not provide this option. This may effect other distro's as well. The end result is lock block devices don't work on the platform.\n/sig storage\nProbably caused by this code:\nDid a strace on losetup -j. looks like the same info can be had by: cat /sys/block/loop0/loop/backing_file\nJust had the same problem with Ceph cluster creation in minikube, the osd-prepare doesn't start because of this:\nFYI: has the same problem in minikube. Replacing the /sbin/losetup busybox link with a \"real\" losetup helped for now, but isn't a solution of course.\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\n/remove-lifecycle rotten\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle stale\n/remove-lifecycle stale\n/help\nThis request has been marked as needing help from a contributor. Please ensure the request meets the requirements listed . If this request no longer meets these requirements, the label can be removed by commenting with the command. <detailsIn response to : /help </details"}
{"_id":"q-en-kubernetes-a41deab71bdfef45c826d3fa1a7402d8f9d34bacbd9fb1750cb066edbe116539","text":"<!-- Please only use this template for submitting enhancement requests --What would you like to be : In block volume tests, wait for the Pod to be deleted instead of letting it be handled by the framework's namespace deletion Why is this needed: I think the regression caused by could have been caught if the tests check for Pod deletion because the Pod object shouldn't be deleted if TearDown fails.\n/assign\nThis issue is not specific to block tests. We should in ideally wait for pod deletion at the end of all types of tests to confirm cleanup logic is functioning\nI checked the tests under test/e2e/storage/testsuites/ where DeletePodWithWait is used to wait for the deletion.\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\n/remove-lifecycle rotten is this still an issue? One way to check is to make a breaking change in cleanup and see if the pull job fails.\nI went through all block volume usage in e2e tests: All of them but one wait for their pods to be deleted. I'm fixing the remaining test in"}
{"_id":"q-en-kubernetes-e7dc1eb4ab260b501e67821b2cafe42d5e17847ff38b47e56ed748a27e21cbee","text":"<!-- Please only use this template for submitting reports about failing tests in Kubernetes CI jobs --Which jobs are failing: jobs primarely, like: Which test(s) are failing: [] Container Runtime blackbox test when running a container with a new image should be able to pull image from [NodeConformance] Since when has it been failing: Since this commit merged: Testgrid link: Reason for failure: The test should always pull the image from (), it is not configurable. Anything else we need to know: /sig testing /sig windows /area conformance This issue was fixed once before:"}
{"_id":"q-en-kubernetes-0aee258daba80738d9b433ae1c1e6667f5282a4d426be899d8be9880dd9041d3","text":"<!-- Please only use this template for submitting enhancement requests --What would you like to be : Add a metric to track number of scheduler binding goroutines. This will be a generic metric with as the only label for now. metric name Description -- -- scheduler_goroutines Number of running binding goroutines Why is this needed: Helps us track how many binding are running and monitor issues such as goroutine leak. /sig scheduling\n/assign"}
{"_id":"q-en-kubernetes-38fb0bcd87a97a6879b6b690a9488c1b8612767736541d17886c023125956f80","text":"In order to initialize a client-go , you need an . and friends initialize a new : launches a goroutine: This goroutine is stoppable with . However, the interface returned above does not expose anything that calls that shutdown method, AFAICT. This means that that goroutine cannot ever be cleaned up. That's not great. It's fine if you're using in like k/k does and never stopping anything, but if you ever want to stop your controllers and start up new ones, it becomes an issue. Tested/tracked down against master (as of ). /sig apimachinery /kind bug\nThe label(s) cannot be applied. These labels are supported: <detailsIn response to : In order to initialize a client-go , you need an . and friends initialize a new : /kind bug </details\n/sig api-machinery I always forget the dash"}
{"_id":"q-en-kubernetes-4b0add5dcc29327fa9cef8cafdca1fb3b31630aa67c15b9cb2d9755a1529f259","text":"A significant number of the of are commented out, possibly was done by mistake in the past. I uncommented the tests to see if they pass, but one of them failed. What we need is to uncomment those tests, fix the one that is not passing and remove the redundant ones if there are any. /sig scheduling /help\n/assign I will be moving some of the tests to unit tests and leave a cleaner set of integration tests.\n/remove-help\nwould you like to take this? I can provide guidance.\n/help /unassign\n/assign\n/remove-help"}
{"_id":"q-en-kubernetes-4bbc21e5e6e46f1eaf0c4befa50aa3b5c4f3922fedf71188022e2ee8574d2e5c","text":"Which jobs are failing: Which test(s) are failing: Since when has it been failing: 18 Oct 13:17 PDT Testgrid link: https://k8s- Reason for failure: check\n/sig network /cc /cc\n/priority critical-urgent /milestone v1.17\n/triage unresolved Comment when the issue is assessed and confirmed. I am a bot run by vllry. ‍\nI believe this is caused by this merge cc.\nMost of this is caused by and should be fixed by . Is there any way to run these tests against that PR? /remove-triage unresolved\nThe job is still failing. Can you check?\n/reopen\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/triage unresolved\nOK, fixed some of the tests, got the rest to the point where they failed differently...\nThanks for the update. In testgrid we still see the same error. We will keep observing the next runs.\n- still failing. This time the error is We probably need to clean up sanityChecks a bit more - this seems to be a similar issue.\nyeah, I'll temporarily un- the tests in my next PR so I can make sure they're actually fixed this time\nAfter the last fix, it seemed to finally recover."}
{"_id":"q-en-kubernetes-5a1b2f312a186542159d468345a1bfab8985056d3009a60c68ed1857cb2684f2","text":"What happened: During the work for reducing invalid dependencies from the e2e core framework to subpackages of the core framework, we found a lot of particular test specific functions. That means these each function is used at a single e2e test module only. These functions are unnecessary to be shared as e2e framework function, so we've just moved these functions to the particular test module like This cleanup work will help us to maintain the e2e core framework in long term, so it is helpful to share this work in volunteers as a community. This issue is for tracking this working items. How to do that?: volunteer selects a single file which is empty (no PR, no assignee) from the following e2e core framework all functions in the file all callers of each function finding the function is called at a single e2e test package only, let's move the function to the module with a PR Implementation rules: If a function is called from 2+ test packages, don't move the function from the core package. If moving of the function requires invalid code dependencies, don't move the function. (Dependency from e2e core framework to other repository) List of files of e2e core framework: [x] ( ) [x] () [x] ( [x] () [x] [x] () [x] get-kubemark-resource- [x] ( ) [x] (Done) [x] ( ) [x] () [x] ( ) [x] ( ) [x] ( ) [x] ( ) [x] ( ) [x] ( ) [x] () [x] ( ) [x] () [x] () References: This working item was under discussion as\n/kind cleanup /remove-kind bug /area e2e-test-framework /sig testing /priority important-longterm\nThanks for taking care of this quickly :-)\nThe proposed policy to follow for what methods are to be kept within the framework: What do you think ? does that sound good?\nThanks for bring it up here. As I said on the slack, I agree with his idea.\nI looked through files , and , and found some of the functions are used in a single module. Then fixed them.\ni will pick up\nI will pick\nI will pick up\n/assign I will pick\ni will pick and\nI will pick up ,,,\nThe PR for is already created by . Please check what other people picked already before posting a PR.\nis already picked/posted by\nThat is a nice point. I wrote point as one of the above Implementation Rules, and another rule comes from which adds another invalid dependency.\nI didn't noticed that already picked by , sorry about that. After looking around, there is noting need to do with: ,,,,(not picked). I will pick the left: . Found that is also no need to move, is the same.\nThanks for your investigation, I've updated the above list based on your input :-)\nI think is does not need to be modified, Cloud you please help me confirm it?\nand\nThere is noting need to do with , and I have filed a PR for\nThanks for your help :-) I've updated the above list based on your input.\nThank you also, I approved the PR and updated the above list from your work.\nand do not need to be modified. Could you help me confirm it? THX.\nI think there is noting need to do with\nThanks for your work, I've updated the above list based on your input.\nI checked but there is no func which is called from less than 2 test packages.\nThanks for your investigation, I've updated the above list based on your input.\nNice!\nYes, thank you so much everyone!! :-)"}
{"_id":"q-en-kubernetes-845759ed726bede9274eacb03ccfa8efddf7c521a8cbdf0a6fbdb1ff712ed56b","text":"Minimal Repro: Create 2 different disks Create 2 Pods each using one of the disks Start both pods on the same node You will notice one of the disks gets ed to a path: , then both disks are ed from to a target path: which is similar but at least uses the pod . Both of these volumes are being treated as the same volume in the Kubelet even though they are definitely not the same volume. Note: it may not be \"treated\" as the same volume - it may just be that since the paths are the same the following volumes see that the path is already a mountpoint and skip the step. This is the more likely scenario based on observations. The value is the \"name\" of the inline-volume. These can obviously be the same across different pods - we should actually be placing it in a more unique directory (based off the translated maybe?). /cc\n/sig storage\nThe inline volume example I'm using (ignore the fact that is a complete misnomer): I've confirmed that changing the \"name\" makes inline volumes work well since there is no longer a collision on the paths\n/assign\nAlright, problem is sort of twofold: The referred to above for inline-volumes is set here: As we can see from the comment: it is set to the because the in-tree CSI Plugin makes a lot of naming assumptions based on the which for PVC/PVs is the PV name (unique in a cluster). Primarily, during unmount the CSI Plugin tries to from a \"well known \": which happens to be some path based on the which for the inline volume is just the ( above). So we run into the problem that if we use the directory is not sufficiently unique for different disks. And if we use something else like then the unmount pathway is unable to find the . I think using is a no-go. So we need to use something sufficiently unique (like ) for the ( above) and fix up some of the other codepaths to either use different \"well-known paths\" to write to, or to fix up path generation code so they give us the correct locations. I'm also going to check to see whether 's refactor fixes this issue (because it might since it translates things earlier)\nOK. Really good news. your refactor in its current state actually fixes this issue if we set the to something unique here: Looks like the early translation is also making the pathway from the \"correct\" dir for inline migration. So I will create a PR that fixes this issue that should merge after"}
{"_id":"q-en-kubernetes-f89bdbc46fcb71d809d84a724d89d62333e62af3f962fa3f2cfb97ed2335b036","text":"What happened: We were running a service with LoadBalancer type with AWS cloud provider and after deleting this service, the cleanup process on AWS removed a wrong ingress input rule which provoked that all traffic between nodes was unable to reach out the destination. The service we were running had a configuration similar to: After applying this manfiest on the cluster, the cloud provider created the associated load balancer with two security groups: sg- (due to extra security group annotation) sg- (a new SG created by the AWS cloud provider with the rules defined on the ) Cloud provider also created a new input rule on the nodes security group which allows all traffic from all ports from (which is the sg created for the newly created load balancer in our example). One particularity of our setup is that the security group we as was the nodes security group. The input rules of this security group after the creation of this service was similar to: Everything worked fine with this setup until we deleted this service and the cloud provider started the deletion process of the LoadBalancer on AWS. We've been debugging the deletion code and we found that the first thing it does is cleaning up the security group. To do it, it first removes the input rule from the nodes security group which allows traffic from the load balancer security group (sg-, on this example) and then, as it's not referenced, it removes the security group. The expected output of this process, would be that nodes security groups would be: But, in our case, it ended up as: which dropped all the traffic between nodes (as it didn't allow any traffic between them). The cleanup process didn't properly remove the input rule from the node security group (the one which allows all traffic from source sg-) and it also did not remove the sg- security group. In our case, as we were using the nodes security group on the extra security groups, this bug provoked a network partition between nodes, but we think that if is used, the cleanup process may be broken and have unexpected side-effects like those we observed. We've been debugging and we found that the process which removes the input rule from the nodes security groups starts here: The first thing this function does is this: : When extra security groups are used, will contain the security group created for the LoadBalancer and all the extra security groups. We also found that AWS returns this SecurityGroups list in distinct orders after changing the LoadBalancer configuration. That code is always assuming that latest security group on the list is the , which may not be true when extra security groups are used. On the cases where the load balancer security group is not returned on the latest position, the cleanup process does not remove the security group and the input rule from the nodes security group is not removed as expected. We've seen that master branch still contains this code and we think this may be hapenning on newer versions: ), so we think that newer versions may be affected by this bug (we didn't try it). In our case, we saw several times this log during the removal process: Additionally, the code tried to remove the security group (but being unable to do so), after wrongly revoking the rule we have talked about before. What you expected to happen: The cleanup process should properly detect the load balancer security group, remove the input rule from nodes security group and remove the load balancer security group, regardless of the order in which AWS returns the security group list on . How to reproduce it (as minimally and precisely as possible): We did not find any consistent process to create the scenario, as it happens randomly (we are not sure if due to AWS or any kind of race condition on AWS cloud provider). Setup: Create a Service with running controller manager with AWS cloud provider with tha annotation containing one or more security groups (as the one the section) Wait until load balancer is created Run Check if does not have the newly created security group for the balancer in the latest position. Repeat until the balancer does not have the balancer security group on the latest position Trigger the bug: Result:The log will appear multiple times on controller-manager logs Load balancer associated security group is not removed Nodes security group input rule allowing all traffic from load balancer security group is not removed If the latest security group on the list was the nodes security group, the input rule allowing all traffic on all ports from all the other nodes is removed and no traffic is allowed between nodes on the cluster. Environment: Kubernetes version (use ): version.Info{Major:\"1\", Minor:\"11\", GitVersion:\"v1.11.10\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2019-05-01T04:05:01Z\", GoVersion:\"go1.10.8\", Compiler:\"gc\", Platform:\"linux/amd64\"} Cloud provider or hardware configuration: aws OS (e.g: ):Kernel (e.g. ): Linux ip-10-2-13-67 4.9.0-7-amd64 SMP Debian 4.9.110-3+deb9u2 (2018-08-13) x8664 GNU/Linux Install tools: kops\n/sig cloud-provider\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\n/area provider/aws /lifecycle frozen\nWorking on a repro for this issue (on 1.11 first, then I'll try on master) and I've not quite observed the same behavior yet. Depending on the order of the security groups on the load balancer, I only see the ingress rule for the new LB security group to the node security group if the new security group comes last. Otherwise, if the NODE security group comes last, the ingress rule is not , and I see in the logs. Even without a full repro yet, I find it troubling that behavior is dependent on the order of security groups attached to the load balancer. One question I have related to your setup, can you describe the reason why you're adding the node security group as an ? I'd like to understand your use case a little better.\nHi Nicholas. I have a working patch for this specific issue: This bug was the subject of a talk for a local meetup I did recently. I also published a personal blog post talking about it: There you can see more details about how it can be reproduced easily. If you think the patch is valid, I'll try to add a few test cases for it and then propose the change upstream. Thanks!\nIt has been a while since we got this error and I'm not sure if we saw this error. What I experienced is that the order is changed by Kubernetes or by some call to AWS. It took me not more than 3 attempts to get this behavior. There's something I can do to help you reproduce the error? Do you need more details of our setup? This was a misconfiguration on our side. When we noticed that this setup was not correct we removed this security group from , which removed the mentioned security group rule and took down the Kubernetes cluster (no communication between nodes was allowed)."}
{"_id":"q-en-kubernetes-06a27ca04ebd97fb6e57fc139dac3b8f43140e976109866ce16fef3b42f75b6a","text":"cc who found it. the reason is the calls from the import here: i suspect the problem surfaced with this bug fix. cc more refs: is an anti-pattern in this case. ideally metrics registration should not trigger warnings for a simple version output. mitigation on the user side is to pipe stderr to a null device (as long as klog is writing to stderr) /kind bug /priority important-longterm /area kubelet /wg component-standard\n/sig node\nshould be fixed before 1.17.0, doesn't block beta.2\ncc\nCould this be caused by the following in init() of ? It was marked I wonder when deprecatedRequestLatency can be dropped.\nlooks like was deprecated in 1.14 - Is that the only one that needs to be dropped? are there more.\nDropping restclientrequestlatencyseconds should remove the warning for --version. In the 1.14 changelog, there were quite a few other metrics deprecated. I guess we don't have to drop them now.\n/assign It's an interesting issue. I will investigate it and get back to you. By the way: I tried this morning, I got 4 warnings, things getting worse as we deprecated more metrics this weekend. (the log can not point the exact name of a metric, we intended to improve it in v1.18)\nit's being invoked from here: the message is not helpful, since it doesn't tell you which metric is deprecated\nYes. It's a known issue and we intended to improve it in v1.18. Now, I suggest removing the log temporarily, Fired a PR , for reasons as follows: Too many binaries have been affected.(kubelet, kube-proxy, kube-apiserver, kube-controller-manager, kube-proxy, kube-scheduler). We should consider more about this issue. v1.17.0 is fast approaching, we should make a quick fix. All components try to register a deprecated metrics will be affected. Maybe we should record deprecated/hidden metrics by a new metric instead of the log.\nSummary the root cause here: According to metrics deprecated policy, if a metric's - = 1, it will become a hidden metric. Hidden means we will not register it to the Prometheus registry by default. If we try to register a hidden metric to the stability framework, the warning log will be printed out. Because some metrics registered by init() calls from the import, so...\nI have tried to print the metric's name at local: kubelet: kube-proxy: kube-apiserver: kube-controller-manager: kube-scheduler:\nideally the registration should happen outside of init(), which means calling it on demand. but i will defer to the owning sigs.\nIn principle, yes, I agree. Otherwise, we need to be very very careful with every log in the stability framework. I think we can track it and work out by iteration. cc for comments."}
{"_id":"q-en-kubernetes-4a920c4f5fce8f7d854e91e7314dbec0dfed4ac35f24324af41bb04310102a6c","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: When developing unit tests related to PDBs/Deletion, mutates pod resource via adding deletion timestamp, and this directly conflicts with checking resource version. What you expected to happen: No resource version mutation. How to reproduce it (as minimally and precisely as possible): I'm attempting to add preconditions to delete call in eviction API: Anything else we need to know?: Happens locally and in CI. Environment: Kubernetes version (use ): Head of master: Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others:\nThere are no sig labels on this issue. Please add a sig label by either: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. See the . The in method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\ncan you provide more details about what you're observing?\nYou can see an example here: Here's the relevant section of the tests: There really shouldn't be anything modifying the pod after creation until the delete request, thus I don't believe we should be hitting this precondition failure. Resource Version in precondition I believe to be correct, but when we contact etcd to verify, it's increased by one. It's not clear why this is happening, it appears to be a bug.\nOkay, I've sorted this out. We have a conflict with 'graceful' deletion with this approach: This updates the object to add a deletion timestamp, thus the first call to eviction will always fail if we need to set ResourceVersion."}
{"_id":"q-en-kubernetes-66dd64e4ff97f98a47f10317650c78fda69dfe8bc0533d47d3a58d91085438f7","text":"Was just perusing the code and looks like we translate regional disks into the \"zonal volume ID fmt\", and there doesn't seem to be any tests for regional disks running so we should confirm that this works and fix it if it doesn't /assign /cc\n/sig storage\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/lifecycle frozen /assign\n/assign"}
{"_id":"q-en-kubernetes-5bcbe5dcb8b5335d46bb282f29b295c3c09cd01dbc8f992a23341393734a7d37","text":"<!-- Please only use this template for submitting enhancement requests --What would you like to be : We should allow collecting both default scheduler metrics (Prometheus) and custom metrics easily. At a high level, these collectors should produce a json formatted result which will then be consumed by perf-dash to present in the dashboard. Prometheus metrics We should provide a metrics collector that collects and summarizes default scheduler metrics defined in . To begin with, we can manually pick a list of interesting metrics. Longer term, we could allow configuring the metric list in test definition. already had a PoC . As a reference, is how clusterloader collects scheduler latency metrics. Custom metrics The most important metric is scheduling throughput. is how clusterloader calculates the throughput. We can start with just calculating the average throughput without the percentiles. We may want other custom metrics that are either common to all tests or specific to certain tests. We should allow these custom collectors to be easily and ideally configurable. Why is this needed: Give more insights into scheduler performance. Part of /sig scheduling /assign\n/reopen\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-541b80bf02b46bee322273efed3dcd702ed6a05dc4bb5f0e424a0423ce019e1d","text":"<!-- Please only use this template for submitting reports about failing tests in Kubernetes CI jobs --Which jobs are failing: Which test(s) are failing: [] [sig-node] Pods Extended [] Delete Grace Period should be submitted and removed [Conformance] Since when has it been failing: The ipv6 jobs have flakiness, but since I had 3 flakies in a row with this test Testgrid link: (IPv6),%20master%20(dev)%20%5Bnon-serial%5D&width=5&sort-by-flakiness= Reason for failure: It fails because is not able to delete de pod in 30 seconds, seems that this is the cause of that Nov 30 07:41:07 kind-worker kubelet[677]: W1130 07:41:07. 677 ] Failed to process watch event {EventType:0 Name:/docker/f35390ea97c6bae92f61d7d297fe7eacb4cfb3f2effafb23f3aca00e1a9d2341/kubepods/besteffort/podd8d966f8-2a03-4529-9d70-/243828b3fed0edf1137a9fa5ade6ffbbe468b79bbf212206480bbd37ab20fc7d WatchSource:0}: task 243828b3fed0edf1137a9fa5ade6ffbbe468b79bbf212206480bbd37ab20fc7d not found: not found Anything else we need to know: The job view seems to report the same test failures twice but the log output shows different timestamps\n/sig node\n/cc\ninterestingly, the serial ipv6 job doesn't look flaky: (IPv6),%20master%20(dev)&width=5&include-filter-by-regex=Delete%20Grace%20Period\nthis one is failing quite a lot in the 1.17 windows job - could someone confirm if it's a bug or a test issue? Thanks\nor is it actually a completely different failure on windows? Since it's being tracked here as well\nwill close this issue because the patch addressing the issue was merged and to avoid having duplicates We can reopen if necessary /close\nClosing this issue. <detailsIn response to : will close this issue because the patch addressing the issue was merged and to avoid having duplicates We can reopen if necessary /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nby the way, this one is unrelated to the windows one :-)\nACK, thank you :)\nTwo failures seen on after the related PR merged, that was only a cleanup, not a root cause. /reopen\nReopened this issue. <detailsIn response to : two failures seen on after the related PR merged Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/remove-kind failing-testing /kind flake\nThose labels are not set on the issue: <detailsIn response to : /remove-kind failing-testing /kind flake Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/remove-kind failing-test\nI think that I'm making progress trying to deflake the ipv6 tests I have submitted PR to fix the portmap plugin issue that should be able to improve the situation\nRetitling because I don't think this specific test failure is IPv6 specific anymore\ntriage: flaking in pull-kubernetes-e2e-gce: /priority important-soon This is the fourth flakiest test for pull-kubernetes-e2e-gce\ndebugging in\nbumped priority, a conformance test must not be flaky. this is now the top flake in the e2e suite.\ncc\ndebugging log messages merged 2019-12-19T08:38:10Z, will check for output in next failure\ndebug output from failure:\ndebug output from success:\nfailure in\nbased on my reading of this, the timeline was: 00:02:23. - we got a pod running 00:02:23. - deleted it with a 30 second grace period 00:02:54. - the kubelet finished handling the delete request (30.18 seconds) 00:02:54. - test failed with a timeout (30.25 seconds) That seems like extremely tight tolerances, especially since the grace period refers to termination of the pod process, not removal from the API. I would be inclined to allow up to 2x the grace period to observe complete removal from the API.\nopened to allow time for the kubelet to remove the terminated pod from the API\ntop flake according to\nOpened to tag as [Flaky] and kick out of [Conformance]\nTest is no longer run in blocking presubmit. - who owns resolving this and reintroducing the test before 1.18 is released? Investigation with combined timeline at shows that the kubelet stopped the container and destroyed it within ~15 seconds (observed deletion at 06:20:11., destroyed container at 06:20:46.), but then took another 30 before actually deleting the pod from the API (06:21:12.).\nTo evaluate the effectiveness of a potential fix, I would recommend opening a PR that removes the tag along with the proposed fix, and repeatedly on that PR to accumulate 10-20 test runs.\n/assign assigning sig-node leads for visibility/routing\n/milestone v1.19\n/milestone v1.18 we need an ack from sig-leads whether this is release blocking or not\nThis issue is opened for 2 releases in a row. I guess the flake is not a release-blocking. In addition, there are many PRs linked to this issue, and many attempts happened to make the situation better. However, we should keep this issue open and make sure that the flake will disappear. WDYT? cc\nThe test has been flaky for multiple releases, but it was marked flaky and completely removed mid-1.18, which is new. I'm wanting to make sure that absence of any test signal around this is known/acceptable to sig-node.\nLinks to the investigations I was able to do: combined e2e/api/kubelet log from failure with extended timeout:\nAdded to agenda for 3/24 sig-node call"}
{"_id":"q-en-kubernetes-1e3d5f814d3d1cd0ae92f2f9adc262569d0e3418b95dbcb250b25bdb3821fc75","text":"This issue proposes the writing of a new test for the API endpoint with the intention of promoting it to Conformance. This issue contains a very masic mock test in order to start the conversation about what a good e2e test would look like for this endpoint. The code below is verified to be hitting the intended endpoint, show as per the queries in APIsnoop's live view of the cluster's audit logs. According to this query, there is only one API endpoint which is for Namespace that isn't being covered. We selected patchCoreV1Namespace to write a test for. Dependencies: The mock test: If a test with these calls is build and merged, conformance will go up by 1 endpoint.\n/sig testing /sig architecture /area conformance"}
{"_id":"q-en-kubernetes-0c58a4fb4c5ae349fea47aac4d01feb301a025031a43485c622d67f8c7ac71a3","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: With client-go v0.17.0, I observed many \"Throttling request\" when DiscoveryClient is used to get API group resources from API server. This didn't happen with client-go kubernetes-1.16.0. It looks like originally this kind of discovery requests are not throttled because no rate limiter provided and QPS == 0. But with rate limiter is always set if nil rate limiter is provided. is particularly suspicious because it is always true because DefaultQPS is nonzero. What you expected to happen: Discovery requests should not be throttled as before. How to reproduce it (as minimally and precisely as possible): Create DiscoveryClient and then call with it. Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others:\n/sig api-machinery\nThe NewTokenBucketRateLimiter was via Always negotiate a decoder using ClientNegotiator /cc\nTo be clear, that was fixing a bug where the documented defaults for rate limiting were not taking effect. I don't think we should completely disable throttling until more of the server-side priority work is completed, but discovery is expected to be bursty, so I could see defaulting to a different amount in the discovery client constructor, similar to\nYeah, everyone must be rate limited. I think we could give discovery a big burst value.\nopened"}
{"_id":"q-en-kubernetes-f02e592daf9802b1402c51394d73fa38959304b6bce5764c2379570be395ac35","text":"What happened: make quick-release-images fails on v1.17.0, works fine on v1.16.4 What you expected to happen: Should work on v1.17.0 How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: The immediate cause of the problem seems to be that the hyperkube image creation fails because it doesn't find the kubectl and kubelet binaries in , apparently because they weren't built.\nThere are no sig labels on this issue. Please add a sig label by either: a sig: e.g., to notify the contributor experience sig, OR the label manually: e.g., to apply the label Note: Method 1 will trigger an email to the group. See the . The in method 1 has to be replaced with one of these: bugs, feature-requests, pr-reviews, test-failures, proposals. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign"}
{"_id":"q-en-kubernetes-e830e3d0767adb87aa2cd365dc2a5abe0e593a8c5b26b9b5bc7af11b88326c30","text":"Which jobs are flaking: pull-kubernetes-e2e-kind Which test(s) are flaking: Testgrid link: Reason for failure: Anything else we need to know: /sig apps\nFailure at\nHi I looked into the failures and think it was because the test required the job succeeds 3 times with pseudorandom 50% failure chance within 15 minutes, while there is an exponential back-off delay (10s, 20s, 40s …) capped at 6 minutes before recreating failed pods. As 7 consecutive failures (1/128 chance) could exceed the timeout, the test failed intermittently because of \"timed out waiting for the condition\". For example, in the failure of , the job pods succeeded and failed in the following orders: After the last failure at , the delay has increased to 640s so no pod was created any more before the test failed at . I think it can be fixed by using a deterministic approach to make the job pods succeed and fail alternately, and its executing time can be reduced to tens of seconds stably. Assigning myself. /assign"}
{"_id":"q-en-kubernetes-2cbe7124f994368121a69d8f0ca559739dc577f71839d3f6da7e2a0a71215679","text":"This issue contains a very basic mock test in order to start the conversation about what a good e2e test would look like for this endpoint. The code below is verified to be hitting the intended endpoint, show as per the queries in APIsnoop's live view of the cluster's audit logs. According to these query results from APIsnoop, we are lacking in tests for the follow endpoints related to Secrets. We chose the follow endpoints for this mock test. Test outline: a Secret in the default namespace with custom values Secrets in all namespaces search for Secret created in  the data in Secret  Secret  Secrets in all namespaces ensure that Secret  doesn't exist Discover useragents Display endpoints hit by the test Review the change in coverage If a test with these calls gets merged in, conformance will go up by 3 endpoints. /sig architecture /area conformance\n/reopen\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nYeah! We got a test written and a promotion (and we linked them to the mock-ticket that started it all). /close\nClosing this issue. <detailsIn response to : Yeah! We got a test written and a promotion (and we linked them to the mock-ticket that started it all). /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-52c9bd0201dca9dc3f0a0243c887b805ff017af46b559b3120ff0dde08ba368f","text":"/sig scheduling /priority important-soon See as an example. Part of /assign\nGitHub didn't allow me to assign the following users: angao. Note that only , repo collaborators and people who have commented on this issue/PR can be assigned. Additionally, issues/PRs can only have 10 assignees at the same time. For more information please see <detailsIn response to : /sig scheduling /priority important-soon Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nthx, I will do it soon."}
{"_id":"q-en-kubernetes-e646bae560ecff79e11d23f16c7cab8c6cae84aaefd504987911a9d2b84ef1b0","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: The apiserver, controller-manager and scheduler are configured to use IPv4 addresses however, are serving on all addresses IPv4 and IPv6 What you expected to happen: To listen only on the IP family specified How to reproduce it (as minimally and precisely as possible): or checking in one of the gce presubmits jobs Anything else we need to know?: it seems that the apiserver, controller-manager and kube-scheduler listen on all IP families by default despite the IP family address configured Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others:\n/assign /cc\nbinding 0.0.0.0 to all interfaces is the documented behavior of net.Listen:\nThe unspecified ip (0.0.0.0) is a special case that means \"bind to all interfaces\". This is currently working as intended. Is there a specific problem this causes?\n/triage unresolved Comment when the issue is assessed and confirmed. I am a bot run by vllry. ‍\nIf (0.0.0.0) means \"bind to all interfaces on all IP families\", then is a misconfiguration problem and not a bug . I was worried because this causes that some components use IPv6 for their communications on dual-stack hosts despite the cluster is IPv4 single family (and viceversa), as we can see in this snippet from one of the gce presubmits job and I don't know if this can impact on some services or from a security point of view\nFrom the listen godoc: That means binding to or or will bind to all interfaces\nI would expect anyone using network interfaces as security boundaries to be binding to a specific interface, not binding to all and hoping \"all\" is family-partitioned\nClarifying the handling of unspecified addresses like that in the flag/config help seems reasonable, but as far as I can tell, this is working as intended.\nyeah, thanks for the clarification, I totally misinterpreted the net.Listen function\n/triage unresolved Comment when the issue is assessed and confirmed. I am a bot run by vllry. ‍\n/remove-triage unresolved\nSo, is there any way to listen on IPv4 addresses only (all of them)?\nhmm, that will requires to expose this setting just curious, what is the use case for having IPv6 on the host and not in the apiserver, is not possible to disable ipv6 in the host?\nI'm considering options to have a load balancer (HAProxy static pod in hostNetwork) for api-server on control plane nodes listening on the same port as api-server. Then I would point api-server VIP to via hosts file on the control plane nodes. This is a workaround for the absence of hairpin in AWS NLB (with instances by id) or Azure LB. I'm looking for a generic solution that will work for AWS/Azure/on-prem with nodes in the same/different subnets. If you have other suggestions, feel free to share, I would appreciate that. For now, my solution is to call , then patch the generated static manifest with set to the node's default ip, and then bootstrap with . Binding the load balancer to IPv6 and and api-server to IPv4, then using in kubeadm's config will be a dirty, but cleaner from the code perspective, approach. Btw, is there any reason behind having set to by default, but not the node's default ip, just like ?\nI can't understand how do you solve the hairpin problem that way, is not the traffic been directed to the LB vip too? or do you mean that you don't use the AWS NLB and use the haproxy instead? what VIP do you use then? I don't remember the details, but the apiserver should be available to the pods too, that don't reach it through the node, they use the the kubernetes.default Service\nExternal clients use AWS load balancer, while requests from control plane nodes go to haproxy on localhost thanks to trick, thus do not reach AWS LB."}
{"_id":"q-en-kubernetes-e65d433be70b14f59611269e2f1c0fa746f312a0f5c96a0c29a874f17f7c06ea","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: If you create a service that reference a Basic IP in a standard load balancer cluster, the service will stuck at \"Pending\" state for days without detailed error from Azure response. When you describe the service, it only has following: What you expected to happen: In this case Azure responded 400 with detailed error message that the referenced IP is a basic IP and cannot be used on standard LB. I would expect the detailed error message show up as events for easier diagnostic. How to reproduce it (as minimally and precisely as possible): Create a service that reference a Basic IP in a standard load balancer cluster Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others:\n/assign /area provider/azure /sig cloud-provider\n/area cloudprovider"}
{"_id":"q-en-kubernetes-2f877173a2a9c18bda24d29b16bbe49e2d322978f8470207d055a2e70ed9f44c","text":"Which jobs are failing: Which test(s) are failing: Testgrid link: Reason for failure: check Anything else we need to know: /cc /milestone v1.18 /sig cloud-provider-gcp\nThe label(s) cannot be applied, because the repository doesn't have them <detailsIn response to : Which jobs are failing: Which test(s) are failing: Testgrid link: check /cc /milestone v1.18 /sig cloud-provider-gcp </details\n/sig cloud-provider\n/area provider/gcp\nThe label(s) cannot be applied, because the repository doesn't have them <detailsIn response to : /area provider/gke </details\nProbably related to\nshows the merge shows the errant merge that inadvertantly disabled the test This test was probably failing since forever.\nAny news?\nthis one is on cloud provider GCP.\nCan you please take a look?\nFollow up from today's release meeting: If the test has been broken for as long as we know it, theres value in just removing it. If the test owners don't have capacity to look into it, CI Signal can investigate.\nAny advice on which team is responsible? It seems there isn't any maintainer for this job.\ntry talking to please\n/assign\nTurning to the question of whether these tests are actually useful... One of the tests that has been continuously failing is this one Which essentially does because From the logs (example ) it waits for 2 minutes (120 seconds) for each node to not be ready and then for the next couple minutes it just checks this condition until it fails So it looks like the e2e tests are not failing because of inherent infra issues (i.e., kube-up decided to do some weird thing) but the issue seems to be with the time to reach a condition. Again if we look at the logs we can see that switching the network interface does have an effect and leads to the nodes reaching a non-ready status but ultimately fail to come back up within the desired 5m time mark - this bit of the tests is where it all goes bad . The question then turns to turning off the network interface for a node in such a way () and then expecting the node to come back to live within 5 minutes of turning the interfac back on something reasonable? If it is, then either we have to look at why it takes longer than 5 minutes or we can increase the timeouts If it isn't, then how can we test the ability of nodes to be rebooted?\n/assign\nWanted to triple check on the test looking for nodes to be back up after being rebooted: I dont think 5 minutes is a hard deadline - only SLOs I could find were for scalability"}
{"_id":"q-en-kubernetes-42ae4c7b8e8f0f236921b5c8c12ce1068419b9c2c8b8b3753423464720d25107","text":"What happened: We enabled the feature gate on to start switching from non-expiring to expiring ServiceAccount tokens as . It works well. We then wanted to completely get rid of the \"legacy\" non-expiring ServiceAccount token Secrets. To do that we configured with (so it wouldn't auto-create the Secrets), removed the existing Secrets and removed the reference from the respective ServiceAccount to the Secret. This works as expected: No non-expiring token Secrets are generated. However, when creating a Pod we now see: This happens because the rejects the Pod because it cannot find a ServiceAccount token Secret (because it doesn't exist anymore as intended). What you expected to happen: When using the feature the legacy ServiceAccount token Secrets aren't used anymore, so a Pod should not be rejected if the Secret doesn't exist. The Pod doesn't even use the Secret at all. How to reproduce it (as minimally and precisely as possible): Enable feature gate Start with Create a new ServiceAccount Deploy a dummy app using that ServiceAccount Observe no Pod being created with Describe the corresponding ReplicaSet to see the above Error event. Anything else we need to know?: Looking at the it seems that at least one legacy Secret must exist in order for the Pod to get admitted. This is enforced even when and we're relying on projected token volumes. Following the code path when is enabled reveals that the token is never going to be used later on as there's no SecretVolume to be mounted in . In other words:  is just fine when . During the transition period it's fine to have the legacy Secrets lying around but ultimately it's desirable to get rid of them because as far as I can tell they are still happily accepted by the API server. Environment: Kubernetes version: tested on but code seems the same in and Cloud provider or hardware configuration: OS (e.g: ): Ubuntu 18.04.3 LTS Kernel (e.g. ): 4.15.0-1054-aws Install tools: /kind bug /sig auth\nFixed in\nThank you!"}
{"_id":"q-en-kubernetes-ccc79f38703779eeea43de8b8711b5df6dfd509e69db05b2dd555127ac20293e","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: After issue is fixed. Cloud provider will automatically reconcile the Azure Public Ip DNS record same as the service annotation. As a result, user finds their DNS record will be deleted if the annotation is not set. Such a change is a block to users who have set the DNS record on their public IP without using this annotation when they upgrade the cluster. Besides, if the public ip is a user's own public ip, it is also not reasonable to delete the DNS record. What you expected to happen: We may need to change the reconcile logic. Maybe only delete the DNS record, when the annotation exists but empty value How to reproduce it (as minimally and precisely as possible): Manually add the DNS label to Azure Public IP resource via Portal Environment: Kubernetes version (use ): 1.15.7 Cloud provider or hardware configuration: Azure\nThe label(s) cannot be applied, because the repository doesn't have them <detailsIn response to : /sig Azure Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig cloud-provider\n/area provider/azure\nThe label(s) cannot be applied, because the repository doesn't have them <detailsIn response to : /sig Azure Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign\nThe fix helm install helm-nginx-ingress stable/nginx-ingress  --namespace ingress-nginx-ns  --set controller.replicaCount=2  --set controller.nodeSelector.\"beta.kubernetes.io/os\"=linux  --set controller.service.annotations.\"service.beta.kubernetes.io/azure-dns-label-name\"=MY-AZURE-DNS  --set defaultBackend.nodeSelector.\"beta.kubernetes.io/os\"=linux"}
{"_id":"q-en-kubernetes-c0e9773b200246da20ddbdade8279e9d92842718ee1cbddaf201518182c2575f","text":"Which jobs are flaking: pull-kubernetes-e2e-gce pull-kubernetes-e2e-gce-rbe Which test(s) are flaking: Stage Testgrid link: Reason for failure: This isn't technically a test that's failing, it's part of the build-stage-up-test cycle used by our e2e jobs. There are two main clusters of failures that I can tell. One is timeouts. The other looks like: Example -gce jobs: () () () () Example -gce-rbe jobs: (source gist ) Anything else we need to know: Triage link for the job that actually reports to PRs: ends up running Some googling for \"bazel error code 14\" leads me to believe that bazel (or its jvm) is getting killed due to OOM. which is a value that's been cargo culted around as far as I can tell. I was unable to find the decision that led to this value. I'm wondering if a bump to 8Gi of memory would help reduce occurrences of this flake. I'm referencing the -rbe jobs to point out that this happens even if we use RBE, but they're not blocking\n/sig testing Top flake according to\nWith merged we'll wait and see if the Stage flakes on pull-kubernetes-e2e-gce go down. If they do, there are other presubmit jobs that could benefit from that. If they don't, we should revert.\n/priority important-soon\n/cc\nStill seeing flakes, so I don't think just bumping memory request helped. ref: my theory is that making this non-privileged broke our ability to free memory This happened via around 2019-04-11 - flakes were still occurring prior to this merge - but not prior to 2019-04-01 - and very infrequently prior 2019-03-15 It could be that bazel's memory usage grew when we bumped it to 0.23.2 on 2019-03-29 (or maybe on 2019-04-02 is what actually deployed the change) Either way, I still think it looks like bazel is eating more memory than the container is allowed, and is getting oomkilled. Two ideas: Instead of allowing the bazel server to exist as a long lived process, have kubetest call in between the Build and Stage steps, in addition to trying to flush memory Constrain bazel's host JVM to a hardcoded memory limit via in bazelrc\nIdea 3: switch build modes and see how expensive / flaky that is. Including: use RBE use the dockerized build instead\nI opened to try and constrain the JVM's memory. Jobs that use RBE behave the same way, so try doing the same to them for comparison's sake.\nthis is still being seen extremely frequently (seen multiple times in a row on the same PR at )\nI opened for the last idea I have for working around bazel's behavior, beyond this I am out of my depth on how to replace bazel for the release/push- script\n... but I'm willing to try with a canary job and see what happens\nbumping priority, this looks like 99/122 flakes according to /priority critical-urgent\nAnother occurrence:\nI'd like to see if we can root cause what changed around 1/27-1/29 that greatly increased this\nkubernetes/kubernetes changes: 1/26, ~6am UTC: 1/28: ~6pm UTC: test-infra changes: 1/25, ~3pm UTC: 1/28: ~8pm UTC: in that delta kubekins changed from v20200123- to v20200127- nothing jumps out at me in those diffs, but wanted to link to them in case it helped anyone else see something\nIt doesn't have to be one, the CI is run overcommitted with no limits, something else could be upping memory pressure..? On Mon, Feb 10, 2020, 08:38 Jordan Liggitt <:\nok... after sweeping k/k and test-infra for that time period, nothing clear emerged, and the bazel bandaids seem to be holding for now. unsatisfying :-/ /close\nClosing this issue. <detailsIn response to : ok... after sweeping k/k and test-infra for that time period, nothing clear emerged, and the bazel bandaids seem to be holding for now. unsatisfying :-/ Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-07859fc11b7535f0f0bd2cd972e6b55c35b142283d2f3b03d681746acc3dc75b","text":"What happened: Currently, the mutating admission happens before updating managedFields in create requests, but in the update and patch handlers, mutating admission happens after updating managedFields. This causes the ownership of fields modified by mutating webhooks to be inconsistent based on the request verb. /sig api-machinery /cc /priority backlog\n/wg apply\n/assign"}
{"_id":"q-en-kubernetes-e1250f4d2f92a22480c3633f206afa6042690071587f7cb7acc16548eaac86ff","text":"<!-- Please only use this template for submitting reports about continuously failing tests or jobs in Kubernetes CI --Which jobs are failing: gce-master-scale-correctness Which test(s) are failing: [sig-cli] Kubectl client Simple pod should return command exit codes Since when has it been failing: 02/05 Testgrid link: https://k8s- Reason for failure: Anything else we need to know: /sig cli\n/priority critical-urgent /assign\nThe test seems to be flaking quite a lot recently ! Friendly ping, sig-cli, could you take a look?\nFTR, these are really blocking tests and it's getting more urgent given the forthcoming 1.18 release\nFull log of the failure: It looks like the finishes with error. Strange, I thought is supposed to create a pod. Would really appreciate if someone from sig-cli could take a look\n4 out of last 8 runs failed: https://k8s- /sig scalability\nAnother flaky, related test:\nSIG CLI Leads -- any update on this? This is affecting scalability signal. /assign\n-- Please keep tabs on this one.\nyep, monitoring\nIt looks like there's a race between when the pod gets created and when we try to connect to it, thus we're seeing but further down in the logs the pod is , I'll be looking more into ti.\nThanks, Any update on this?\n/cc\nSorry I got sidetracked yesterday and I'm back investigating this just now...\nProbable fix in\nWill continue to monitor here and reopen if needed :+1:\nUnfortunately - it seems that the problem still exists - the test failed today (an hour ago): :( ^^\nI'll check it, thx Wojtek for pointing that out.\n- I'm not familiar with kubectl code, but I took a quick look and what I found is that this waitForPod() thing is happening only for pods with RestartPolicyNever: I must be missing something here, but for the case that failed it's not really the case, which means we don't wait for it?\nThis particular one is using , see the actual invocation that failed: That's why I started with this approach.\nSorry - the one for which the bug was originally opened. But the today`s failure is actually slightly different (the same test, different case - sorry): The restart policy for this one is\nGotcha, so it looks like we progressed, I'll check the possible paths for this one, something tells me we're missing another timeout somewhere :wink:\nOK - I think I know what is going on here. Here is what is happening (for the test case from two comments above): the test is setting \"-i\", so interactive mode is switched on that option triggers that Attach option defaults to true: over the Run() command, we get to the handleAttachPod: handleAttachPod is calling waitForPod And waitForPod is broken. The problem is the preconditionFunc: which is below passed to the UntilWithSync function UntilWithSync underneath is creating an informer and calling that function against its state: But the informer is initialized from apiserver watchcache, so it may be lagging. In large clusters we are observing lags in the order of small hundreds of milliseconds and that seem to be enough. What is happening is that: informer is initialized (but the pod creation is not yet reflected in it) the preconditionFunc is called the state doesn't contain that pod (yet) the whole thing fails with that error I actually believe, that just removing the preconditionFunc should actually solve the problem. I opened for that.\n^^ - please take a look at the PR that I assigned to you (it's literally one-liner as of now - if you want to add tests or anything like that, feel free to pick it up).\nActually I went ahead, and fixed couple other commands too - so it's no longer one-liner. But still short one :)"}
{"_id":"q-en-kubernetes-447a18fff2918a999c9e2fd1b3e323060cc20854e937e0b5de508546374b3af6","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: , from whatever combination of inputs we stuck into it, segfaulted. The segfault, which originates from a slightly modified fork of v1.17, looks as such: Translated to master, where the problem appears to be present, the segfault originates from this call: And that is called from here: This seems odd, because The only pointer involved in the segfaulting call is itself The caller nil-checks before calling a method on . I did some digging and found , which is a fix for this problem. However, this fix appears to be undone in . There doesn't seem to be a reason for this to have been changed in the latter PR. What you expected to happen: should not segfault from any bizarre combination of user input. If the situation is invalid, it should fail gracefully. How to reproduce it (as minimally and precisely as possible): Honestly, I'm not 100% sure what combination of inputs we crammed into , but I can locate them and post them if necessary. Anything else we need to know?: I don't feel confident opening a PR to fix this myself, but in addition to recreating the fix from , a comment should be explaining why there is this weird, seemingly pointless layer of indirection when assigning to the interface, so that this doesn't get un-fixed again.\nMaybe decompose the above statement into multiple lines which should give us more idea which field was nil.\n/sig network These SIGs are my best guesses for this issue. Please comment if I am incorrect about one. I am a bot run by vllry. ‍\ncan't be nil, as it is of type , which is not an interface or pointer is an interface. However, when built from the exported constructor, , it is set to a literal , which is likewise not a pointer Therefore, the only remaining field is itself. Plus, I don't quite recall my go segfault stack trace syntax, but I'm relatively sure that indicates that the first argument to the method is , nil, and that the first argument to methods is the object being called on.\n/triage unresolved Comment when the issue is assessed and confirmed. I am a bot run by vllry. ‍\nWhen config.HealthzBindAddress is empty, we would encounter the scenario depicted in the stack trace. I have proposed a fix.\n/assign\nThat commit, , is part of the PR that I believe introduced the problem. I strongly believe the problem to be one of a non-nil interface with a nil receiver. This is the only way that could be nil in the snippet of code from . Before PR (the changes from which are present in the version I ran), there was this odd looking set of code: The variable healthzServer, which is a pointer of type , was assigned to the variable , which is an interface of type . Then, later on, the variable was passed to , as an argument of type . This looks unexpected, and it appears that the variable is extraneous. After the changes in PR , the variable was removed, and was passed directly to as the argument for interface type . The former set of changes, with the intermediate variable, are meant to avoid the situation of a non-nil interface with a nil value. If the variable is never set in the conditional statement, then it is a nil interface. If is passed directly, then the result is the argument to is an an interface who's value is and who's concrete type is . This means that is called with as the receiver. I strongly believe this to be the root cause of the issue, but I do not know enough about kubernetes to open a PR with the correct fix, which would be to reinstate the variable (after the renaming in , this would be a variable of type ) to use as a go-between, to once again avoid this situation of a non-nil interface with a nil receiver. I believe that the removal of the variable was out of a misunderstanding of its purpose. If my assessment of the problem is incorrect, please let me know so I can understand where my logic is wrong. It is very possible that I am misunderstanding the phenomenon of a non-nil interface with a nil receiver. I would open a PR myself, but I do not know enough about Kubernetes to be confident in my ability to correctly fix and test this problem.\n/remove-triage unresolved"}
{"_id":"q-en-kubernetes-c0df0534fd17c133fb6f328c49b78b58cbd5eecd8d01bc7e2006ccc94f7a590c","text":"Seen in the KIND e2e test: a) maybe we shouldn't log this more than 1x per second or something? b) maybe the error could make it clearer that the problem is the user's data? c) maybe we could fix the test triggering this? /sig api-machinery /wg apply Logs that I saw this in:\nThe 'duplicate entries for key' error came from validatingObjectWalker#visitListItems.\n/assign\nYeah, they have multiple environment variables with the same name. I don't think that should be an error though ...\n/cc\nAt the same time, it's not really useful to have multiple times the same variable and is probably a recipe for trouble.\nIt's definitely user error. Spamming apiserver logs about it doesn't help things, though.\nI don't want to hide it, it's really not supposed to happen :-) it's not really user-data though. The fact that the error triggers here rather than in the actual validation means that there is a discrepancy between our validation and the regular validation. That's probably a problem that we have to fix. In that specific example, we can see that Kubernetes allows multiple environment variable with the same name, while we have assumed that it's not OK. It's a very legitimate error and we have to think about it to fix it. I think the test is relevant, and people out there are doing that. We could consider doing ratcheting if we want to stop allowing multiple variables, but I don't think that's how we want to fix that problem.\nFor history: should fix that.\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nWe've definitely improved that, maybe let's close it for now.\n/reopen We still see this happening. Kube users get nervous, I would too. Please fix the root cause.\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-9b4e57f70d867b81062475b4c5e9d584cb2ea943db8f6a1ec73aa8cb21e83139","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: when i installed k8s using kubeadm and my custom config like then output is It is a bug that the default value(Node,RBAC) is entered but a warning message is displayed. A function must be to compare two string slices. What you expected to happen: If the default value and custom value are the same, no warning message should be displayed. How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): 17.3 Install tools: kubeadm /area kubeadm /feature bug /sig cluster-lifecycle\n/lifecycle active"}
{"_id":"q-en-kubernetes-aec36b7f770d9404a1db4677e90d433651191647f402976249b7f4561af1fad0","text":"Which jobs are flaking: pull-kubernetes-integration Which test(s) are flaking: TestStressingCascadingDeletion Testgrid link: Reason for failure: Maybe related: Anything else we need to know: links to appreciated links to specific failures in spyglass appreciated\n/sig api-machinery\nI'm seeing lots of these.\npost-submit failures (all of these are flakes):\nfirst failure seen in postsubmit is , part of ... preceeding many runs looked clean\nI verified the delete options as seen by the server are identical before and after , so I don't think it is a functional change related to that PR\nnotably, many other tests in this suite started flaking on 3/7 and after ()\nfigured it out, dropped the detection of empty deletion options which would trigger the test to wait for all pods to be observed by the GC controller"}
{"_id":"q-en-kubernetes-9f9a5b66cafbfab35661f2396d44bffdf5a8d33c0b45a5ff7ba71eeb1dbdc8cd","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: If gid is configured in Pod's securityContext.fsGroup, csi inline will panic. log: E0320 16:50:59. ] Operation for \"\"\" (\"-ae1b-4b5c-be91-\")\" failed. No retries permitted until 2020-03-20 16:51:31. +0800 CST m=+. (durationBeforeRetry 32s). Error: \"recovered from panic \"runtime error: invalid memory address or nil pointer dereference\". (err=<nil) Call stack:ngoroutine [running]:(0xc001497e70)n +0xb5npanic(0x3b713e0, 0x7bd2050)n +0x1b5(csiMountMgr).applyFSGroup(0xc0005372c0, 0xc001a2f100, 0x4, 0xc00205ed70, 0x44, 0x0)n +0x64(csiMountMgr).SetUpAt(0xc0005372c0, 0xc000b6cb80, 0x78, 0xc00205ed70, 0x0, 0x0, 0x0)n +0x1180(csiMountMgr).SetUp(0xc0005372c0, 0xc00205ed70, 0x0, 0x0, 0x0)n +0x4d(operationGenerator).GenerateMountVolumeFunc.func1(0x8, 0x453f9f8, 0xc001497e70, 0xc0025ba750)n +0x4a4(GeneratedOperations).Run(0xc0009742a0, 0x452f628, 0xc000d56480, 0xc00135a050, 0x45)n +0xaf(nestedPendingOperations).Run.func1(0xc000d56480, 0xc00135a050, 0x45, 0xc001106d50, 0x24, 0xc0009742a0, 0x0, 0x0, 0x0, 0x0)n +0xcbncreated by (*nestedPendingOperations).Runn +0x378n\" code: If using inline volume, c.spec.PersistentVolume will be nil What you expected to happen: fsGroup is successfully applied Environment: Kubernetes version (use ): 1.16.3 Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others:\n/sig sig/storage\nThe label(s) cannot be applied, because the repository doesn't have them <detailsIn response to : /sig sig/storage Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig storage\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-testing, kubernetes/test-infra and/or . /close\nClosing this issue. <detailsIn response to : Rotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nran into this bug recently on v1.18, and I don't see anything that would fix this issue on later releases either. /reopen /lifecycle freeze /assign\nReopened this issue. <detailsIn response to : ran into this bug recently on v1.18, and I don't see anything that would fix this issue on later releases either. /reopen /lifecycle freeze /assign Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/triage accepted\n/lifecycle frozen\nI tested this scenario with the hostpath driver on k8s 1.24 (built from master) on a local VM using a slightly modified version of the examples/csi-app- spec ( fsGroup and fsType): And this does not panic kubelet It does not panic because of this check: Which was in However, I do believe we can and should support fsgroup for CSI inline volumes. Instead of checking the PV spec (which is not set for inline volumes) we could pass the directly into , which will be RWO for inline volumes.\nSorry, it was not obvious that the panic was already fixed when I first looked at this. Re-purposing this bug to address the issue that fsgroup still does not get applied for inline volumes. /retitle CSI inline volumes do not apply fsGroup\nTHX, also link to this\n/retitle CSI inline volumes should support fsGroup"}
{"_id":"q-en-kubernetes-e5193060d64822a343afb837f2656faef432e7db14c37f1e690e2ab28b3f00fb","text":"After reading up on some corner cases with etcd client behavior (connections, streams, etc.) I came across the following option that has been introduced in etcd a while ago: From the official etcd docs on client behavior: Source: I do not see this option being used in the API server when is . Is there code in the API server that deals with these cases (etcd/API server node paritioned) and disconnects clients (REST, SDK) appropriately? Or are we at risk of having dangling (hanging) consumer-side client watches due to not using this option in the etcd client? Related discussions: /sig api-machinery cc/\nwas this something that was considered already?\nWe should be using for watches. (see also the ). It should just need to be . It would also be valuable to confirm that the behavior works as expected both when establishing watches and when a etcd member is partitioned after the watch has been established.\nIf anyone wants to own fixing this please self assign, else feel free to assign to me.\nFor this method: It seems we should do:\nThanks! This is a good question. I am not exactly sure what is the semantics on kubernetes watch. It is correct that without , etcd server will not cancel watch stream when itself is isolated from the cluster. However, when this happens, other requests to etcd (including health checks on etcd) will fail. So the control plane is already in a bad state and is likely to be replaced / repaired either automatically or manually. IIUC, this means apiserver will restart, the its clients will restart their watches.\nOh I just saw Joe's earlier comment. If using aligns with k8s watch semantics, we should use it:)\ni'd support moving to 3.4.5 to pick this up!\nI am happy to contribute but want to do some more testing and validation on the current behavior of and with the proposed fix. Please give me some days.\nQuick update: after some quick tests it seems that does not solve the issue with an active after the connected etcd endpoint becomes partitioned. I think we have the following scenarios to consider: 1) Single node API server/etcd 2) Quorum-based cluster (e.g. 3 nodes) where K8s control plane is colocated with the etcd nodes (e.g. one API server and one etcd instance on one node) 3) Quorum-based cluster (e.g. 3 nodes) where K8s control plane is not colocated with the etcd endpoints (separate nodes for K8s control plane and etcd nodes) We can rule out (1) because technically a partition with just one node in the cluster (K8s control plane and etcd) cannot happen. (3) is mostly addressed with existing gRPC timeouts and keep-alives so when an API server connected to an etcd node that gets partitioned this API server will reconnect to other reachable etcd nodes (endpoints). However, this is only true if the API server itself is not within that partition (e.g. multi-zone deployment where a whole zone becomes partitioned; read on for my observations). (2) is the interesting one: an API server is connected to an etcd node and both become partitioned from the majority. But the API server has active watches against this etcd node and serves these (Kubernetes) watches to Kubernetes clients (controllers, etc.). I was expecting that the gets evaluated even after establishing the . But it seems my assumption is not correct, i.e. it is only evaluated (checked) when is called. At least I did not see an existing being dropped after partitioning the nodes from the quorum (I did check the underlying gRPC transport status and it remained in stayed for as long as the partition was not resolved. So what does that mean for the issue raised here? One could argue that if more than just one etcd node is partitioned, i.e. also the API server and perhaps other Kubernetes controllers running in that partitioned environment (e.g. a zone), it makes no difference using because: the clients (custom controllers, etc.) are also partitioned and thus cannot connect to any other API server anyways; would be nice to return an error here instead of hanging but the system won't self-heal during the partition even if the clients (custom controllers) would not be in that partition, the corresponding API server could not be reached anymore and client-go session handling should address this (e.g. reconnect to another API server) you write: I did not see an existing watch being dropped, e.g. using this simple code: Once I partition the etcd node including this client it will not drop the . Looking at the definition of it seems that this is performed only during establishing the but not reevaluated as long as the client has a successful connection to the etcd node (e.g. node is partitioned but not dead). Since I am by no means an expert on the etcd semantics here, I might be doing something wrong. It would be great to get some clarification whether this issue (I raised, I know :) ...) needs to be addressed, documented, or can actually be addressed at all with the current implementation of .\nPlease see analysis above. should this work when the etcd + apiserver are on the same node or not?\nThanks for the careful analysis My take on this is: For case (2) where the apiserver and etcd are co-located (and the apiserver only talks to that etcd), I don't see any reason why dropping the watch would help. In this type of setup, when etcd becomes partitioned, the co-located apiserver becomes unavailable and apiserver clients should route away from that apiserver entirely until the partition is healed. Even if the apiserver were to drop the watch, it is still unavailable. For case (3) the behavior is already better than I expected. optimizes the case slightly by ensuring the initial watch connection is against a non-partitioned etcd, so it's still worth enabling, but not urgently needed since failover is already in place. We should document in etcd that only applies to establishing the initial watch connection, the documentation is currently misleading.\nIt is correct that without WithRequireLeader, etcd server will not cancel watch stream when itself is isolated from the cluster This is correct. Edit: Sorry I missread the test. IIUC, this is not correct. After receiving the watch request with , etcd server will continue to monitor its leader for this watch stream. When there is leader lost observed for \"maxNoLeaderCnt election timeouts\", server will cancel the watch stream.\nThanks that means the documentation is correct then. So while we haven't found any pressing issues that would fix, it does do what it should do, enabling it in the kube-apiserver won't make anything worse, and might make failover slightly faster?\nMake sense to me.\nBTW, I just did a quick test. an etcd cluster of 3 members: A, B and C. member B and C some time, the etcdctl command prints\nSorry for my delayed reply, wanted to do some more tests before replying. I can confirm this scenario. However, the tests that I did and which caught my attention because it was not what I expected from reading the doc string was to not shut down the peers but isolate one peer, i.e. simulate a partition which technically is also a \"lost leader\" scenario. In this case I saw the watches not being closed (neither for nor programmatically). Not sure if this is a bug or by design but I think it's more likely to have partitions than loosing the whole cluster where a majority of nodes is terminated/crashed (stop-failure). If you think it's worth and not expected behavior I am happy to open an issue in the etcd repo and provide more details. Even though in etcd makes it sound should address the partition scenario. Note: I'm currently reviewing the network partition tests in etcd which state that the results I'm seeing are not expected. Update: looks like I'm experiencing what was reported a while ago in but it's unclear to me whether it's fixed, unfixable or what the workaround (from K8s perspective) would be.\nI'm in favor of you opening a new against etcd for the watch not being closed in the lost leader scenario (and linking to this discussion). If it turns out it is a duplicate we can always mark as such.\nhas some detailed watch behaviors.\nFYI. There was a bug in . The fix was backported to = v3.3.19 and v3.4.5\nThank you and for your feedback. After reviewing the docs and I think the issue with and how to deal with partitions is still open (see TODO below). The scenario where clients (Watchers or any client indirectly consuming from an etcd ) would hang (forever?) is when the client gets partitioned with the corresponding partitioned etcd node OR when a whole zone/dc is partitioned. Since etcd picks safety (consistency) over liveness (availability) in this case at least not data corruption occurs but the overall system would be stuck when using watches instead of PUT/GET RPCs with timeouts. It would be ideal to throw a useful error, e.g. \"lost leader\", instead of hanging in \"gRPC TRANSIENT_ERROR\" state (which is what I'm seeing). Until this is resolved in etcd, perhaps the docs could be updated as is a bit misleading here. I will comment on the etcd issue above and file a PR for the K8s watcher implementation to make use of as it is still a useful option when (re)establishing the stream. Depending on the environment, API server instances could have a health/liveness probe (or systemd equivalent) which could restart the process in case of a partition. Upon reconnect to the etcd node with the option it would immediately fail and error out accordingly."}
{"_id":"q-en-kubernetes-6abaff1c06fb0fe2ce1c10cb553e48c751ac045e2d865fcb8c21a5cd2c228909","text":"Hi, Since my upgrade to 1.18 version, I have errors in kube-proxy: I have ipv4/ipv6 dualstack enable. No problem with cluster and IPVS works despite errors. Do you have also this issue ? OS: RHEL7 Kubernetes version (use ):\n/sig network\nRelated to cc\n/cc\nTaking one of the outputs it gives me the following addresses? can you share more details about your environment, Is it possible that the CoreDNS pod has the address ? what are your cluster-cidr and service-cidr? Indeed that new net.Utils change looks suspicious , a quick grep in the repo gives me some suspects, maybe the service/ipallocator? It's important to mention there were similar issues with IPVS before, interpreting IPv4 addresses as IPv6, has some experience with this kind of issues, maybe he can see something:) /cc\nA bit tricky to trace since libnetwork/ipvs is moving around at the moment. Problem seems to be in; The \"family\" must be wrong. It is called from;\nPlease do a and attach the output.\nHi, Some informations: cluster-cidr and service-cidr: fd00:10:245:0:0:0:0:0/64,10.245.0.0/16 fd00:10:96:0:0:0:0:0/112,10.96.0.0/12 ipvsadm output:\nRef:\nThe commit in the ref above should be in v1.18.0 which is the version used in this issue. So the commit does not fix the problem.\nMaybe something wrong around here Before , it looks like this: Now it looks like this: However is not initializedit is zero :(.\nI do not get these faults. Has the netlink API changed? What kernel version are you using?\nd.AddressFamily is supposed to be set from the \"NetlinkRouteAttr\" slice; It should definitely not be hard-coded to ipv4.\nMade a trying to reproduce this in my environment but didn't got this error. I may try this on some older kernel like from CentOS 7 Added exactly the same IPVS configuration here: And when running this 'gist/program' pointing to the IPVS from the DNS (fd00:10:96::a port 53 UDP) got the expected result: Maybe trying to run this also in some other scenarios could clarify if this is something related to Netlink API changed between Kernels, something with the used library, etc.\nAfter some study, I find that this issue most likely is caused by the low linux kernel version. I have reproduced this problem on . Accroding to the code, netlink will try to get attribute, but I find the kernel just does not support it :( This is the defined in (kernel 3.10) No attribute is defined! But in new kernel version, the is defined like : Obviously, the kernel has some attributes (IPVSDESTATTRADDRFAMILY, IPVSDESTATTRSTATS64...). That is why kube-proxy works well on systems with a higher version of the kernel. So we can address this issue by upgrading our linux kernel. Not sure which kernel version is the minimum requirement, maybe we should document something about this.\nHere is the defined in :\nGood finding. So, we can conclude that this really is kernel version dependent. It is hard to know what to do about it. We must check the family. I would propose that we make it work for ipv4 with some fallback and put a requirement on kernel version for ipv6-only/dual-stack. But the decision is others to make. /cc\n/cc\nThe symbol in the kernel tree The for service family seems very old. Since the service must have the same family as the real-servers perhaps it can be used for compatibility with old kernels?\nThe commit that adds the symbol; Seem to go into linux-3.18\nI did not know the minimum required for the kernel version, I was using the stable and official version of RHEL7. My kernel version is: Currently the max version in the repository is: According to the documentation (), I should switch to RHEL8 if I want a recent kernel :-\nThere is no minimum kernel version requirement afaik. So I think this must be fixed for ipv4-only clusters IMO. However ipv6-only and dual-stack are in \"alpha\" so a requirement on kernel version might be ok, but I can't say myself. A complication is that the problem is not in k8s itself but in a 3rd party lib.\nipv6-only is beta since 1.18 :smile:\n/assign\nI opened a PR to fix this in moby/ipvs If someone w/ a cluster on the older kernel version can test it that would be great, otherwise I will setup a cluster to validate this some time this week.\nInstead of bootstraping a whole cluster right now (pretty busy here) I can use the Gist I've mentioned to check in a CentOS 7.1 if it faces the problem before and after the patch. At least to have some more test right now\nwith a small test, it worked fine. Environment: CentOS 7.1 - Kernel 3.10.0-229.el7.x86_64 + ipvsadm 1.2.1 Pre-loaded some IPv6 rules into ipvs: I've used code, searching for the following structure: I then call ` to get the IPVS object, and then try to fetch the destinations. The following are the results: Before patching the ipvs module: After patching the ipvs module (pinning to your repo/branch) Also I've made those tests with the structure but searching for TCP, and for kube-apiserver (that have only one destination) and it worked fine in all those scenarios. If wanted I can improve the gist to pass the parameters as command arguments instead of hardcoded (sorry, did this in a rush to test)\n/triage unresolved Comment when the issue is assessed and confirmed. I am a bot run by vllry. ‍\n/remove-triage unresolved\nThanks just waiting on now\nsame issue,sloved by update kernel to 5.X\nI didn't want to wait, I updated to kernel-lt in the version 4.4 . The problem is solved on my side :)\nThanks for confirming. We'll still try to get in since I'm sure other folks with older kernels will run into this issue.\n/retitle parseIP error with IPVS proxy\nupgrade your system kernel to 4.xx edition\nFYI: just merged, need to run some validation & cut a new release (v1.0.1).\nFYI\n/retitle DualStack / IPv6-only: parseIP error with IPVS proxy on CentOS 7\nv1.18 cherry-pick\nHi, it seems that you're using v1.18.0 and this has been corrected in v1.18.3, can you please update and check? Tks"}
{"_id":"q-en-kubernetes-1a4314174503923db08a51edfe0f79a1a6baa297ceb5f23812a8fb28b8f36c2d","text":"<!-- Please only use this template for submitting reports about flaky tests or jobs (pass or fail with no underlying change in code) in Kubernetes CI --Which jobs are flaking: pull-kubernetes-integration Which test(s) are flaking: TestSubresourcePatch Testgrid link: Reason for failure: Unknown. A diff with the original object would be helpful. Anything else we need to know: /sig api-machinery\nOK Trying to reproduce, see what's going on.\nClearly this is because SSA is updating the tracked entry's time. It happens when the two request don't happen in the same second. Having a between the two requests makes this into a definitive error.\n/wg apply\nThis seems to be because of . A repeated patch like will always trigger a new managedFields entry to be created, since it's always going to change the object (status or spec will be wiped depending on the endpoint), so the patch will always cause the object to change, which will always cause the entry to be created.\n/assign\n/assign\n/assign\nStill seeing this, bumped into it over in\ndeflake in while the underlying issue is being worked on in"}
{"_id":"q-en-kubernetes-a352408adf15ef110289db611e1e32e8429277034907655174e7203ab9fa896d","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: When creating a new cluster using , panics with SIGSEGV on first run: Note that after the crash the pod restarts, and the second run is successful: and everything seems to be working as expected. What you expected to happen: starts successfully the first time. How to reproduce it (as minimally and precisely as possible): Create a new cluster using , with in mode. Anything else we need to know?: Environment: Kubernetes version (use ): 1.18.0 Cloud provider or hardware configuration: Hetzner VPS OS (e.g: ): Ubuntu 18.04.4 LTS Kernel (e.g. ): 5.3.0-45-generic (Ubuntu's default HWE kernel) Install tools: kubeadm Network plugin and version (if this is a network-related bug): Others: using CRI and mode Please let me know if you need any additional information.\n/sig network\n/assign\nAccording to the panic output, the crashed code is For some reason, might be nil, we should always check it before using.\nDo we understand WHY it is nil?\nOne more thing I've noticed while fiddling with the cluster is that if I remove the node and re-join it, starts just fine, so I'm guessing it has to do with the initial state of the server / ipvs.\nThat would be my guess too. And that sort of race is worth fixing :)\ndo you see anything in the logs in the form: ??\nI think somewhere here we're failing to initialize the IPVS interface. We probably shouldn't ignore the error from also :grimacing:\nNope, I posted the (full) logs in the issue description.\nThanks, will look into it deeper\ndo you happen to have a start up script that loads the necessary IPVS modules or is it preloaded on your nodes? Wondering if there could be a race between when the modules get loaded and when proxy starts up. Either way we should log the error ().\nNeither, but if you give me the module list I can try to load them before creating the cluster to see if I still get the panic.\nFor kernel <4.19: For = 4.19:\n/triage unresolved Comment when the issue is assessed and confirmed. I am a bot run by vllry. ‍\n/remove-triage unresolved\nI can confirm that, if I explicitly preload the modules, starts successfully the first time.\nAccording to this log , we can confirm second check success, missing modules loaded after first failed check. Because missing modules loaded but not put to"}
{"_id":"q-en-kubernetes-57bd8a52613b2d9eb488199535053c58e9b54764a050d4dce81013f5f996a398","text":"Hi, I have a issue with a dualstack cluster for my Kubernetes Dashboard deployment. I opened a ticket here (), but I am informed that it is not related to Kubernetes Dashboard. What is the problem? Errors in kube-api: The service is created with the correct ipFamily (=IPv6). My endpoint is fully in IPv6. Indeed, my pod has one IP in IPv4 and another in IPv6. Do you have also the issue? Kubernetes version: 1.18 OS: RHEL7\n/sig network /sig api-machinery\n/triage unresolved Comment when the issue is assessed and confirmed. I am a bot run by vllry. ‍\nI think the following should fix the error you saw:\n/remove-sig api-machinery"}
{"_id":"q-en-kubernetes-10e0800334eabc101ba3e719bde6addc76496fd00734e09bb571fcc5b6be7d48","text":"What happened: In a dual-stack cluster when specifying both ipv6 and ipv4 addresses in the entries are inserted in ipvs for the ipv6 nodeports but with an invalid ipv4 address; The example shows the ipv4 loopback address but any ipv4 address in will cause a faulty entry. What you expected to happen: No ipv4 entry shall be inserted for the ipv6 nodeport. How to reproduce it (as minimally and precisely as possible): Configure with both ipv4 and ipv6 addresses Bring up a dual-stack cluster with proxy-mode Create a service with ipFamily:IPv6 Do Anything else we need to know?: Proxy-mode=iptables seems to work. With no ipv4 entries are inserted for ipv6 nodeports. In it is recommended to use but doing so makes kube-proxy enter rules for address (in addition to \"127.0.0.1\"). While these entries cause no harm they can't be used and should not be there. These entries are created for both proxy-mode \"iptables\" and \"ipvs\". This should not be handled in this issue but in another (new) issue if needed. Example; Environment: Kubernetes version (use ):Cloud provider or hardware configuration: Kvm's on Ubuntu OS (e.g: ): xcluster Kernel (e.g. ):Install tools: \"from scratch\" Network plugin and version (if this is a network-related bug): calico/cni:v3.12.0 Others: /sig network\n/triage unresolved Comment when the issue is assessed and confirmed. I am a bot run by vllry. ‍\n/area ipvs /area ipv6\n/remove-triage unresolved /assign I take this myself if no one objects.\nThe ipv4 address is obviously a ipv4-converted ipv6 from .\nAccording to the code here (), we may need to filter the before passing it to .\nLooks that that's it. Please go ahead and make a PR, or I can do it if you prefer. I will help test the PR of course. Here are some review comments right away; Use v1.IPFamily as param to the filter function rather than The filter function seem usable and generic and may be placed in some \"util\" package, but leave it as-is and let the maintainers decide that.\nHi feel free to send a PR please, as you have took this already :) I'd like to help review it!\nThe \"addresses\" are really CIDRs which make the fix even simpler."}
{"_id":"q-en-kubernetes-2d86227263004d2306fadc99355cf0a88aac1d38d7e2576116201cd81e3141b9","text":"What happened: In handler#handleSummary, return value from request.Request.ParseForm() is ignored. What you expected to happen: The error return should be checked.\n/sig node\nHi this seems to be easy to fix. Could I solve this?\nPlease go ahead."}
{"_id":"q-en-kubernetes-70a8102d9ccc47715542393177cb108296cc3d43660031402155c7b49d89619f","text":"What would you like to be : Please see Implement Instances::NodeAddresses so it doesn't ignore the node input. Why is this needed: The current implementation assumes it will be only run on the Kubelet. That is no longer a valid assumption and in fact it the nearish future it will no longer be run on the Kubelet.\n/sig cloud-provider /area provider/gcp\nThis issue is about the GCE implementation found at This problem cannot be fixed in the cloud node controller.\nAccording to is indeed only used for the address of the calling instance. Maybe we should use instead? Note that implements\nOr, we should definitely remove the if its TODO item is no longer true.\nI believe the TODO is no longer true as there is also a call in"}
{"_id":"q-en-kubernetes-f2f057b2b0803650cc4c635fc72b7b97c855411c53c3a5c781802ce26cf15068","text":"What happened: In , there is mismatch between method name and the method in logs. e.g. RemoteRuntimeService#Attach has the following: What you expected to happen: The method names should be consistent with what appear in the logs. Otherwise people would be confused when troubleshooting. /sig node\ncould i try to fix this?Thanks\nYes. fixes method name for ContainerStats. You can cover the rest. thanks\n/assign"}
{"_id":"q-en-kubernetes-4a931e19141565ae059dfa88de9dd4579f1331d05e6aac83653229ef31ddc6ca","text":"For external cloud provider - Kubelet setting the TaintExternalCloudProvider taint: The results in calling function in cloud node-controller: This (in a healthy case) sends 4 GCE API calls to: node addresses: (as part of getNodeModifier call) instance type: (as part of getNodeModifiers call) instance: (as part of updateNodeAddresses call) node addresses (again): (as part of updateNodeAddresses call) All those requests are read requests and could be replaced by a single call. We could do that by introducing a simple struct like: and adding a method (or rather replacing the existing InstanceTypeByProviderID which is used only for this purpose) to Instances cloud-provider interface; With that change, we would be able to replace all that 4 calls with just a single API call. This would be useful especially for creating large clusters (like 5k nodes), which may saturate the quota for API calls for quite some time. This is quite significant size of refactoring, though it should be relatively mechanical task to do. /sig cloud-provider /sig scalability /priority important-longterm /help-wanted - FYI\nThe InstanceMetadataByProviderID would be to every cloud provider, such as aws, azure, etc. right ?\n/assign Hi , if you're interested in it, I can work with you.\nYes - we need to do that (which is why I wrote it's quite significant refectoring). It also makes reviews harder, because we will need to ask maintainers of all cloud-providers. So it would be good to separate changes to different cloud providers into separate commits.\nI have one question: and first query by providerID, if failed then query by nodeName. While the logic in is converse(if providerID is nil, get providerID by nodeName, then ensure its existence by providerID). If we merge these three method into , how to deal with the query sequence of nodeName and providerID?\nI bet it's a consequence of different people adding different code paths. Those conceptually seem to be the same to me. Unifying this will be additional benefit from this change imho. So I would say that choose whatever is easier to implement.\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nI think and are the only PRs left for this. One other optimization that can be made is to cache the instance metadata but I think that's a provider specific implementation\nPlus enabling it for GCE if we decide to remove it from the first PR. But yeah, in general I agree.\n- can you please address ?\nYes, I have seen that.\nI think this issue has been resolved, feel free to reopen it. /close\nClosing this issue. <detailsIn response to : I think this issue has been resolved, feel free to reopen it. Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-bbf94b5e4f8874c9821c52cd68737c64ffa5b29bd76bbeec678d5c88e760b105","text":"We have some scripts that install docker/containerd etc and are having trouble installing on focal based images as does not yet support focal fossa. So in we have a patch that uses debs for focal. Additional context, please see ci test failures in and This issue is to track fixing the script back up when focal support is to\n/sig release /sig testing\n/priority important-soon /area release-eng\n/assign"}
{"_id":"q-en-kubernetes-b92e3c87d8a9bbce5e888fb81596443a14f95cbfa3a1fc2c1859b99b77537ddf","text":"<!-- Please only use this template for submitting reports about continuously failing tests or jobs in Kubernetes CI --Which jobs are failing: sig-windows jobs: Which test(s) are failing: [sig-network] Services should be able to change the type from ClusterIP to ExternalName [Conformance] [sig-network] Services should be able to change the type from NodePort to ExternalName [Conformance] Since when has it been failing: It was a false positive until recently. Testgrid link: Reason for failure: Partially Qualified Domain Names are not currently supported on Windows containers. The tests are checking PQDNs for validation. The test was passing until recently because was still returning the exit code 0, even if the given DNS name was not found: [sig-network] Services should be able to change the type from NodePort to ExternalName [Conformance] [BeforeEach] [sig-network] Services STEP: Creating a kubernetes client Apr 28 21:24:38.587: INFO: kubeConfig: /home/ubuntu/kube/config200424 STEP: Building a namespace api object, basename services STEP: Binding the e2e-test-privileged-psp PodSecurityPolicy to the default service account in services-2633 STEP: Waiting for a default service account to be provisioned in namespace [BeforeEach] [sig-network] Services [It] should be able to change the type from NodePort to ExternalName [Conformance] STEP: creating a service nodeport-service with the type=NodePort in namespace services-2633 STEP: Creating active service to test reachability when its FQDN is referred as externalName for another service STEP: creating service externalsvc in namespace services-2633 STEP: creating replication controller externalsvc in namespace services-2633 I0428 21:24:42. ] Created replication controller with name: externalsvc, namespace: services-2633, replica count: 2 I0428 21:24:45. ] externalsvc Pods: 2 out of 2 created, 0 running, 2 pending, 0 waiting, 0 inactive, 0 terminating, 0 unknown, 0 runningButNotReady I0428 21:24:48. ] externalsvc Pods: 2 out of 2 created, 0 running, 2 pending, 0 waiting, 0 inactive, 0 terminating, 0 unknown, 0 runningButNotReady I0428 21:24:51. ] externalsvc Pods: 2 out of 2 created, 1 running, 1 pending, 0 waiting, 0 inactive, 0 terminating, 0 unknown, 0 runningButNotReady I0428 21:24:54. ] externalsvc Pods: 2 out of 2 created, 2 running, 0 pending, 0 waiting, 0 inactive, 0 terminating, 0 unknown, 0 runningButNotReady STEP: changing the NodePort service to type=ExternalName Apr 28 21:24:55.544: INFO: Creating new exec pod Apr 28 21:25:06.371: INFO: Running '/home/ubuntu/rune2e/kubernetes/output/dockerized/bin/linux/amd64/kubectl --server=https://kubetest--83ef-11ea-ae60- --kubeconfig=/home/ubuntu/kube/config200424 exec --namespace=services-2633 agnhost-pod -- /bin/sh -x -c nslookup nodeport-service' Apr 28 21:25:09.610: INFO: stderr: \"+ nslookup nodeport-servicen kube- can't find nodeport-service: Non-existent domainrrn\" Apr 28 21:25:09.610: INFO: stdout: \"Server: kube-rnAddress: 10.0.0.10rnrn\" STEP: deleting ReplicationController externalsvc in namespace services-2633, will wait for the garbage collector to delete the pods Apr 28 21:25:10.429: INFO: Deleting ReplicationController externalsvc took: 423.628373ms Apr 28 21:25:10.830: INFO: Terminating ReplicationController externalsvc pods took: 400.686463ms Apr 28 21:25:15.792: INFO: Cleaning up the NodePort to ExternalName test service [AfterEach] [sig-network] Services Apr 28 21:25:16.230: INFO: Waiting up to 3m0s for all (but 0) nodes to be ready STEP: Destroying namespace \"services-2633\" for this suite. [AfterEach] [sig-network] Services â¢{\"msg\":\"PASSED [sig-network] Services should be able to change the type from NodePort to ExternalName [Conformance]\",\"total\":2,\"completed\":2,\"skipped\":2913,\"failed\":0} The issue was observed with an updated, which doesn't return 0 if the given DNS name was not found. Anything else we need to know*: /sig testing /sig windows"}
{"_id":"q-en-kubernetes-b93e267deb56d23bda0d28a0263a32df5ffd0f9938ba97576aec76ce8a64ccb2","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: On running kubetest with external storage driver option with driver name \"\", encountered error of flag redefinition. The issue happens because: As part of this patch , a flag is defined for number of inline volumes. The driver name is used to construct the flag name. Thus, when running tests on an externally provided driver (in this instance the Persistent Disk CSI driver) with driver name \"\", a redefinition error happens since the ephemeral test suite code path is executed once for and once for External PD CSI driver. [example run command: kubetest --test --check-version-skew=false --testargs=\"-- --storage.testdriver=/path/to/gcp-compute-persistent-disk-csi-driver/test/k8s-integration/config/test-\" Test config looks something like this: { StorageClass: FromFile: ... SnapshotClass: FromFile: ... DriverInfo: Name: ... } ] What you expected to happen: Test should run and succeed. How to reproduce it (as minimally and precisely as possible): Use an test-, with driver name as \"\", and run kubetest with focus as \"External.Storage\" kubetest --test --check-version-skew=false --testargs=\"-- --storage.testdriver=/path/to/gcp-compute-persistent-disk-csi-driver/test/k8s-integration/config/test-\" Anything else we need to know?: Possible fixes options: , (), when using a FromFile or FromExistingClassName option to pass in SnapshotClass information, pickup the provisioner name from the class definition's \"driver\" field. the flag which has a dependency on driver names. (need to understand the impact of removing the command line flag) Environment: Kubernetes version (use ): latest master Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others:\n/sig storage\n/assign Do you know why this parameter is set as a flag and not a part of the ?\nTo me, is something that defines what the driver can do. This flag is about what the test does and not related to driver capabilities. Making it depend on the driver name was probably unnecessary. A single flag for all incarnations of that test probably would have been sufficient and thus better (\"keep it simple, stupid!\" - the stupid being me here...). It also shouldn't have used the global flag set. In other words:\n/reopen I don't really want to get into the habit of defining per test case flags in There are thousands of test cases and it would be unwieldy to discover all the flags available and understand how to use them. What factors would influence how many volumes you want the test to create? Can we just hardcode a value in the test case?\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThe defaults must be usable so that users do not need to understand them. The flags can be discovered with - not ideal, but at least they show up. That's effectively what the default value for the flag does. This is a way to to run a custom stress test: the default is so low, it just tests that the feature works. By increasing it as high as the underlying CSI driver and storage system supports, one can run this (and potentially other tests) as part of stress testing. Not sure whether it's worth supporting this. It seemed worthwhile at the time, and better than hard-coding just one value.\nThe bigger issue is: how do users of the test suite actually learn what tests they run and what those tests do? I suspect the answer to that is simply \"read the source\", in which case they'll also find the knobs that can be used to tweak testing.\nSince increasing the parameter depends on the csi driver/storage system, then I would recommending splitting this up into 2 test cases: 1) test basic functionality 2) test stress scenario, and add the parameter to the driver test config similarly to other stress tests:\nAt least for the storage tests, because we are intending that they be run externally, having all the tunables in one place I think makes it easier to discover.\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle rotten\n/remove-lifecycle rotten\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale /help\nThis request has been marked as needing help from a contributor. Please ensure the request meets the requirements listed . If this request no longer meets these requirements, the label can be removed by commenting with the command. <detailsIn response to : /remove-lifecycle stale /help Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle stale\n/remove-lifecycle stale\nI saw that the flag was removed in , I think we can close this one if doesn't apply anymore\n/close The flag was removed in #diff-20517805986874f69e3254bed93d59996d5a6fd571a70ab8120736ded5aafa24\nYou can't close an active issue/PR unless you authored it or you are a collaborator. <detailsIn response to : /close The flag was removed in #diff-20517805986874f69e3254bed93d59996d5a6fd571a70ab8120736ded5aafa24 </details\n/close\nClosing this issue. <detailsIn response to : /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-04298a57bb6bbadab2ae389c253044418da070e519876e8a987add6c4a3d1714","text":"[x] Original approval issue: Start 9 June 2020 3 Flakes No Flakes for 2 weeks According to this APIsnoop query, there are still some remaining ReplicationController endpoints which are untested. a ReplicationController with a static label the ReplicationController with a new Label and updated data the ReplicationController's status /Scale the ReplicationController's Replica count to 2 Replicas the ReplicationController the ReplicationController's Status all ReplicationControllers in all Namespaces find the ReplicationControllers(1) ensure that the ReplicationController is found and is patched Namespaced ReplicationControllers(1) via a Collection with a LabelSelector Discover useragents: See all endpoints hit from this test Display test coverage change If a test with these calls gets merged, test coverage will go up by 7 points This test is also created with the goal of conformance promotion.\n/sig testing /sig architecture /area conformance\nCan you please add in PR\n/reopen Merged. Promotion to be created\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale"}
{"_id":"q-en-kubernetes-a6d16665f1e6762b35bdab01dd8cab20b8ce126e21fc5362b5abf78e7b1ae18f","text":"[x] Original approval issue: [x] two weeks soak end date: 2 July 2020 This issue is created to allow edit by According to this APIsnoop query, there are still an endpoint which is untested. /wip /hold a set of 3 Pod Templates with a static label within a Namespace. that all 3 Pod Templates with the label in the Namespace are created. the set of Namespaced Pod Templates with a label via DeleteCollection. that all Pod Templates with the label have been deleted. Discover useragents: List endpoints hit by the test: Display endpoint coverage change: If a test with these calls gets merged, test coverage will go up by 1 points This test is also created with the goal of conformance promotion.\n/sig testing /sig architecture /area conformance\n/area conformance\nDuring Conformance Meeting: Is there a way to write this test so it is resource agnostic? If we are just changing the resource... can be parameterize the resource-type\n/assign\nGitHub didn't allow me to assign the following users: heyste. Note that only , repo collaborators and people who have commented on this issue/PR can be assigned. Additionally, issues/PRs can only have 10 assignees at the same time. For more information please see <detailsIn response to : /assign Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign\nGitHub didn't allow me to assign the following users: heyste. Note that only , repo collaborators and people who have commented on this issue/PR can be assigned. Additionally, issues/PRs can only have 10 assignees at the same time. For more information please see <detailsIn response to : /assign Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/reopen Merge 17 July 2020\nReopened this issue. <detailsIn response to : /reopen Merge 17 July 2020 Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-cebd916f3f284efd9f87fc49f447a1a462f9211f370f6b99d9d95f3fd2a1ed17","text":"What happened: I checked if kubelet () has permissions for with SelfsubjectAccessReviews and I saw that it has. But when I checked its permissions with SelfsubjectRulesReviews it didn't show it. What you expected to happen: I expected that the API SelfsubjectRulesReviews will show that has permssions for . How to reproduce it (as minimally and precisely as possible): You can check it on the master or the worker. Test SelfsubjectAccessReviews: You will get answer that it has permissions: Test SelfsubjectRulesReviews: It won't show the permissions. Anything else we need to know?: Environment: Kubernetes version (use ):Cloud provider or hardware configuration: AWS OS (e.g: ):Kernel (e.g. ):Install tools: kubeadm /area/api /kind/bug /sig api-machinery\n/sig api-machinery\nThis is a more clean display that show the bug: I think the problem is related to the fact that the file contain the kubelet certificate which is part of the group :\n/sig auth\nWhat is the SelfSubjectRulesReview response from the server ()? Note that not all authorizers support enumerating permissions, and the API object status includes an attribute which explicitly indicates the result is incomplete. See\nIt has Details:\nI noticed that it also doesn't show that it has permissions to view and . It probably related to the Node authorizer: Because I assume that these rules doesn't appear in RBAC.\nThe node authorizer does not enumerate permissions, so it does not contribute to SelfSubjectRulesReview results\nfixes setting correctly when the API is requested with a kubelet's credentials"}
{"_id":"q-en-kubernetes-779de3284752604a311018afe5a276b30f7b4d99b18a50c633ee35fec5876197","text":"<!-- Please only use this template for submitting reports about continuously failing tests or jobs in Kubernetes CI --Which jobs are failing: Which test(s) are failing: Since when has it been failing: Testgrid link: Reason for failure: Anything else we need to know: /cc /priority critical-urgent /milestone v1.19\n/sig api-machinery\nSuspect it has to do with cc /assign (for CI signal ownership)\nSeems unlikely since multiple kind e2e tests passed on the PR itself\nI was thinking that as well, but I don't see any other relevant changes\nActually, I take that back. Looks like the same failure was encountered as a flake on some of the PR runs. Looking at the test, it appears to assume the resource version of the kubernetes service will always be in the watch window, which is a bad assumption. I'd expect us to fix that assumption in the test, but also make sure we didn't unintentionally shrink the watch window for resources.\n/assign\nNote: this also appears to be breaking\nFailure seems to be timeout related as well shows the runtime ~tripled after merged\nrevert open in can investigate test failure and timeouts and bring the change back in\nI take it back, the test tolerates an expired error explicitly.\nshowed 17 of the 22 runs since merged took ~1 hour. The first run after the revert at returned to less than 30 minutes for the run. Will keep an eye on that to see if subsequent runs match.\n- thanks a lot for looking into it. I don't think my PR changed anything that anyone should rely on. I will take a deeper look into it and try to fix that (hopefully later this week).\ncontains the explanation and describes the bug (in fact this bug still exists in the codebase, that PR was only increasing the probability of this particular test hitting it)"}
{"_id":"q-en-kubernetes-e050be9174826390929e6ad32b4aeb5b625e5d3f65d51a76662ce8da7a709c6c","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: UT TestValidateNodeIPParam does not really tests . The first IP and interface are always 127.0.0.1 and lo and will break the loop. So will never be appended into . What you expected to happen: gets tested correctly. How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others:\n/sig node\n/sig node"}
{"_id":"q-en-kubernetes-ca234716f87c4e407844c0bad1bee9d9159a3c006a9970465d985a03768552c6","text":"apiserverrequest (total, latency histogram) metrics are likely the two most important metrics emitted by the kube-apiserver, aggregator, and aggregated-apiserver binaries. has been open for a year regarding some of the issues that have been ongoing problems. I want to open up a discussion on: (1) cleaning this metric up so that it can be used reliably and consistently for many versions addressing as many of the cardinality concerns as possible (2) explicitly defining a set of prescriptions for adding new metrics which may require joins to this very centrally important metric These metrics are critically important because people use them for SLIs/SLOs. In fact we define SLOs for them here (). Yet, due to the fact that people can inject arbitrary webhooks into the request flow process, we can never actually guarantee that the apiserver response time actually meets our stated objectives. Which brings us to the other dimension in which these metrics are flawed. They do not account for apiserver request processing time, exclusive of webhook processing time. As long as we do not break out internal request processing time, it is possible that arbitrary clusters will never be able to meet our internally stated service level objectives given the one metrics we provide.\n/sig api-machinery instrumentation /assign\nThe cardinality issues hit us here pretty badly, just one example for this, I suppose: Soo :100: % agree.\n/milestone v1.20\nHeyho I wanted to check how it is going? The code freeze is starting 12.11.2020 (about 3-4 weeks from now) and while there is still plenty of time, we want to ensure that each relevant issue has a chance to be considered. As the issue is tagged for 1.20, is it still planned for this release?\nAh, this is going to be slated for 1.21. /milestone v1.21\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\n/open\n/re-open\nHello! I am from the Bug Triage team! Friendly reminder that code freeze is starting March 9th, 2021 (today!). As the Issue is tagged for 1.21, and the PRs linked are merged and marked for 1.21, can we close this one?\nYes! :)\nThanks !"}
{"_id":"q-en-kubernetes-7a28318c84bd39d7ccfbb272be61cb2824eb24ba88bc20a89720f22fd15fa079","text":"What happened: does not handle \"holes\" in interface device-number correctly, dropping later interfaces. Create instance with eth0, eth1, eth2 ENIs as usual. Remove (eg: by selectively deleting pods) eth1. Only the addresses on eth0 are reported by NodeAddresses. What you expected to happen: Addresses on eth0 and eth2 should be reported.\n/sig cloud-provider These SIGs are my best guesses for this issue. Please comment if I am incorrect about one. I am a bot run by vllry. ‍"}
{"_id":"q-en-kubernetes-6184445b7c28cc3e6bc3480e701994fd5543b639f72f4fc32cc9643d4062c630","text":"See - currently there's just one test that triggers Unreserve through a pre-bind failure.\n/sig scheduling\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-testing, kubernetes/test-infra and/or . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle stale\n/remove-lifecycle stale\n/good-first-issue\nThis request has been marked as suitable for new contributors. Please ensure the request meets the requirements listed . If this request no longer meets these requirements, the label can be removed by commenting with the command. <detailsIn response to : /good-first-issue Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign I shall work on it.\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle stale\n/remove-lifecycle stale\n/assign\ni saw commented the only test code, do you have any idea why this going on?\nwell, I found some functions missing, maybe because of the code refactor.\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\n/remove-lifecycle stale\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\n/remove-lifecycle stale"}
{"_id":"q-en-kubernetes-cb5567adfea5d580e9f4efa2cb249bb7332e9b2a05351c6940569eebb0a18ebf","text":"<!-- Please only use this template for submitting enhancement requests --What would you like to be : The ebtables utility in is missing 3 useful features: method in its Interface table chain Why is this needed: The missing features are useful to allow for: rule deletion as there is a way to add (ensure) a rule but not removing it currently. manipulating the broute table and brouting chain The first feature would also align ebtables with iptables, as a DeleteRule method is available there.\n/sig network\n/sig network\n/triage unresolved Comment when the issue is assessed and confirmed. I am a bot run by vllry. ‍\nIs there some feature/bugfix you're working on in kubernetes that needs this?\nYes, in kubernetes/dns (node-cache) actually. That PR is working, but it's only in my fork of kubernetes/dns at the moment as it depends on this. It solves this: Adding this code to ebtables would help me avoid re-inventing the wheel there.\n... not sure if you've seen that ... but I have created a PR to solve this issue already:\nYes, I saw the PR first. So, kubernetes/kubernetes isn't really intended to be vendored into other repos and kubernetes/dns is being bad here. It's dubious to add methods to that kubernetes itself isn't using because someone else might just end up removing them later when they notice that they're not being used... OTOH the PR needs access to the internals of the ebtables code so you couldn't just add it to kubernetes/dns directly without forking all of ...\nThanks for your comments. I saw iptables used that way in node-cache and I thought it would be good for consistency to go down the same route with ebtables. To be honest I am not very keen on forking the entire package, for obvious reasons, but I can understand your reservations about this change.\n/assign\nHi Did you manage to discuss this?\nyes, sorry, didn't get a chance to get back to this on Friday. (I /lgtm'ed the PR) /remove-triage unresolved\nThanks!"}
{"_id":"q-en-kubernetes-a49c37eb1701bd6842c7cc14a5401a7d3c1131c4fad04457f9a240141a4d7e5c","text":"/sig testing /sig release\nI'll work on this one /assign\nthanks! /remove-sig release (on second thought this is just SIG Testing)\n/lifecycle active"}
{"_id":"q-en-kubernetes-5641f3f29b124384c57cf073606440c648ac75ef3d0c6c559319a60adb8444b2","text":"What happened: Running manually includes this file in LICENSES. Running complains if it is present. The file was incorrectly dropped in This needs to be resolved before 1.19 ships /priority important-soon /milestone v1.19 /cc\nThere are no sig labels on this issue. Please add an appropriate label by using one of the following commands: - Please see the for a listing of the SIGs, working groups, and committees available. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/area code-organization\nfixed in"}
{"_id":"q-en-kubernetes-3a8743db3bf29458e5a27c17c50bcd7d0d79fac70a15136fee44de5d7d666517","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: kubelet fails to create kubepods cgroup on larger machine with 512 cpus, fails with the following error: What you expected to happen: Run kubelet properly without any error How to reproduce it (as minimally and precisely as possible): Start the kubelet on a machine which has a more processors Anything else we need to know?: Environment: Kubernetes version (use ): 18.5 Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others:\n/sig node\nWhen I changed CgroupDriver to cgroupfs, kubelet is running is fine:\nAnd here is the sample code which is creating this issue:\nso this is essentially a runc/libcontainer problem or a problem with the host configuration - e.g. the host having less than the requested ? this line: hints that your snippet is for an older version of runc. rc8 renamed it to \"Manager\": is that the version of the kubelet too? k/k 1.18 is using rc10: so ideally the match test for kubelet 1.18.5 should be against rc10 of runc. and current master of k/k is at rc90: are you able to reproduce the same with rc90?\nI tried running with master kublet which does use rc90 build and kubelet fails with the same error: complete log: I think this is happening because we are trying to create a cgroup more than a supported range for the Linux kernel: 2^18 = is the max value we can use for the CPUShare, so wondering whether can add logic into the kubelet to address this issue when we run on bigger machines(more than 256 processors)\n/cc\ninteresting snippet in moby: ! where\ni think we have a misconception here that CPU shares == # of CPU. \"cpu shares\" is a weighting proportion to give a container more resources in comparison to other running containers: the reason for 2^18 in the kernel is to have a larger integer scale to work with. unlike docker that has (nanocpu / fractional cpus) (): does not have such an option. but it does expose (CpusetCpus) which accepts a such as . i have a wild guess that this same problem is not present if you use dockershim. to see more on what is happening you could: run the kubelet with --v=4 or more. patch the kubelet to add debug code.\nsorry for the late reply. I think the issue is not in libcontainer, but in the Kubelet asking for a out of range value for cpu shares. I am not sure how it makes sense to set the number of shares in relation to the number of available CPUs, since it is a weight and not a limit. The error you are seeing is coming straight from systemd: Since there is already a check for the minimum value, I think it makes sense to add also a check for the maximum allowed (while we figure out why cpu shares are used this way). Could you give a try to the current patch? I've nothing close to 512 cores to validate it really solves the issue you've seen :smile: I'll open a PR once we are sure it solves the problem\nYou are right, this is coming from the systemd library which is getting used in the libcontainer With the above patch, I see kubelet is running fine but I don't think this is just good enough because while requesting the cpus in the pod is in terms of milicores which is 1000th of 1 cpu which is getting converted into and assigned to the pod's cgroup. e.g: if the user requests a cpu of which 1024/ = ~2 cpu share out 512 which brakes the current logic in the code. I think we need to add an additional logic for allocation based on the number of cpus present in the machine. Simple math should work fine if cpus are more than 256 in the system: e.g: if the request is 1024m 1024/(()/) = 512 shares\nthe value doesn't really limit the cpu time a process gets. It is a weight used for the cgroups at the same level. It looks to me that the Kubelet is using this mechanism simply to maintain some sort of priorities among the different pods (i.e. if pod A has double millicpus than pod B, then A gets double weight than B). Before adding more complexity, I'd make sure why it is done this way, I think it is just wrong to put in relation to the number of cores. The cpu shares assigned to each pod though are not dependent on the number of cores, but they are converted directly from the spec file. You've reported the issue on systemd, and it happens since systemd adds another check that is not present with cgroupfs or the kernel. The kernel does the same clamping I've proposed above: so the patch is just to make sure the systemd cgroup driver behaves in the same way as cgroupfs; and I think it doesn't have any impact on the pods\nI've opened a PR here:\nThanks for the PR, my only worry about - how k8s will prioritize when we have two pods requesting more than the MaxShares how this logic will prioritize it!\nfor now it works in the same way as in the cgroupfs case. Pod requests are not in relationship with the # of cores, that AFAICS happens only when setting up the node cgroups. I think a static allocation would work better for such cgroups, as anyway is a weight."}
{"_id":"q-en-kubernetes-32f4d04dfcc1bd22e814c5d0afd87b3c1dd279d24073534f782acdae857c0e4d","text":"These tests have flaked consistently since merged: in many cases, they are the only failure in a test run: Either the functionality introduced by that PR is not reliable, or the test is faulty. /sig node /priority critical-urgent /assign\n/milestone v1.19\n/assign\nLooks like this is failing ~5% of pull-kubernetes-node-e2e runs\nFYI\nLooks like there might still be a race somewhere looking at the kubelet logs Success case: Fail case:\nLooking at it more, it seems like is unexpectedly empty here Success: Fail:\nStacks of interest I think we might be dealing with a stale If there is a call before the container is to the Pod in the runtime and the is called within the 2 second default cache period, it seems possible that the in the could have no containers.\nwhat do you think of this analysis? If so, any obvious solutions occur to you?\nIf I'm correct, didn't introduce this; just the e2e test exposed an existing bug in the kubelet.\nThat all looks correct. I'm taking a look at the proposed solutions"}
{"_id":"q-en-kubernetes-f038e31e526d8c4211d1068605609b210c266a046b40220b41574e24d66d5e2b","text":"Which jobs are failing: ci-kubernetes-node-kubelet-conformance Which test(s) are failing: Variable Expansion should succeed in writing subpaths in container Variable Expansion should verify that a failing subpath expansion can be modified during the lifecycle of a container Since when has it been failing: Since Fri Jul 10 15:41:46 2020 -0700, when was merged Testgrid link: https://k8s- Reason for failure: Used DeletePropagationForeground option to delete test pod introduced by Anything else we need to know: The test pod stuck in Terminating state until 5m timeout of DeletePodWithWait expired. It happens with empty pod as well. Here is a simplified ginkgo test case that triggers this failure: The pod phase is \"Failed\" and container reason is \"Error\" for some reason: Reverting above commit or changing delete option to DeletePropagationBackground should fix this.\n/cc\n/sig node\n/sig storage cc\n/assign\nI think change just triggered the underlying issue. DeletePropagationForeground option doesn't work as expected it seems.\ndoes the ci-kubernetes-node-kubelet-conformance job run the controller-manager? the kube-controller-manager is responsible for deleting objects with a foreground finalizer on them, but I don't see any kube-controller-manager logs gathered in the artifacts for one of those test failures -\nI could reproduce the issue for commit (before 'generic ephemeral inline volumes' PR was merged) + 'use foreground deletion' commit.\nyeah, I tend to think switching that e2e helper method for all callers was not a good idea\nit looks like node e2e runs only etcd, api server and namespace controller: I don't think it makes sense to run controller-manager just to fix this failure. Would modifying test framework to set DeletePropagationBackground option for e2e node tests be correct solution?\nFor a quick fix, I would suggest just reverting the test change for now. Long term, yes, it should be configurable per test.\nAlso do we have plans to make this job a blocking presubmit test so we can find these problems earlier?\nI agree, we should revert , both to avoid changing behavior for existing callers and to match how pods are normally deleted\nif we need a PR for that I can prepare it.\nthat would be great, thanks\nputting my hat on, probably not? more hour+ tests that run on every PR are a non-goal\nThe SIG / OWNERs should enable testgrid alerts on failures, if periodic testing is not sufficient you can use a postsubmit to run whenever we merge code. We're not looking to rely further on presubmit, as is we're running too much there leading to excessive resource usage and inability to merge fixes like this in a timely manner between one thing or the other flaking heavily (and our odds get much worse with each new presubmit)"}
{"_id":"q-en-kubernetes-52f08f62336e8ee6f7a74524b939a90dafc9cb70ce2ec377c78b275b974d673d","text":"Which jobs are flaking: ci-kubernetes-kind-e2e-parallel pr:pull-kubernetes-e2e-gce pr:pull-kubernetes-e2e-gce-ubuntu-containerd pr:pull-kubernetes-e2e-gce-ubuntu-containerd ci-kubernetes-kind-ipv6-e2e-parallel Which test(s) are flaking: Second flakiest test in the e2e job according to Testgrid link: Reason for failure: Anything else we need to know: /sig network /assign\nneeds investigation since we're switching kube-proxy to endpointslice in this release\n/triage unresolved Comment when the issue is assessed and confirmed. I am a bot run by vllry. ‍\nmade me think of , but that's in already.\nIt's possible it's a test timeout that's just too close, but we need verification of that checking the controller and apiserver logs to make sure there isn't a latent bug\nThanks for catching this! It looks like the timeouts are actually associated with waiting for 2 pods to be ready, not actually anything related to EndpointSlice functionality - timeout is here: Given that, it seems like the obvious solution would be to update the Services to set to and update this block to wait for the Pods to have IPs, not for them to be ready. If that sounds reasonable, I'll work on getting a fix in tomorrow.\n/remove-triage unresolved\nIf that is still exercises the same conditions (I vaguely remember publishNotReadyAddresses changing code paths in the endpoints controller... does it do the same in the endpointslice controller?), that seems reasonable to me.\nI think the code path is very similar here, the only real difference should be how the EndpointSlice controller calculates readiness. I've gone ahead and filed a PR that uses this approach, hopefully it helps:\nThis flaked again this morning /reopen\nReopened this issue. <detailsIn response to : This flaked again this morning Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThis should be /priority critical-urgent given\nThe top level error message does not make this clear but the issue here is that too many EndpointSlices are being created - 3 instead of 2.\nInteresting... that condition should not persist for O(minutes), right? that seems like it might be a bug, not just a test issue\nThere's potential that it's not consistently merging 2 ports with 1 endpoint into a single slice and instead creating a slice per port. It also could represent a larger bug. Trying to find a way to consistently reproduce this.\nDespite my best efforts I have been unable to reproduce this on my own e2e cluster. I've had 10 test runs going at the same time repeatedly without any luck, any ideas for how to reproduce flaky test failures local would be welcome. In the meantime, I've created a follow up PR that will provide more helpful information if/when this test fails again: My leading theory right now is that this could be a caching issue. The EndpointSlice controller runs a sync, sees that we don't have an EndpointSlice for the given port combination in cache and then goes ahead and creates one, even though a matching EndpointSlice already exists and has not been cached yet. That would result in an annoying situation where an essentially duplicate EndpointSlice was created due to a stale cache. I would think the EndpointSliceTracker would resolve this issue for us though, so not sure. Getting the above PR in will help validate this theory. It's worth noting that if this is actually a duplicate EndpointSlice, clients like kube-proxy will handle this seamlessly because they already have to handle potential duplication in EndpointSlices. That would potentially explain why a test failure like this does not seem to be causing other things to break.\nIt appears that my theory was correct as witnessed here: I create a . The key points I've noticed so far: 1 EndpointSlice references pod2 (example-named-port-f7vq5) 2 EndpointSlices reference pod1 (example-named-port-fkgcc and example-named-port-vfp2q) 2 EndpointSlices were created at \"2020-07-28T04:57:51Z\": example-named-port-f7vq5 (references pod2) example-named-port-vfp2q (references pod1) 1 EndpointSlice was created 1 second later (\"2020-07-28T04:57:52Z\") example-named-port-fkgcc (references pod1) All 3 EndpointSlices are: On generation 1 - no updates have occurred With a state of ready: true The 2 EndpointSlices referencing pod1 are effectively identical  Of course this is something the EndpointSliceTracker should be handling for us, but even that approach seems less than ideal here. It wouldn't actually prevent the duplicate EndpointSlices from being created, it should just be cleaning up duplicates after they happen. This combined with the race conditions referenced in make me think we may want to find a way to slow down the sync cycle for each Service here. This also feels at least tangentially related to the need for better rate limiting in both this and the Endpoints controller discussed in . /cc\nHmm - actually I don't understand why this is happening. Maybe I'm missing something, but I don't think slowing down controller is needed to solve this problem. Even though the controller usually is not lagging my more than 1s, it is possible that it can be lagging by a minute or more in pathological cases. So we can't just rely on the fact that it's guaranteed to be synced. Can you explain it a bit deeper why this is happening? i.e. some more detailed timeline like: T1: state of cache is XYZ T2: controller computes the ES x1, x2 T3: state of cache in ABC T4: controller computes ES y1 [why?]\nAgree, I'm not in favor of slowing down the controller globally to address this. Delaying resync of a particular service until we observe the existence of new endpointslices we know we created (with some max delay) could be reasonable.\nThat's exactly where I was hoping to end up here, although I'm not quite sure how to do that cleanly yet. To be clear, this is not just to fix this flaky test. This flaky test is a symptom of a much larger and more annoying issue with the current implementation. The EndpointSlice controller is consistently running into errors where it's trying to update EndpointSlices that are stale. This is caused by follow up calls to running before the informer cache is up to date, which happens all the time during rolling updates where rapid churn of EndpointSlices is expected. Although this eventually resolves itself, it's causing a lot more load on API Server than necessary and can potentially result in some bad states where an update fails but a delete succeeds. In this specific instance, it looks like a syncService run is occurring after the first 2 EndpointSlices have been created but before the one referencing pod1 is in the informer cache. That results in the controller creating a new EndpointSlice to reference pod1 and thus we end up with a duplicate. If anything in the future triggers another syncService call that would be resolved. I'd initially thought that EndpointSliceTracker would be thing to trigger that follow up syncService, but I was wrong. That tracker just tracks the latest resource version for each EndpointSlice the controller has created/updated and compares that with EndpointSlices received through create/update events. If the controller doesn't think that EndpointSlice should exist or the controller's resource version for that slice does not match it triggers another syncService. In this case neither of those conditions would be met so no additional syncService call would be triggered. So that's a long way of saying that I think our best possible solution here is to introduce some delay between when Services can be synced. I think the easiest solution that would solve 99% of our problems would be to add something like a 1s delay in between syncs simply by delaying when was called. This would not be ideal, but it would result in an immediate improvement and be easy to backport. It's worth noting that although duplicate EndpointSlices like this or even attempts to update a stale EndpointSlice are not ideal, they are not actually invalid and will eventually resolve themselves. Consumers of this API already have to do some form of deduping. With that said, I'm hopeful that there's a clean way to implement a solution that actually waited for each EndpointSlice change the controller has made to be reflected in the cache because that would be a better solution. I'll spend some time looking into possible implementations here.\nI've created a draft PR that contains my initial progress towards the approach I described above. I don't particularly love this approach so I'm very open to other ways we could solve this issue. It's also still very much a WIP and has significant work remaining. I opened the PR early just so I could get some initial feedback on the direction and potentially inspire other ideas around approaches we could take here.\nis still the best approach I can come up with here. It is ready for review if anyone has time to take a look.\nQuick update here, I've limited the scope here to the following changes as discussed in Update the flaky test to consider duplicate EndpointSlices valid. Update the docs to make it very clear that consumers of the EndpointSlice API will have to do deduplication (as kube-proxy already does). Update the controller error handling to return immediately if a create, update, or delete errors (in that order). Reviews would be appreciated on all of the above.\nReopening since is not yet merged.\nThanks for catching that Closing now that is merged. /close\nClosing this issue. <detailsIn response to : Thanks for catching that Closing now that is merged. Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-6517593412935133397bc229bce4cc11c59f77134aa8a443efc63d74523f8d97","text":"<!-- Please only use this template for submitting reports about flaky tests or jobs (pass or fail with no underlying change in code) in Kubernetes CI --Which jobs are flaking: gce-master-scale-correctness Which test(s) are flaking: [sig-network] Ingress API should support creating Ingress API operations [Conformance] Testgrid link: https://k8s- Reason for failure: Anything else we need to know: /kind flake /sig scalability /sig network /cc\n/triage unresolved Comment when the issue is assessed and confirmed. I am a bot run by vllry. ‍\nthis is caused by the issue described in a different test is creating a default ingressclass, which breaks creation of ingress objects if there was already a default ingressclass set up in the cluster /assign\nthis should resolve the issue:\nI take it back, looks like the conformance test is also creating default ingressclass objects... it should not fix in"}
{"_id":"q-en-kubernetes-594015d051aa924e759c9f3f32851dfa4ee62c4bcf0b744cd10ad2563942c773","text":"The statement \"returns [...], or nil\" cannot be true since the return type is a struct: Suggest it could say \", or an empty struct if ...\" /sig api-machinery\nI can file a PR to fix the incorrect comment if you don't mind :)\nSure, I didn't get round to it yet.\n/assign"}
{"_id":"q-en-kubernetes-22019b12e21f03cfc3866a548a042db93f8788d8668821091518f2505b967935","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: not fixed the SMB mount issue totally, when there is a SMB mount with same remote path, original SMB mapping would be removed and SMB mapping re-established, that would cause temporally data inaccessible with the original container. should change to should use to check whether original smb mapping is valid, here is an example showing that after SMB server password changed, could validate whether connection is ok: the main diff on smb mount between linux and windows is that on LInux, every smb mount is independent, while on Windows, one remote smb server address(e.g. ) can only have one SMB-GlobalMapping. So on Windows, there is a difficulty: in some cases, we should skip create SMB-GlobalMapping since it already exists in other cases, original password is incorrect or expired, user wants to delete SMB-GlobalMapping, and then create a new SMB-GlobalMapping (fixed by ) On Windows, there is no way to check whether original SMB-GlobalMapping is valid or not, that was the answer about 1 years ago from windows team, will update this thread(also about the max SMB-GlobalMapping limit) if there is any new update recently. What you expected to happen: How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others: /sig storage /assign"}
{"_id":"q-en-kubernetes-741ba7915552c883482f87169b8e962f2e8ba01f4d776c4d410444365339445c","text":"What happened: I two CSI plugins to my Kubernetes cluster via a single call to . When I did this, the annotation was improperly set on 2 out of 3 nodes. Only one of the plugins appeared: What you expected to happen: Both and plugins should have appeared in the annotation. How to reproduce it (as minimally and precisely as possible): Apply two or more CSI plugins via a single call to . Anything else we need to know?: Kubelet logs indicate that it registered both plugins at the same time: The CSINode object looks complete: Only the node's annotation did not include Restarting kubelet fixed the issue. It looks like a race condition when InstallCSIDriver is called concurrently for multiple drivers. The code here: could lead to each thread observing the annotation as being empty, and then setting its own annotation that includes only itself. Environment: Kubernetes version (use ): v1.17.11 Cloud provider or hardware configuration: AWS OS (e.g: ): Ubuntu 20.04.1 LTS Kernel (e.g. ): Linux ip-172-31-17-163 5.4.0-1021-aws -Ubuntu SMP Fri Jul 24 09:42:29 UTC 2020 x8664 x8664 x86_64 GNU/Linux Install tools: juju\n/sig storage\n/assign\nI can reproduce this issue with commands below ! The correct csi nodeid would be like this"}
{"_id":"q-en-kubernetes-8d68651f6dddb98b5802137df4a357d3d5e45450aa6ccc1a8beea1f415e790bd","text":"What happened: I download fluentd-elasticsearch in addons. The following error occurs then I use \"kubectl apply -f es- \" Then, i find the elasticsearch-logging pod restart lots of time, and keep status with CrashLoopBackOff. [root ~]# kubectl logs -f elasticsearch-logging-0 -n kube-system export NODENAME=elasticsearch-logging-0 NODENAME=elasticsearch-logging-0 export NODEMASTER=true NODEMASTER=true export NODEDATA=true NODEDATA=true export HTTPPORT=9200 HTTPPORT=9200 export TRANSPORTPORT=9300 TRANSPORTPORT=9300 export MINIMUMMASTERNODES=2 MINIMUMMASTERNODES=2 chown -R elasticsearch:elasticsearch /data ./bin/elasticsearchloggingdiscovery I0902 14:25:06. 9 ] Kubernetes Elasticsearch logging discovery I0902 14:25:06. 9 ] Found [] I0902 14:25:16. 9 ] Found [] I0902 14:25:26. 9 ] Found [] I0902 14:25:36. 9 ] Found [] I0902 14:25:46. 9 ] Found [] I0902 14:25:56. 9 ] Found []\nunclear whether the addon is up to date and supported at this point. /sig instrumentation\ncc\n/assign I will take a look this week.\nI have the same problem k8s version:v1.18.8\nsame\nI've pushed a change to the stateful set that should unblock people as there was a 1.18 compatibility problem. however, please make sure you are also creating the service via , the discovery program finds the other elasticsearch nodes via service registration.\nplease feel free to re-open this issue if the latest version doesn't work for you.\nWhy was this PR to v1.20?\nprobably because that is the next release. most people who use this addon do so from master. Do you need this packaged in some other way?"}
{"_id":"q-en-kubernetes-b5dd420e9256a2076d24a57e38777d86254ed53c2469c911295cd2867dd2a197","text":"Is this a valid test? My understanding is we cannot rely on events for programming against. Which jobs are flaking: ci-cri-containerd-node-e2e-features ci-kubernetes-node-kubelet-features among others, mainly 'node-e2e-features' based tests. Which test(s) are flaking: Testgrid link: Reason for failure: expected total number of events was 4, actual events counted was 7 Anything else we need to know: It doesn't fail in a specific environment, should this test be running at all outside of this job? The comments inside the test aren't very helpful for understanding why we're checking what we're checking.\nI heard this is blocking you. /sig node /cc\nI see flakes going back at least a month In addition to an earlier issue clarifying the error message This might be a spurious match,\nNodeProblemDetector is broken on nearly every run:\nMaybe they should be marked [serial] or [disruptive]?\nWe can mark it and move it to the flake job, but I fear it will work fine over there.\nEdit in upstream issue says \"needs to be marked serial\" I'll make a PR"}
{"_id":"q-en-kubernetes-66e5bd92e1e67ce92c49e7e79cedef3be454c739166778adbb794720ca773090","text":"Which jobs are flaking: pull-kubernetes-bazel-test Which test(s) are flaking: TestDevicePluginReRegistrationProbeMode Testgrid link: Reason for failure: Anything else we need to know: Reproduce with /sig node\n/assign\nSee\nnot sure this is actually fixed, I see a failure from today"}
{"_id":"q-en-kubernetes-cecee0a132f84dbf6d5783d328e94b49c8666ebba8eef4ded1687e53b0451e97","text":"/kind failing-test <!-- If this is a flake, please add: /kind flake --/sig autoscaling Looks like it just started happening today (@ 2020-09-15 03:54:19 +0000 UTC) !\ncc the namespace from the provided object \"kube-system\" does not match the namespace \"horizontal-pod-autoscaling-830\". You must pass '--namespace=kube-system' to perform this operation. the namespace from the provided object \"custom-metrics\" does not match the namespace \"horizontal-pod-autoscaling-830\". You must pass '--namespace=custom-metrics' to perform this operation. the namespace from the provided object \"custom-metrics\" does not match the namespace \"horizontal-pod-autoscaling-830\". You must pass '--namespace=custom-metrics' to perform this operation.\nrelated fix:\nNo I don't think that's the fix for this (that is just a little cleanup). Actually I think we have to delete namespace parameter in line : Because before , the will have no actually effect. But now with the change of , it will have namespace : at the end of the command. But the namespace in yaml file is which is different from . We should not have parameter when applying from file because this test didn't have before, right? Related PR:\n/assign"}
{"_id":"q-en-kubernetes-d95498099d2ecd3a5959c7138a5f239c853bc0619206403c627c11bff3650513","text":"Which jobs are flaking: https://k8s- Which test(s) are flaking: Services should have session affinity timeout work for service with type clusterIP Services should have session affinity timeout work for NodePort service We have other flaky cases, like , that should also bu resolved. But cases also run in , so I suggest we solve this problem first. Testgrid link: https://k8s- Reason for failure: In ipvs mode, we set session affinity timeout as 125s, the test case is: run host command to send a request, mark returned hostName 125s another request, if return different hostName, test succed. I think the reason is, in the second request, ipvs connection may still in sessionAffinity, send another request will result in seffinity restart count, eventually result in this test case flaky.\n/sig network /area ipvs\n/triage unresolved Comment when the issue is assessed and confirmed. I am a bot run by vllry. ‍\n/assign /remove-triage unresolved\nThis keeps failing https://k8s-"}
{"_id":"q-en-kubernetes-7f02b998dfe3209626d954e195f8980700695bbf3d02ce8cc11b2df175854505","text":"Inspired from Our default Filter plugins have the following order: IIUC, this order was chosen to have the fastest Filters or the most probability of reporting Unschedulable first. Arguably, some are pretty much equivalent and could be reordered. The major problem with this order is that plugin is almost at the beginning. This plugin only reports . A Node that fails this plugin will be evaluated during preemption. But perhaps this Node doesn't satisfy other UnschedulabeAndUnresolvable requirements (taints, node selectors, nodeaffinity). Theoretically, having those Filters run first could improve preemption latency. The only open question is whether or not the ordering we have today is an API guarantee. I believe it's not (at least in component config), because the only way to change the ordering of Plugins is disabling all of them first. Decided order: /sig scheduling\nfor comment\nand :)\nEven we changed the order of plugins, the framework.RunFilterPlugins still exits early when returned by plugin. So should we change the current behavior to run the filter plugins until there is one plugin returned ?\nThat is correct, and that's what it should be.\nNo, that's what the original conversation suggested and it's likely not ideal in the average case.\n+1 to reordering. Taints, affinity, node name can all be executed before checking for resource fit. No, it is not.\n+1. Lightweight plugins that may return status can be moved ahead of other plugins that only return status.\nFYI: from who also had this concern: So how about , is it also a good candidate to be moved ahead of ? NodeAffinity and NodeSelector both incorporated in , I'm pretty sure NodeSelector logic is simply, but not 100% sure NodeAffinity logic falls into the category of \"lightweight\".\nLet me propose an order, for extra clarity:\ninternally composes and invokes , which IMO is a bit more expensive than . should be cheaper than . So my 2 cents is:\nBoth proposals are fine by me.\n/triage accepted /help /good-first-issue\n/assign\nI want to help here :)\n/remove-help"}
{"_id":"q-en-kubernetes-0c1300fddbd9f4a9f4b956efe1ec6f383c16bbfee6e62efdaecf062c42c5ed85","text":"What would you like to be : Scheduler VolumeBinding plugin reports error when a PVC doesn't exist for example: We need to take a second look at the VolumeBinding plugin and make sure that errors are returned only when there is a system error, not a user error. Why is this needed: To fix the scheduler's error reporting metrics. /sig scheduling /sig storage /cc\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign I'll take a look.\nIf the PVC object does not exist, the plugin will return status. I guess you mean the plugin will return an error when a referenced PV does not exist? I've created a PR for this:\nThanks! The errors I am seeing are on 1.18 clusters, can we look at 1.18 and try to patch? This is still throwing the scheduler's error metrics off.\nsure\nI've submitted a patch for the 1.18 release branch:\nI guess on the master branch should be reviewed and merged too? cc\nhello! the fix you made only apply for 1.18? how about 1.19 and 1.17? thanks!\nafter the fix on the master branch () is merged, I'll cherry-pick it into the 1.19 branch (the code are almost the same). I'm not sure if it's necessary to fix the 1.17 branch. It's 3 versions behind 1.20.\nhi, This is because of this patch: Do you think this fix needs to be cherry-picked into 1.19 and 1.18 branch too?\ncan you send a patch for 1.19?\nah, here it is\nto sum up: master: 1.19: 1.18: and I think this change should be backported to 1.19/1.18 too: I'll send PRs soon.\nsummary of the issue that an error will be reported instead of UnschedulableAndUnresolvable when PVC not found: master: already implemented, see 1.19: 1.18: can't backport as it does not implement the PreFilter extension point"}
{"_id":"q-en-kubernetes-d1edb6ddb7fb597cb6c072af22b1a8eaf52b05f34b8fb8536b7efd18524ab8a5","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: When I restart containerd.service, kubelet can not connect to it again and cause 100% cpu usage. there are many logs like this: I can show container info via crictl at the meantime and restart kubelet.service can resolve the problem. What you expected to happen: kubelet can reconnect to containerd after containerd restart. How to reproduce it (as minimally and precisely as possible): cpu usage and syslog. Anything else we need to know?: Environment: Kubernetes version (use ):Cloud provider or hardware configuration: armbian(debian on arm64) OS (e.g: ):Kernel (e.g. ):Install tools: kubeadm Network plugin and version (if this is a network-related bug): Others:\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig node\n/assign\nReproduce: can reproduce the issue by restarting when is the container runtime another setup where was used used as container runtime, didn't experience the issue by restarting , but restarted the underlying , same issue shown up again and had to restart to resolve the issue. Will continue with the investigation, might need to reconnect to after the socket is broken.\nTo me it seems there are a couple of problems here. 1. kubelet does not reconnect to the containerd socket. 2. The reconnect is agressive and consumes a lot if cpu. I don't know the cause for 1, but for 2 this might be an indication that the retry interval should be relaxed.\nSpent a few hours on the investigation, it's likely to be some internal states being messed up in .\nTL;DR: It should be related to gRPC. Hacked the vendored gRPC mod and changed its to something like this: Then CPU only spike for a few seconds after restart then comes back to normal instead of failing the all the time. kubelet passes a customized dialer to the gRPC client. I logged the address during kubelet startup: As you can see, the address of the is at . Before restarts, I can see from log that every time, the gRPC function is using the customized dialer that kubelet passed: However, once got restarted, the address of the dialer is changed: And that function at keeps failing. As I forced to use the original customized dialer, the issue is gone. Without too much knowledge on gRPC, my blind guess is as is passed by reference to gRPC client, some component could modify it during the process lifetime and the change got persisted. Anyway, I'll check with gRPC team on this issue and see if there's already a fix on their side. If so then I can port the fix here.\n/cc\nHi, are there any updates on this. I am experiencing what appears to be the same issue. If I restart containerd while the kubelet is running, my kubelet logs get flooded with: Restarting the kubelet appears to be the only way to re-establish the connection with containerd. Version information is:\nHi thanks for the info, I'm still working on the root cause. As I mentioned above, it's something related to gRPC. I collected the object info, this is the object before containerd restart: ! And this is the object after containerd restart: ! There's one thing made me concern is, the connection state is still 2 (i.e. , see ) after the failure. I'm not sure if this is related to the cause but anyway it still should be something related to gRPC. Btw I've followed their to use the latest gRPC client without customized dialer and still see this issue, will still work on the root cause.\nWe're seeing this now when we update our nodes to 1.19 where our Kata Containers install does a . Our current workaround is to also do , but would be good to get this fixed.\nIs this MR relative to the problem?\nNo I don't think so, I've cherry-picked that change and tried in my environment weeks back, and the issue is still there. I only get around the issue by modifying the gRPC code as I mentioned in the above, I'll continue to check which component modified the function in after thanksgiving.\nFor the sake of priority. This hit us hard these days too, as security updates for container.d triggered a restart and a ton of nodes flooded the central log management. ;) Thanks for working hard on a fix.\nSame here. Two systems were affected by automatic container.d upgrades & restarts. But now I wonder if - generally - \"unattended-upgrades\" (Ubuntu 18.04) is a wise thing to do? Or should the OS for kubernetes deployments always be updated manually & controlled? I can not find best practices for this in the documentation.\nAfter the recent CVE-2020-15257many users will upgrade and restart containerd. It would be a big problem.\nI see, will prioritize on the root cause and fix. At the mean time, you need to restart after you restart .\nHad another round of debugging, and made one step towards the root cause. I following snipple to the as I mentioned in previous : When I reproduced the error, I got following information for the dialer function: That leads to cadvisor, which is the only component that uses the in the repo: So I'm convinced that this should be related to and will update the thread once I've made further progress.\nEOD updates: issue is indeed caused by cadvisor, reproduced the issue in this . confusing timeout in log should be . root cause might be a bug for using context that already exceeded deadline in cadvisor. To validate that, I need to explore all methods of this : Will work on the fix on cadvisor for the rest of the week.\nFix was made in Copying from to ensure we can capture it on the issue here: I've cut new cAdvisor releases (v0.37.3 and v0.38.6) with this fix cherrypicked. We should get the cAdvisor fix back into kubernetes. We'll need 3 PRs to k/k with following updates: [x] k/k master to update cAdvisor to v0.38.6 -- [x] k/k release-1.20 to update cAdvisor to v0.38.6 -- [x] k/k release-1.19 to update cAdvisor to v0.37.3 --\nThanks for fixing this one, definitely a tricky one to track down.\nThanks for all the patience and team efforts on getting this issue fixed, also credits to on the guides and discussions, won't be able to get it done without your helps!\nThe fix was applied on master as well as cherrypicked into and releases.\nI remember you told me, but I forgot. Does issue exist in 1.18 and 1.17?\nI don't believe 1.18 or 1.17 should be affected. I think this issue was introduced with which commit was included in cAdvisor v0.37 release. cAdvisor v0.37 was vendored into k8s 1.19."}
{"_id":"q-en-kubernetes-2b1a8571774c2d7125d3f12097a1d441ecea49fa20b41cd9289c4da8d07b4033","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: Windows nodes running with high (90%) CPU usage are having calls to /stats/summary time out with \"Context Deadline Exceeded\". When this happens metrics-server cannot gather accurate CPU/memory usage for the node or pods running on the node which results in HPA not functioning returning <unknowninstead of values What you expected to happen: Windows nodes should be able to run with high CPU utilization without negatively impacting Kubernetes functionality How to reproduce it (as minimally and precisely as possible): a cluster with Windows Server 2019 nodes CPU intensive workloads into cluster ex: workload to put node into high CPU state node metrics with Anything else we need to know?: This issue appears to be easier to reproduce on 1.18.x nodes vs 1.16.x or 1.17.x nodes Ex: ! ! metrics-server logs show E1020 10:58:24. 1 ] unable to fully collect metrics: unable to fully scrape metrics from source kubeletsummary:akswin000000: unable to fetch metrics from Kubelet akswin000000 (10.240.0.115): Get https://10.240.0.115:/stats/summary?onlycpuandmemory=true: context deadline exceeded Kubelet logs show and Environment: Kubernetes version (use ): 1.18.8 Cloud provider or hardware configuration: AKS / Azure OS (e.g: ): WIndows Server 2019 Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others: Docker EE version 19.3.11\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig windows /area kubelet\nI was able to reproduce on 1.18.8 during a scale operation: where it fully subscribes all CPU's in Allocatable: Metric server logs: I do not see \"context deadline\" in kubelet but I did find some logs assocated: Once the pods are scheduled the node remains at 100% cpu and I no longer see errors but is not accurate: on the node (run after the pod deployment stabilized)\nAdding to kubelete's flag and I no longer see the errors. This reserves some cpu for kubelet and other processes. We call this out in the but do not do the same for cpu.\nAlso, without setting system reserved CPU, it looks like does not accurately reflect CPU usage on the nodes when CPU utilization is very high. When I got some nodes into this state I did not see any error messages in logs from metrics-server and also I noticed that that querying the metrics API directly from showed recent timestamps.\nI run into this again during scaling, collectiong logs for analysis here: Metric Server logs: Kubelet logs:\nhave you experienced this at all?\nI believe has seen high CPU issues before. Not particularly on this. I've heard there were kernel issues in the past that triggered high CPU cases.\nSome investigations on our end are showing that the DiagTrack service running in the infra/pause container can consume non-trivial amounts of CPU. I'll create an issue to address that problem as it might help out some here. That certainly will not eliminate the problem we are seeing tho.\nis an issue for DiagTrack\nThe differences in behavior between 1.17. and 1.18. clusters looks like it is caused by (which was NOT backported to any prior releases) K8s version <= 1.17 behavior Prior to these changes CPU weights for containers were inferred based on limit values set in the pod specs. The observed behavior on Windows system when CPU weights were used was that under high-CPU load pods/containers were not guaranteed their scheduling time. The good news here is that system critical services like kubelet/docker/etc were still able to get enough CPU time for the nodes to function properly. K8s version = 1.18 behavior After these changes CPU limits were enforced for pods. The observed behavior on 1.18 clusters is the when a Windows node enters a high-CPU state is that containers are guaranteed their scheduling time but host resources start to get starved. Usually this manifests as the node entering the state, metrics not working correctly (because calls to /stats/summary are timing out) or both. Suggestion / possible fixes Set the Windows process priority of the kubelet to  to help ensure it gets CPU time Kubelet can do this itself though calls (GetCurrentProcess() and Get/SetPriorityClass()) Set reserve CPU with with --system-reserves=[cpu=500m] This needs to scaled based on number of CPUs on host based on my experimentations Currently guidance on for Windows only mentions reserving system-memory It is still possible to over-commit CPU and starve kubelet/docker/HCS if users schedule pods with mis-matching requests and limits Update documentation.\nKubelet can do this itself though calls (GetCurrentProcess() and Get/SetPriorityClass()) I can work on this, I am wondering if we should do the same for CRI(like docker) and CNI processes.\nSure, if you want to work on this go ahead. I think we should probably try to make this be toggleable via a kubelet flag. I believe there are other windows specific flags already implemented in the kubelet (like one to run as a windows service IIRC)\ndocker/containerd should probably have a flag that allows them to run as high priory as well but would likely be different issues in the respective projects\n/reopen I don't think this was ever fully addressed\nReopened this issue. <detailsIn response to : /reopen I don't think this was ever fully addressed Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nalso related to\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\nThe Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle rotten [1]:\nThe Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Reopen this issue or PR with Mark this issue or PR as fresh with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /close [1]:\nClosing this issue. <detailsIn response to : The Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closedReopen this issue or PR with Mark this issue or PR as fresh with Offer to help out with [Issue Triage][1] Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-83e56b03cffa63a801b9c551dc6f83c72b33470abc0b17a71a3174855c0b306d","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: In k8s 1.19.2 If the seccomp annotations of psp doesn't contain , pod with seccomp annotations would not start with error: psp with works psp without doesn't work Is this due to ? What you expected to happen: In k8s 1.18, psp annotations without works. I think in k8s 1.19, it should also work.\n/sig auth\n/assign I think this was a regression with the seccomp graduation to GA. Since we auto-convert 'docker/default' to the runtime default in fields, but PSPs that only allow 'docker/default' aren't accepting 'runtime/default'. I think the easiest fix is to just add both and to the allowed profiles map whenever either is set (here: ) This will need to be cherry-picked /priority important-soon\nI'm happy to work on this one tomorrow if is not tackling the issue earlier :upsidedownface:\nFix is incoming in\nThanks for the quick fix ! Can you open a cherry-pick for v1.19 too? (see )\nForgot to link it, but this is already done: :)\nIs there an easy way to check in which version of k8s this fix is applied? Label, milestone, etc? Another question, what might be a workaround without updating cluster?\nFind the fix commit and run And there's also a cherry-pick commit: The workaround could be add annotations to your PSPs, take a look at this PR:"}
{"_id":"q-en-kubernetes-8b0b4f034e42c86a48caeab39efcee6b051c611ae2503e735f6aee9e4ff213df","text":"<!-- Please only use this template for submitting reports about continuously failing tests or jobs in Kubernetes CI --Which jobs are failing: - Also flaking on: Which test(s) are failing: Since when has it been failing: ~Since Oct 21? Testgrid link: CAPG conformance: CAPA conformance: gce cos master default: Reason for failure: Example spyglass link: Anything else we need to know: Triage shows that this is happening across several jobs (not just the ones I listed above) and the spike started happening earlier this week. /sig cli /cc\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nFixed by\nThis is affecting ppc64le jobs as well.\n/remove-kind failing-test /kind flake This is not a continuously failing test, it is instead flaking\n/priority critical-urgent the top flake for per /area conformance this is a conformance test, it definitely shouldn't be flaking, and should be demoted if it can't be fixed\n/reopen For CI signal to observe before marking as resolved\nReopened this issue. <detailsIn response to : /reopen For CI signal to observe before marking as resolved Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nLast failure 7 days ago 2020-11-03 05:40:32 +0000 UTC\nPlease review to see if we can close this now.\nLet's close this - also looks like it is no longer flaking as well"}
{"_id":"q-en-kubernetes-9f0489ba90a5e45d3a3b7e0a8ee80754276a572689d99ff3a3f98b8b892cc5b4","text":"Instead of relying on the clock: we should do something similar to the golang rate-limitter is doing:\n/sig api-machinery /priority important-longterm\n/help\nThis request has been marked as needing help from a contributor. Please ensure the request meets the requirements listed . If this request no longer meets these requirements, the label can be removed by commenting with the command. <detailsIn response to : /help Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nHi, I would like to help on this.\n/assign Thanks! /triage accepted\ncan you please provide more context or some guidance on how to proceed?\nSorry for delayed response. - what we should do is to basically: instead of having an additional go-routine that every 1s refreshes the budget change it roughly to: (a) keep the value of budget and the last time when we were refreshing it (b) on takeAvailable method: first recompute takeAvailable, which will be \"min(current budget + time elapsed since last refresh * refreshValue, maxAllowedBudget) update last refresh time to now return that value [We also would need to update the existing tests to match the new logic.] - would you be willing to work on it?\n-- I can take this up. One question I had was should the method rely on or the current time should be passed on by the caller like in the following: Having the caller pass the time completely removes the dependency on the clock at least for the implementation.\n/assign\nLet's may be start with: passing on budget creation to it relying on\n-- thank you for the thoughts. I drafted a rough implementation in which still depends on . I will make the change as you suggested and rely solely on a instance passed on initialization. I have to rethink the testing strategy a bit, possibly using a .\nYes - FakeClock is the way to test this.\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle stale\nStale issues rot after 30d of inactivity. Mark the issue as fresh with . Rotten issues close after an additional 30d of inactivity. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle rotten\nRotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . Send feedback to sig-contributor-experience at . /close\nClosing this issue. <detailsIn response to : Rotten issues close after 30d of inactivity. Reopen the issue with . Mark the issue as fresh with . /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nReopening - this didn't happen."}
{"_id":"q-en-kubernetes-3fd3f003e8b9c95013faf5e9493422dc066d2c12c2867630994871159ed7ad67","text":"/sig storage What happened: When you interrupt the e2e storage stress test, namespaced objects are cleaned up, but non-namespaced objects like StorageClasses are not cleaned up. What you expected to happen: Non-namespaced objects like StorageClasses are cleaned up so long as the user doesn't interrupt a second time. How to reproduce it (as minimally and precisely as possible):\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-c6227add66bc5f03787cb3392aefa0f44a98a0ee8635bd04e13da5a47ee9df31","text":"What happened: ExecUtil in has mandatory logging side effect hardcoded to INFO What you expected to happen: Since INFO level logging may or may not be an implementation detail of a test that is worth showing all the time, it would be better if we can turn this logging off optionally with a Quiet (false by default) boolean field How to reproduce it (as minimally and precisely as possible): Run any test that execs across pods, i.e. exec tests\nThere are no sig labels on this issue. Please add an appropriate label by using one of the following commands: - Please see the for a listing of the SIGs, working groups, and committees available. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-4a4ec1a49fbabe86c6d090e1ab7913dc9461e7b5190552e5bbe6102febf15f15","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: With a /22 pod network and a /24 assigned to each node, a third node cannot get a pod CIDR if kube-controller-manager has ever been restarted. What you expected to happen: Once a third or fourth node joins, it gets a /24. How to reproduce it (as minimally and precisely as possible): 150% usage on the IPAM controller is not a good sign. Anything else we need to know?: During the creation of the rangeAllocator, Occupy() is called twice on the pod networks. The first time is before subscribing to the node informer. The second time is during the node informer callbacks because on subscribing it calls Add for all existing nodes. I will submit a proposed fix shortly. Environment: Kubernetes version (use ): 1.19.2\n/sig network\n/triage accepted thanks for finding the bug I 've introduced :)"}
{"_id":"q-en-kubernetes-6340c3ef6ea462c626be4450573121f0269c27f7fc373237e8e1fa675a8ec0f3","text":"What happened: Test create method for non-namespaced subresources generate invalid code. I have the following code generation annotations: which generated test method with generated fake client: What you expected to happen: It should be: How to reproduce it (as minimally and precisely as possible): Generate clients for non-namespaced resources with subresources (like scale) like this: /sig api-machinery /area code-generator /assign\nThe label(s) cannot be applied, because the repository doesn't have them <detailsIn response to : What happened: What you expected to happen: It should be: How to reproduce it (as minimally and precisely as possible): Generate clients for non-namespaced resources with subresources (like scale) like this: /sig api-machinery /area code-generator /assign Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign\n/triage accepted"}
{"_id":"q-en-kubernetes-084b04dfbf25870909925faa5be0f6f44aa534bd7c78956a8d5debd4f56b4523","text":"When configuring new behaviors section in HPA I two scenarios that I think are incorrect. First one: stabilizationWindowSeconds set up to 5 minutes for both scale up and scale down min replica count 2 Application consumes big amount of resources during warm-up 0 minute: current replica count: 2 1 minute: desired replica count: 3 - no scale up stabilization window worked 2 minute: desired replica count: 1 (!) - scale up to 3 (max number from previous recommendations during stabilization window period) ! Second one Stabilization window for scale up 0 Stabilization window for scale down 5 minutes min replica count 2 behavior set to scale up new pod in 2 minutes scaleUp: policies: type: Pods value: 1 periodSeconds: 120 What happens: 0 minute: 2 replicas 1 minute : performance spike scale up to 3 (desired value 4) 2 minute : no spike no scale up 3 minute: no spike scale up to 4(!) - supposedly desired value 1 and supposedly worked this way because of scale down window set to 5 minutes picking max value from that 5 minutes window that was 4 replicas Expected behaviors 1 Scenario no scale up for 5 minutes 2 Scenario correct measurement and no scale up when there is no need for it Environment: Kubernetes version (use ): client 1.19.3 Server 1.19.0 Cloud provider or hardware configuration: Azure aks\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig autoscaling\n/assign josephburnett\nThe HPA controller keeps a flat history of recommendations for stabilization. ! When downscale stabilizing the maximum recommendation is selected. Purple is pre-stabilized recommendation. Green is final recommendation. Blue is max recommendation within window. ! When upscale stabilization the minimum recommendation is selected. ! However when both up and down scale stabilization is configured, the interpretation of the history changes depending on the direction of movement. Red is the undesired, actuated recommendation. ! What we want is to keep the stabilized recommendation within the envelope of the minimum and maximum over configured stabilization windows. We should only move when the envelope forces a move. ! We first stabilize up and down, and check if the current value fits within it. If it does (or the range is \"inverted\"), we do nothing. If it does not, we emit the \"nearest possible\" value, which will be lower or upper bound, depending on where the current size is in relation to the range. (thanks and for the paragraph above)"}
{"_id":"q-en-kubernetes-80b85acd6231ee235cafa6f27894587bc197e63dfd9d9867bbb28bcc18cf7d30","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: Tried to create an ingress with a backend serviceport pointing to a service with a port name with length 15 chars that is valid from the service point of view. But failed because the validation for Ingress does not match the one for service. Service Validation (uses DNS Label): ) Ingress Service Port Validation (uses IANA format like for container ports): What you expected to happen: Service port validation should match Ingress ServicePort validation How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): Client Version: version.Info{Major:\"1\", Minor:\"19\", GitVersion:\"v1.19.4\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2020-11-16T17:18:52Z\", GoVersion:\"go1.15.4\", Compiler:\"gc\", Platform:\"linux/amd64\"} Server Version: version.Info{Major:\"1\", Minor:\"18\", GitVersion:\"v1.18.10\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2020-10-15T01:43:56Z\", GoVersion:\"go1.13.15\", Compiler:\"gc\", Platform:\"linux/amd64\"} Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others:\n/sig network\nIt's not friendly to users who create service with long service port name, to create ingress with port reference. Although, the port name of service is editable. Maybe, we can add warning message for service creation and update when the port name is not a valid ingress port name.\nI'm fine with a doc change mentioning the restriction similar to the restriction in the doc for container port () where it explicitly says about the IANASVCNAME constraint\n/triage accepted /assign"}
{"_id":"q-en-kubernetes-21bf7b43af757a45e2c2cf965564fa22ba871a5ed69674dbd9c8734885a2946e","text":"What happened: The latest release of the debian-base image, v2.1.3, contains a vulnerable version of the apt package. CVE-2020- impacts apt version 1.8.2.1. What you expected to happen: Image should be updated to use at least apt 1.8.2.2 to clear the vulnerability. How to reproduce it (as minimally and precisely as possible):\n/sig release\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nkube-proxy is using debain as base image. /assign Let me check other components\nchecked the base image in and I seems that no new debian or ubuntu official image is built for this CVE ?\nThat is correct, the purpose of this issue is to get a new image built and released, one which contains a resolution for the CVE.\nper comments of @ tianon in Will update it once debain releases it."}
{"_id":"q-en-kubernetes-0f88a3153dda567e9533f504c976fd64b77ab7325499fe31f150f7406ad1ec6b","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: ListByResourceGroup method in the azure diskClient mock makes a call to Delete What you expected to happen: ListByResourceGroup method in the azure diskClient mock should call ListByResourceGroup How to reproduce it (as minimally and precisely as possible): Anything else we need to know?: Environment: Kubernetes version (use ): Cloud provider or hardware configuration: azure OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others:\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign\n/sig cloud-provider-azure\nThe label(s) cannot be applied, because the repository doesn't have them <detailsIn response to : /sig cloud-provider-azure Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig storage\n/area provider/azure\nfixed by /close\nClosing this issue. <detailsIn response to : fixed by /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-a447d5b22258c459bf0b5135f1bf8ec8c889c8a7717de7adc483e574a01e514d","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: Cluster was upgraded from 1.19.5 to 1.20.1 using kubeadm. After that, kube-apiserver cpu usage went from 50% to 300%. What you expected to happen: Expected that apiserver has the same cpu usage. How to reproduce it (as minimally and precisely as possible): Update a cluster from 1.19.5 to 1.20.1. Anything else we need to know?: The cluster is running on sbc hardware, thats why the original 50% cpu usage, and thats why 300% cpu usage is noticeable and causes service disruption. Environment: Kubernetes version (use ): (the apiserver is rolled back to 1.19.5)Cloud provider or hardware configuration: bare metal on SBC hardware. OS (e.g: ):Kernel (e.g. ):Install tools: kubeadm Network plugin and version (if this is a network-related bug): kube-router Others: Please check profiling of each apiserver.\n/sig api-machinery\nplease do a and show the contents here. (hide sensitive IPs if needed). this is the static pod for the kube-apiserver. also, anything interesting in the logs of the api server?\nOn another cluster I could not reproduce the issue. But, on the original one, it happens again if i switch to the newer apiserver image. So, now I am bisecting apiserver to find the commit since it behaves this way, hopefully that will help. I could not spot any interesting in the logs. I will be able to provide more after the bisecting procedure.\npost the above output whenever possible.\nThe static pod manifest has been moved to avoid kubelet starting it: Still bisecting, trying to find the cause. May it relate to the go1.15 upgrade? Building the apiserver with go1.15.6 seems to be working, while building with 1.15.5 results in a wrong binary. But, still I am working to confirm this.\nWith apiserver built with go1.15.5 I get 300% cpu usage most of the time, and the following log entries appear:\nunclear whether we have perf tests to catch similar problems: /sig scalability for other readers, go1.15.5 is the version used to build 1.20: /sig auth this and similar come when doing rest storage on various apis. e.g.\n/retitle performance regression: 1.20 apiserver has high cpu usage\n/kind regression\nDefinitely it seems that the binary built with go1.15.5 is broken. It ends up on 300% cpu usage and with log messages posted above. Note that I am building apiserver this way: Building with go1.15.6 is better, even if not all of the time. Perhaps I am on the wrong way?\nthis is a detail i've missed from the OP. since we don't have k8s ARM tests, it is likely why we are not catching this and it could be a golang compiler issue on ARM only. in the past we have seen some weird issues from ARM users, that have hinted about compiler bugs.\nSorry for that, I did not mention that explicitly. If it turns out to be a compiler bug, will it trigger a golang upgrade for k8s images?\nI will make a build with golang 1.15.2 too and post the results later.\nif we know that a certain golang version is buggy, we can update a specific k8s version to use a golang version that works. in this case the latest 1.20.x k8s can be updated to not use 1.15.5.\nSimply compiling 1.20.1 with golang 1.15.2 still produces misbehaving binaries. So, I will do a bisecting again with building binaries as documented, and try to find the bad commit this way. This will take longer on my hardware.\nSo, apiserver behavior is a bit random regarding this issue. What is true that on my hardware (orange pi pc) the apiserver needs around 85 cpu seconds to boot up, after that it consumes ~10% cpu time. When a build is bad, apiserver passes the 85 cpu seconds and still uses more cpu than 10%, sometimes even 300%. So, based on this, bisecting shows is first bad commit.\nReverting that commit on 1.20.1 makes apiserver behaving as before. Consistently it finishes its bootup in 90 seconds, and then cpu usage gets to its normal 10-50%, it depends on load-balancer.\nBy the way, the problem is not just high cpu usage, but during bootup and high cpu usage, apiserver stays in not ready state, thus kubelet restarts it after failed liveness probes. So it may cause service outage.\nwe had long discussions in the past where RPI users complained about the kubeadm probe defaults. we claimed something else is not right if the probes are failing, so i think you are confirming that. i can ping the release team and api-machinery about this ticket, but note post the holidays responses will be slow.\n/retitle performance regression on ARM: 1.20 apiserver has high cpu usage\nI suspect, that the mentioned commit fixes things in a way that it actually places more cpu load, I mean expression evaluates to true more often. The attached profiling data may be a proof for this.\nIt seems that on slower hardware, operations are distributed into more sample periods (time.Millisecond), which makes more calls to the expensive metrics.Observe functions. Before the mentioned fix, due to quantization bug, those calls were not made every time.Milliseconds. Now they are, and profiling data proves this. Can we raise the samplePeriod, can it be made configurable?\nsig api machinery can provide feedback on this.\nis that accurate? my reading of the fix is that the quantize method returns results with the same granularity. The bug that was fixed was that quantization could return numbers that resulted in negative calculations if wall clock time changed, and now it respects the monotonic clock.\nPerhaps not. That was just my assumption. But still, profiling data shows (at least for me) that metric.Observe calls take most of the time, while reverting the mentioned commit solves my problem. Somehow less calls are made to them.\n/assign\nWe have numerous K3s users reporting perf issues (and crashes in the cronjob controller) on armv7l. I can reproduce it locally on my Pi nodes as well. I hadn't gotten around to digging into it yet, so I'm glad someone else has done so. We also had issues getting arm to build properly starting with Kubernetes 1.19.4; some builds would crash with what looked like a corrupted stack on armv7l unless we set for the build. See: Prior to doing that it was fairly random if a build would work or not; changing unrelated bits of code (ie bumping a vendored module version) would give us a clean build, and then changing something else would break it, and then changing a 3rd thing would give us a clean build again. It felt very much like golang arm toolchain issues. This sounds similar to what reported with good vs bad builds.\nI've a little debugging: And I got strange results, a snippet from stdout: Check out the high on the 4th line. That causes a long-long loop to start. And there are more of such lines. If I refactor the code to remove the inline func() and do the Unlock() and the end of the block, this behavior goes away, apiserver goes back to normal. May it be a compiler bug?\nearlier you've mentioned: Building the apiserver with go1.15.6 seems to be working, while building with 1.15.5 results in a wrong binary But, still I am working to confirm this. did you confirm that the newer toolchain does not produce code that exhibits the problem? seems likely. in the past i've seen at least one compiler bug-like report from a k8s user (on ARM)\nUnfortunately, I think I got confused regarding compilers. Now it seems that I got the same results with go1.15.2, 1.15.5, 1.15.6. So, now I think it does not matter which toolchain I use.\nI've applied the following patch on release-1.20: And this produces strange results: It can be observed that from the third line, the function receives strange values. Later normal values appear, but sometimes the calculated is a large positive integer, which triggers a long running loop. Unfortunately, I am not familiar with golang's internals, and could not reproduce this in a few lines of code. But indeed, it seems that something weird is happening. If I place a before the call to , then the function gets the correct value passed all the time. How to proceed?\nI think this is an arm32 int64 division overflow/underflow issue? I'm currently testing a minor change to confirm/deny this hunch, but building for arm takes forever so it's still running.\nIt looks like the time value is being corrupted when being passed into quantize somehow, would be nice to try and reproduce with a very small executable. k3s has a long history of arm32 issues like this that I suspect is a core golang issue, but hard to reproduce and isolate, arbitrary code changes could change the result.\nyou could try running something like this, which exercises the local logic of the old and new quantize methods and checks that they output roughly similar values (some drift is expected because one uses wall clock and the other uses monotonic clock) I was able to run this for several minutes on linux/amd64 with no problematic values output\nYes, it looks like parameter corruption rather than a division problem. very strange.\nYeah my simple change to the quantize didn't make any difference. That combined with the strange output from the debug printing does indicate some sort of corruption to the when time struct that's being passed in, rather than bad math on the parameter.\nPresumably all the unit tests are passing on the arch & OS you are using.\nBut there are no unit tests focused on that. It's on the to-do list.\nIf we created unit tests for this file, would somebody be running them on the arch & OS involved here?\nI put in a check in quantize to grab a stack trace if when is before t0:\nThis seems like papering over a compiler problem, but this fixes it for me:\n: that is interesting info. I would prefer to not merge that change, we have another mystery where the monotonic time seems to go backwards, and I would hate to lose half the debug info from that. I will add unit tests for the sample-and-watermark histograms; can somebody please run them on the arch & OS involved here?\nThe time going backwards stuff all seems to be related to VM time sync, isn't it? In this case the compiler seems to be doing something that's corrupting the time struct out from under the anonymous function; the fact that it appears to be going backwards is not (as far as I can tell) due to host time actually jumping backwards into the dark ages with a weird TZ offset. FWIW I considered moving the Before check-and-err into the quantize function, but it wasn't necessary for a minimal fix for this issue don't I left it alone. I have a couple armhf (Pi3b/Pi4b) systems I can run tests on, assuming it's not too difficult to set up.\nI am not sure whether the monotonic time going backwards is the same bug. It was observed on Intel arch. I suspect another way to paper over the compiler bug would be to switch the argument to from pass-by-value to pass-by-reference. I suspect a compiler bug would be a valuable thing to find. I am working on unit tests for this package. They will involve dependency on the Prometheus go client library, but nothing more. Will that be simple enough?\nThe incidents I have personally seen or helped users with time sync issues have all been VMs on Intel. If there are non-virtualized platforms with this issue it'd be interesting to see. Simply preventing the sampling time bucket from going backwards seems reasonable, which appears to be what the fix was attempting to do. Changing it to a pointer is an interesting idea. I can give it a try later this evening. I was figuring I'd have to run the entire Kubernetes test suite, so anything less than that is great!\nYour small program is running on arm for minutes now without any errors. It needs more to trigger the compiler bug.\nI will be able, though I will need instructions how to do that.\nHere is a typescript of me running the new unit test: (the test does exercise the case of the clock going backwards a little) That's using the branch in\nOn ARM, I get similar results:\n/remove-sig auth\n: dang. Can you try it a few times? If that does not provoke the problem, perhaps you can edit the test to do a lot more iterations? You can also cut down on the warning noise by making a much smaller fraction of the cases have time going backwards.\nRunning multiple times results in some failing. Sometimes there are cases when 3 consecutive runs are successful. So basically, the test triggers the issue, but not deterministically. Even with iterations set to 1000, there are still successful test runs.\nIs that close enough to submit as a compiler bug report? BTW, what do the failures look like?\nis the logic non-flaky on AMD64? for a compiler bug, the maintainers are likely to request the failing code disassembled, so that they don't have to clone k8s, or a minimal code example that fails.\nI am getting non-flaky results on this: I think I have gotten the test to be about as simple as makes sense for a kube unit test. I will have to leave the compiler bug hunting to people more familiar with ARM.\nI tried to make a stdlib-only file that exercises the same call patterns w.r.t. variable scope as the original at what does that do when run on ARM?\nInteresting, running the tests by on an rpi4 (using arm32) I am able to trigger two different issues: clock is NOT monotonic and can go backwards time IS corrupted periodically when being passed into I am also not able to trigger any issues running those tests on Darwin (using x86_64). However, running the stand-alone go file by I am not able to trigger any problems on the rpi4\n: sorry if it wasn't clear, but the test introduced by uses a fake clock that is set backwards in some cases just to test how the metrics code reacts to that.\nSame here, on an OrangePI armv7 board, the single stdlib-only test does not produce any errors.\nLater I will do a test on arm64 as well.\nwoops, ah thank you I had to add to use the flag with go test to see those messages in Darwin and missed that part of the test code. At least the arm32 corruption issue is reproducible with that unit test, but I suspect that attempts to reduce the code space for the problem may cause it to go away, so maybe an issue of memory management or executable code layout. Unfortunately seems like debugging tools can be hard to come by on arm32 (eg no delve), am open to any ideas or suggestions on how to dig into it more\nBTW, I updated to always log an info message explaining what's up with the warnings.\ni don't know how to handle go mod imports with and some of the golang intrinsics are weird, but you could instead: build the unit test as a binary with the binary and for the object code of interest log a ticket @ golang and share the dissasembly you might also be able to use gdb over the test binary - set some breakpoints and try to catch what instruction is stomping the register / stack address that holds the value that ends up being bogus:\nThanks a bunch Was just looking into the compiled test, that is a great guide & will submit an issue with golang.\nOn arm64 the tests succeeded multiple times, no failures.\nCompiled the test code for i386 resulted in: On i386 the test code triggers the bug too.\nDifferent GOGC settings results in different test failure rate:\nI am able to reproduce that. ... and if I build without the variant GOARCH then the test passes all 300 times.\nWhen I merge and , the problem is visibly gone on the 386 arch on linux.\nshould we merge the pass-by-reference bit of in the meantime until golang/go is resolved?\nI would say that we could wait a few days for response on the bug report. On the other side, I dont know how fast a resolution will get into a stable golang release, and what is needed for k8s to be built using that. Can these take a long time?\nyes, I would not expect it faster than O(weeks), potentially O(months) if the golang investigation takes a while\nIt would be great if the fix could be merged and backported in time for 1.20.2; as it currently stands 1.20 is basically unusable on arm.\nBTW, contains only the pass-by-reference change and logging of an error if there is a wild jump anyway; I would recommend taking all of that PR.\nthe deadline for cherry-picks to be merged for 1.20.2 is tomorrow\ni wonder if it's also related w/ , the latter issue seems to be stambling on the memory pressure tho. but cpu&memory consumption can be also connected.\nI think yes, I've seen similar log messages.\nFor what it is worth this seems like a very high impact bug, as it can potentially effect 64-bit architectures also. The following example of the bug that was submitted as part of the fix also fails on Darwin/x86_64: It seems like the best mitigation until a new version of golang is released is to disable optimizations for the compiled binaries.\nI created as a cut-down alternative to , containing only the work-around and not the checking and logging for the argument passing problem.\n: What do you recommend for how to proceed on this? We should probably get something into 1.20.2\nagree. I'm fine with a temporary commented pass-by-reference workaround.\n/triage accepted\nping , On arm64 I got the same error like . and kube-api-server restart many times.\nthe error you are getting looks like a connection problem, so not clear if related. the bug discussed here is in golang. so if we update to the latest PATCH version golang in our k8s PATCH versions it will be resolved.\nthanks. but In my 3 masters kubernetse clusters. Only one master got this Error . others works well. this confused me. this problem like connection working with ipv6 not ipv4?\nyou can try asking in the #sig-network channel on slack or log a new issue."}
{"_id":"q-en-kubernetes-c1d13bb18e183f878ed8214b5be51aa974fcabd28c6374bb8a8397af786b67ff","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --What happened: Creating a pod with a nested subpath volume will cause the pod to fail to start on Windows containerd nodes. What you expected to happen: The pods should start normally. How to reproduce it (as minimally and precisely as possible): Create the following pod on a Windows node with containerd: apiVersion: v1 kind: Pod metadata: annotations: e2e-test-privileged-psp name: pod-subpath-test-downwardapi spec: containers: args: mounttest --filecontentinloop=/test-volume --retrytime=20 image: claudiubelu/agnhost:2.21 name: test-container-subpath-downwardapi volumeMounts: mountPath: /test-volume name: test-volume subPath: downward/podname volumes: downwardAPI: defaultMode: 420 items: fieldRef: apiVersion: v1 fieldPath: path: downward/podname name: test-volume Anything else we need to know?: This issue is causing the following tests to fail on Windows Containerd nodes: [sig-storage] Subpath Atomic writer volumes should support subpaths with downward pod [LinuxOnly] [Conformance] [sig-storage] Subpath Atomic writer volumes should support subpaths with projected pod [LinuxOnly] [Conformance] The tests have the ` tag because we did not support Windows Containerd nodes at the time, and the tests still do not work for Windows Docker nodes. After a fix for this bug is merged, these tests will have to be renamed to enable them for Windows Containerd environments but not for Windows Docker environments. /sig windows Environment: Kubernetes version (use ): v1.21.0-alpha.0.493+ Cloud provider or hardware configuration: N/A OS (e.g: ): Windows Server 1809, Containerd Kernel (e.g. ): 10.0. N/A Build Install tools: N/A Network plugin and version (if this is a network-related bug): N/A Others:\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle stale\n/remove-lifecycle stale\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle stale\n/remove-lifecycle stale"}
{"_id":"q-en-kubernetes-392791a00eb0d458d65d485406650be237b4fd573add6c43f25d80455a2233cb","text":"<!-- Please only use this template for submitting enhancement requests --What would you like to be : Remove the folder in favor of . Why is this needed: With the refactoring from the Netpol validation , the tests from folder are deprecated and duplicated at this point. Action items: tests tests\n/area testing /sig network\nThe label(s) cannot be applied, because the repository doesn't have them <detailsIn response to : /area testing /sig network Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nthank you ... !\n/triage accepted"}
{"_id":"q-en-kubernetes-254a7f50124a0fe157119f98cc1640ad309a623d1243da6408a85398946022f6","text":"Many e2e tests today have a tag in their name, which lets us know which sig is responsible for fixing failing or flaky e2e tests. However, not all tests have such a tag. We should: assign tags to all e2e tests that don't yet have them enforce that no new e2e tests can be without a tag in their name EDIT: proposed path forward (pulled out of and ) rearrange e2e tests into sig-owned packages, - e2e tests that run as part of nodee2e - e2e tests that run as part of e2e and nodee2e add that imports sub-packages so e2e binaries can continue to import to pick up test definitions () ensure each sig-owned package has: a package-local () an OWNERS files with sig members and ( follow these conventions: top-level calls are required, nested calls are not allowed remove usage of , , , and in favor of nested calls are ok, nested calls are not enforce these conventions via some combination of: static analysis regex enforcement against names output during a run while we are here, is vestigial and can be removed (ref: ()) EDIT: Everything now conforms to policies defined above will close this out by adding a blocking test and docs to /sig testing /area test for enforcement (FYI /sig architecture /area conformance some of these are conformance tests (FYI /wg reliability for helping determine appropriate sig ownership (FYI ) /sig release since this is relevant to figuring out who should fix release-blocking tests (FYI /priority important-soon /milestone v1.21 I think it's reasonable to aim to accomplish this prior to code freeze for v1.21\nThe label(s) cannot be applied, because the repository doesn't have them <detailsIn response to : Many e2e tests today have a tag in their name, which lets us know which sig is responsible for fixing failing or flaky e2e tests. assign tags to all e2e tests that don't yet have them enforce that no new e2e tests can be without a tag in their name /sig testing /area test for enforcement (FYI /sig architecture /area conformance some of these are conformance tests (FYI /wg reliability for helping determine appropriate sig ownership (FYI ) /sig release since this is relevant to figuring out who should fix release-blocking tests (FYI /priority important-soon /milestone v1.21 I think it's reasonable to aim to accomplish this prior to code freeze for v1.21 Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign Assigning an arbitrary lead from each SIG/WG mentioned, please assign to someone more appropriate if you don't have the bandwidth\nI can't see anything sig-network related :thinking: there are a lot of NodeConformance things though ,\n/assign for visibility - lots of these are node /sig node\n/assign (for sig-node)\nI've just sent out for autoscaling tests. It seems that most of the remaining tests are node-related. We should try to fix most of these to see if there is anything that is hidden between those.\n/unassign\nSent to address rest of the tests. Echoing the review comment from I agree, we can remove prefix once all the tests are tagged. wdyt ? I'll raise a follow up to clean the prefix.\nre: /assign (also maybe knows) at one time I thought Openshift was relying on to filter out upstream tests for certain jobs? I don't see that it's consistently applied some I'm assuming we could drop from our tests? (I said as much here: )\nWe synthetically apply that (if the source package is in ) so can be dropped. Also note in generally I am massively in favor of this change - I forced this to happen in openshift so some of the unassigned upstream tests got blanket assigned to equivalent teams (i.e. any test in that didn't have a sig had to get one synthetically ) so this will remove the need for it. Having owners just clarified everything, and then the ratcheting \"you may not add one without\" also ensured folks thought even a tiny bit about ownership.\nFleshing out what I suggested in and Can we agree to do the following? rearrange e2e tests into sig-owned packages, - e2e tests that run as part of nodee2e - e2e tests that run as part of e2e and node_e2e ensure each sig-owned package has: a package-local () an OWNERS files with sig members and ( remove usage of , , , and in favor of add that imports sub-packages so e2e binaries can continue to import to pick up test definitions () Enforcing test ownership then looks like: an e2e test must live in a sig-owned package (or subpackage) an e2e test must be defined by that package's PRs to that package will be assigned / auto-labeled to the appropriate SIG changing test ownership is a move from one package to another, requires approval from both the old SIG and new SIG Implications I don't have the foresight to see as blockers, can you think of any? we disallow the concept of e2e packages containing tests owned by multiple sigs some moves may require imports to change or things like helper functions and fixtures to get extracted Folks seemed to like the idea at the time but I'm seeing PRs move in different directions. How can I help get us on the same page?\n+1 for that I think that the PR above as an example is an intermediate step (it's not a different direction, rather just one step instead of multiple at once).\nThere is one question that I have at this point - what do we do with nested Described, e.g.: this one: [there is KubeDescribe nested in SIGDescribe]\nI think that having SIGDescribe twice in those cases doesn't seem to be the best option, and those cases seem to be significant percentage of remaining cases...\nI don't have strong opinion here. Mostly curious why prefer the non-explicit variant? Is it to optimize for tests migrations?\nThere is one question that I have at this point - what do we do with nested Described I can't think of a strong reason to ban them, and I think it would be more trouble than it's worth to come up with alternatives for how they're used now. I can't think of a compile-time way to enforce that \"[sig-foo]\" doesn't show up in their text, so I'm not sure it's worth keeping/enforcing use of wrappers (they definitely make error reporting worse) So something like: replace nested calls with follow these conventions: top-level calls are required, nested calls are not allowed nested calls are ok, nested calls are not enforcement of these conventions can be done via some combination of static analysis and regex check against the full names generated by a run\nThat makes sense to me. I will start looking into migrating existing tests to satisfy these rules in free cycle, but i'm leaving the enforcement to you :)\nI do consider not having to modify test code to change ownership a benefit. But it's more that we'd have to change a lot more code to conform to than . Since we mostly already have the pattern in place, I'd rather stick to it.\nThinking of timeline (ref: ) This work can continue past Code Freeze on Mar 9, but that's a useful milestone to evaluate our progress I'll update the issue description as we get tests owned My goal is to get enforcement landed sometime after Code Freeze This work can't continue past Test Freeze on Mar 24\n- also one more thing to do would be to ensure that no tests are tagged with more then one SIG (which I think isn't true now for a few tests)\nalso one more thing to do would be to ensure that no tests are tagged with more then one SIG (which I think isn't true now for a few tests) Good point. There are tinier corner cases, but let's start with the big one. /sig storage /sig windows /assign There are a bunch of storage tests that look like the following: Which SIG should own these?\nWe make us of for many e2e tests that are intended to run on windows machines. My suggestion here is probably to have sig-storage own these tests as they reside under /test/e2e/storage/* and switch to using instead of in the test names. SIG-windows will always be happy to help debug issues here. - thought?\nI think it makes sense for storage to own the windows storage tests. Using [Feature:Windows] seems reasonable although we'll have to update a number of test jobs.\nWorking on in-tree enforcement before I update description. tl;dr the majority of tests listed in the description are still not tagged (with at least one SIG), waiting on SIG Node to sign off on usage of KubeDescribe has gone way down, but not at zero yet\nis the biggest bit now (once I get feedback from SIG node, we will be almost there with both)\nrefreshing description based on script in\nRegarding \"multi-ownership\", with windows test figured (the PR didn't merge yet, but hopefully will soon), the biggest question is now about the upgrade tests. - there are a number of apps related upgrade tests. Even though they are currently running on GCP, I suggest that we officially make them sig-apps tests (I'm sure sig-cloud-provider-gcp will be happy to help on infrastructural problems). Thoughts?\ncan you be more specific about which particular tests? I quickly checked and those are already marked as sig-apps. I don't have any objections as long as they are part of sig-apps territory but I'd like to know a concrete list before saying so :wink:\n- sorry, I should have been more specific. Many of those tests are actually being run from here: and the tests there are marked as [sig-cloud-provider-gcp] As part of ensuring that each test has a one single sig ownership, I suggest making it explicit that all of those are sig-apps. Does that make more sense now?\nmost - yes, but I see a few exceptions: upgradeTest: ServiceUpgradeTest - seems like sig-network SecretUpgradeTest - looks like assigned to sig-storage/sig-apimachinery ConfigMapUpgradeTest - looks like assigned to sig-storage/sig-apimachinery HPAUpgradeTest - sig-autoscaling PersistentVolumeUpgradeTest - sig-storage AppArmorUpgradeTest - sig-auth (?) VolumeModeDowngradeTest - sig-storage gpuUpgradeTests - looks like sig-node/sig-scheduling kubeProxyUpgradeTests - both sig-network kubeProxyDowngradeTests - both sig-network serviceaccountAdmissionControllerMigrationTests - looks like sig-apimachinery the remaining less than half :+1: for sig-apps.\nSure - I didn't mean to put all of them to sig-apps. There is just the biggest group of apps related one so wanted to start with this group.\nI'm just making sure :wink: Go for it and thx :100:\nHello! The code freeze is starting March 9th, 2021 (about 1 week from now). As the Issue is tagged for 1.21, is it still planned for this release?\nYes this is planned for this release, and PRs related to this are not subject to code freeze. This will be completed by test freeze (see timeline here: )\n- calling your script from head, I have: So there are just couple exceptions now (and a mess with upgrade tests :) )\n- when you have time, there are couple PRs waiting for you now\nNot reviewing on the weekend, but will check them out first thing next week. Also ok if someone else gets to them before me\nRun in master. The is due to The are all from It seems it is done.\nThe two opened PRs + the one that I have internally (close to ready, though dependent on the firs from the above) should actually close this.\nwill close this out by adding a blocking test and docs to\nNice work ... that jq stuff is something :)\n/triage accepted"}
{"_id":"q-en-kubernetes-d6237202a7c4cbc2f8b6753b131788cd5c8097147495a990999752a426439104","text":"is a very useful option in scheduler perf test framework to define an operation without waiting for it to finish. However, the current testing logic still waits for all the pods in this operation to be schedulable, in the last step: the number for Pods in this operation to be schedulable This can be problematic as some case just want to generate Pods (for example, to test internal queuing or sorting behavior/latency). The following test command times out as those large-cpu Pods are expected to not fit on any node: So, given is commented as: I'm proposing to relax the semantics to not wait for the Pods to be schedulable if the operation is specified as . /sig scheduling\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign WDYT?"}
{"_id":"q-en-kubernetes-42665a877e4575e36f3bb694c9e3c1b9184bb837416571eeaf4386a287c8cb67","text":"What happened: Sweeping uses of extensions/v1beta1 in preparation for , the addon manager stood out as using extensions/v1beta1 Ingress by default: What you expected to happen: All components shipped in 1.21 must use Ingress in order to work properly against a 1.22 API server. /sig networking cc per cluster/addons/addon-manager/OWNERS cc I'm not sure what the addon manager does if given a prune list of Ingress objects while still using a manifest containing extensions/v1 Ingress objects... this needs to be tested. /milestone v1.21 /priority important-soon /lifecycle frozen\nThe label(s) cannot be applied, because the repository doesn't have them <detailsIn response to : What happened: cc /priority important-soon /lifecycle frozen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig network\n/triage accepted\nhow \"urgent\" is this? Can wait until next tuesday? Otherwise I can work on the tests this weekend :)\nimportant, not urgent. it should happen early in the 1.21 cycle, but does not require work out of business hours\nHello! I am from the Bug Triage team! This issue has not been updated for a few weeks, so I'd like to check on the status. Code freeze is starting March 9th, 2021 (about 1 week from now). We want to ensure that each PR has a chance to be merged before code freeze. As the Issue is tagged for 1.21, is it still planned for this release?\nwe're waiting to the PR to get merged :) I guess it's important for the v1.21 release, for the sake of extensions apiGroup deprecation :) Will ping reviewers on the PR"}
{"_id":"q-en-kubernetes-842dae7bf3431a97dcb89bde5e79459b21ecf7e42395a13e818bcee10e8c9425","text":"When both and parameters are set on Kubelet, they act as the upper and lower thresholds that define how the state is entered and exited. When trying to remediate a , Kubelet performs different operations to free up space. In order to decide if it has freed up enough, it compares the currently available disk space with the triggering threshold instead of the exiting threshold. Kubelet ends up in a long lasting state which can only be left with Kubelet restart, or if by luck the Pods that were using disk space are rescheduled somewhere else. Kubelet periodically does some housekeeping in order to ensure its node has enough resources available. If that isn’t the case, it tries to free up resources, updates the node’s condition (e.g. ) and in last resort starts evicting Pods. By default, the housekeeping is done every 10 seconds for a set of “signals” which represent node level resources (e.g. , , , ). There are a couple of parameters that comes into play in this process and I'll ignore all but two for simplicity: Signal thresholds: defines at which point there is not enough of a resource available, e.g. Signal minimum reclaim: how much resources Kubelet should aim to free up before a threshold is considered as not met. It’s the lower bound of an hysteresis, the threshold being the upper bound (e.g. ) The housekeeping is implemented in Here is my understanding of what this method does, filtering out what I think does matter in this issue: the signals (e.g. how much is available) a list of thresholds that are met (e.g. observed < ) there were thresholds met on a previous run: a. Check if those previous threshold are still met taking the into account. If = 15% and , our value to unmet the threshold is actually 25% percent b. Merge them with (2), the thresholds that are met from the new observation (1) out thresholds for which the individual grace period is not reached. They might be taken into account in another run if observed again. the previously met threshold with the new computed list of thresholds. Those met threshold are stored across runs. no threshold is not actually met, exits. If at least one is met, Kubelet has work to do! Try to reclaim node level resources. In other terms, try to free up resources without having to evict any Pod. For that’s done by removing stopped containers, and unused images. a new observation. If the threshold are not met anymore, ignoring , exit with Pods as last resort Ignoring the in (7) seems wrong. If the amount of resources freed is somewhere between and , Kubelet can get stuck. On the next run, as part of (3a) the threshold taking into account is not unmet. Therefore Kubelet tries again to free up node level resources in (6), eventually does nothing but exits (6) pretending it free up enough resources without doing anything. after having cleaned up the node and in order to decide if Pod eviction is the step, seems to do the trick. Kubelet should try to get out of a by evicting almost all Pods. In this case, it means it should try to evict Pods until the threshold is reached and the can be removed. I expect Kubelet to stay in permanent only if the resources that are used are beyond its control. Preparation: a node with a different from (the problem might be the same when they are mixed but I didn't try) to 15% and to 30% in the Kubelet’s configuration into a Pod running on that node, and use to write a file inside the container. The amount of data should be enough to bring the underlying disk at 86% usage for Kubelet to notice the threshold and to cleanup stopped containers and remove unused images Remove your file. The node is now clean. Restart Kubelet and notice how it sees everything is normal Getting Kubelet stuck: into your Pod and write a file to reach 86% usage again for Kubelet to notice, cleanup stopped containers and images. Since the cleanup was done in the preparation step a minute ago, there is almost nothing to be removed and your usage should stay more or less the same the Pod's file by a new one to reach between 55% and 85% disk usage. how every 10 seconds, Kubelet pretends it was able to . The Node is stuck in until you restart Kubelet and it forgets it had to ensure there is at least available in . Kubernetes version (use ): 1.18.8 Cloud provider or hardware configuration: vSphere OS (e.g: ): 18.04.2 LTS Kernel (e.g. ): 4.15.0-130-generic We have been running a slightly modified version of Kubelet on one of our clusters for the past 5 days, and since then we had no node staying in state for more than 20 minutes. From the disk usage graph, it looks like both image garbage collection and the disk pressure remedy operations are working properly. I'll open a PR with the little patch we're using.\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig node"}
{"_id":"q-en-kubernetes-7bd844af0d2175d75b9cd7af255da901ebe734b41b62990f2447aeb3ad83fc45","text":"<!-- Please only use this template for submitting reports about continuously failing tests or jobs in Kubernetes CI --/sig windows /cc After was merged I think.\nI agree that looks like the issue. It started failing between the two commits: on this run:\n/triage accepted\n/assign"}
{"_id":"q-en-kubernetes-7502f6fe7f97023e3c49d1500e19b5d826a5bad20ba6d4677a3bc38b3c03cd46","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --Start the scheduler with a scheduler config file which has a custom configuration for a specific plugin, the scheduler will fail to start and throw below exception, kube- error message, custom plugin config should take precedence over default plugin config. start a scheduler with a config file. Kubernetes version (use ): Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others:\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig scheduling /assign\nDoes disabling the plugin and re-enabling the same plugin with a different weight work as a workaround?\nThis is WAI. You have to explicitly disable the default configuration.\nI think so.\nI not sure I agree with this, particularly, for someone who doesn't know details of code IMPL, e.g. an operator, Why bother to disable the plugin first and then re-enable the same plugin? there is no pointer to show they are different in a config file, as a developer, we know the reason behind the scenes, can we assume operator also know what that means to define a schedule config? By disabling the plugin and re-enable it means the operator has the knowledge this plugin is a default enabled plugin, should the operator check the source code to confirm this plugin's configuration? When this happens, given a different config for a specific plugin, should this deserve a corruption? I think scheduler should more robust to tolerate this.\nWe can improve the docs to include the default plugins. It's better for the API to be more predictable (always do the same thing). One thing to think about is order: if we do what your PR proposes, we are altering the order in which plugins are run, not just modifying the \"weight\" of the plugin.\nI'm leaning towards closing this ?\nIt's the key point I'm concerned about: commented here So let me ask a question: for a given input like: What's the expected output? {DefaultPlugin1, 1}, {DefaultPlugin2, 200}, {DefaultPlugin3, 3}, OR {DefaultPlugin1, 1}, {DefaultPlugin3, 3}, {DefaultPlugin2, 200}\nToday, it is 2.\nSo in other words, if a user wants to customize a plugin's weight and keep its order, the only way is to use and give a full list of plugins? If so, that's not a good UX IMO.\nWell, order doesn't matter for Score, so not really a problem.\nAll right, so the only case users could disable PluginXYZ while also enabling PluginXYZ is Score plugins, I'm fine with the order issue. Back to this issue, my 2 cents is to change the semantics of to a more enforced manner (proposed in this issue) in a new API version - v1beta2 or GA. As IMO it's a real UX improvement. WDYT?\nI'm not convinced it's a UX improvement. If the user \"enables\" a filter plugin that was a default (say, by mistake), then its order changes. At least first disabling the plugin makes it more explicit. cc who has been working on decisions based on CC version\nI guess it depends on the implementation. What is the full proposal?\n/remove-kind bug /kind feature\nThat's a fair point, esp. if we adjust default plugins from version to version. I think a good UX is to throw an error/warning to indicate that a default Filter plugin is being re-enabled - this's useful no matter which semantic employes.\nI'd like to: keep current logic to warn users that \"PluginXYZ\" is already registered (by default), if it's a non-Score plugin apply the \"overridden\" semantics to directive, for Score plugins only btw: let's get a consensus first and better adopt it in new CC version.\nI propose (only for new version): Fail when plugin is already registered in the custom configuration, and only warning if the plugin was already part of defaults. Override any weight, preserving order I think that keeps everything consistent. I think this is what the last iteration of the PR is doing, but I only skimmed through it.\nSGTM. To confirm, by \"Fail when plugin is already registered in the custom configuration\", this's only possible when users implement an out-of-tree plugin with a name identical to a default one, right?\nI'm just referring to failing when the user repeats the same value twice. I think we can't prevent what you say.\nThis might be the reason why the order of the plugins should be reserved.\nUnless it's score plugin like this issue, I don't know if there is any point to enable the default plugin in a config file. I think there is no need to distinguish score plugin with others, this will make the life easier, correct me?\nYeah, overall, is following this ideas, except that PR doesn't handle the repeated configuration, because if there is repeated value after the default plugin is overridden, it will hit the error below,\nthanks for those valuable suggestion! So, we are still with beta1, and you suggest we can make some change in the new version, beta2? Is there any plan when this could happen? v1.22?\nYou are correct, there is no point, but it's good to have consistency. If it's allowed for Score, it should be allowed for other extension points. I agree Yes, that is the plan. It was 1.21, but we fall behind due to newer plans for deprecating plugins\nlet's say there are 3 default non-Score plugins A, B and C (in that order). I want to move B to the end of the list but I want to custom enable C too (for whatever reason; since we want to consistent with score plugins, let's allow this for other extension points too as said). if a non-Score plugin is custom disabled AND enabled, we should move it to the end of the list but if it's simply just enabled, then we should preserve order. if it's a score plugin, disabling and re-enabling has the same semantics as just simply enabling; they both will only change the weight. (idk if this has already been proposed; if it has been, consider this a vote of support :))\nAre you saying that \"disabled and enabled\" should be treated the same as just \"enabled\"? That is, always move to the end of the list?\nNo, just \"enabled\" would NOT move things to the end of the list (it will preserve order), but \"disabled and enabled\" would move to the end of the list.\nok, that sounds good\nAnother problem here is if we don't want to overridden anything but just config a plugin which is accidently a default plugin, something like below, scheduler will fail as well, I think we need to De-dup here instead of just a warning here. I guess this PR should already cover this, I will go back to check it again, at least one more testcase is needed to cover this case. keep current logic to warn users that \"PluginXYZ\" is already registered (by default), if it's a non-Score plugin\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle stale\n/lifecycle stale let me revive the PR later.\n/remove-lifecycle stale"}
{"_id":"q-en-kubernetes-132a80ff98e1e6dd0a89d4bfcef8ef92e7c22e3237cafbba2867719c6b2e36dd","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --kube-up tries to add metadata to the wrong project. It is trying to add metadata to the prow project running the tests instead of the target project. This is because is missing a kube-up uses the target project to add metadata Don't set/set a different default project and run cluster/kube-up. originally in kubetest hides the project acquired from boskos under the covers by setting it as the default project uncovered as part of migration to kubetest2 xref: /cc\nThere are no sig labels on this issue. Please add an appropriate label by using one of the following commands: - Please see the for a listing of the SIGs, working groups, and committees available. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-facb8043a3df3974473d7c79b50f1e5dce5028f3ba86dc13c34e2a3b213f89a1","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --When is passed a ref while logging, it panics: or an empty reference should be logged. I have fixed this in so we will need to bump klog. ht for reporting in /priority critical-urgent /sig instrumentation\nneed update klog dependencies later. milestone 1.21\n/milestone v1.21 Will set the milestone for the issue and let people review the PR during regular burndown.\nScope check: potentially affected components are: Kubelet Controller manager Scheduler (grep for KObj: )\nI looked in current implementation of handling object. Please remember how checking works. Current fix should work on simple nil (): but it will fail on structs implementing KMetadata interface, but not handling nil pointer (). Reason is that interface with nil value is not nil. Resulting call to 'Log' function will panic on call. In stacktrace presented above it was function that panicked which leads to thinking that the problem is in function which panics if pointer is empty. Potential solutions: Replace all calls with until we solve this issue Fix this particular case where pod can be nil (solves only local problem, but not the long term issue that every log call can be a panic) Use reflection to verify if object under interface is an nil (expected overhead EDIT: got results only 7% increased cpu usage) Fix all and functions to check for nil (not enough time before the freeze, some overhead on calls) Do runtime checks for K8s types (would require klog to import api types) Looking at all options I think that first one makes most sense as all other have bad impact on whole k/k repo\nDid benchmarking for using reflections, looks like performance overhead is not as significant as I previously anticipated, around 20%. I think we could consider using reflections.\n/cc For feedback\n/cc\nThere are hundreds of instances of in k/k now, and every single one would need to be carefully checked when obtaining the pod name/namespace, and we'd need a top-level approver to do the replacement. I think this really needs to be fixed in klog.\nWith benchmark results being not as bad as first estimated I think we could use reflection in klog like proposed in kubernetes/klog\n/reopen needs to get bumped in k/k to close\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/triage accepted"}
{"_id":"q-en-kubernetes-e9d472d5a01abca954a4ab598d29e5b2f82ddd11a788a2b2ecc287e7ab34e0d8","text":"The method doesn't check taints and tolerations. This is problematic in 2 scenarios: If the users bypass scheduler by setting a value in If the node restarts. The node would try to restart scheduled pods once it comes back up, even if it's not ready (has a not-ready taint). Kubelet to check taints and tolerations. a taint in a node. a Pod with set to the Node above, without a matching toleration. /sig node /sig scheduling\nthis reminds me of your kubelet sync PR ...\nI don't think this is a syncing issue. The logic doesn't check taints at all.\nscenario 1 is pretty bad, i will let the #sig-node slack know about this.\nIt doesn't sound like a syncing issue, but we seem to have multiple issues with pod admission/management on restart currently\nI'm aware of that issue. This should be unrelated.\nI would like to address this. /assign\nI think it would be great to proceed after merging ref:\nPlease add me as reviewer\n/triage accepted\nIssues go stale after 90d of inactivity. Mark the issue as fresh with . Stale issues rot after an additional 30d of inactivity and eventually close. If this issue is safe to close now please do so with . Send feedback to sig-contributor-experience at . /lifecycle stale\n/remove-lifecycle stale PR in review:\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\n/remove-lifecycle stale\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\n/remove-lifecycle stale"}
{"_id":"q-en-kubernetes-b79f17d4ba155bb2a820042336c3b9d69b344618f78a325028d3c17b0646adf1","text":"[x] test approval issue : [X] test pr : 26 May [x] two weeks soak end date : 9 June [x] test promotion pr : kubernetes/kubernetes According to this APIsnoop query, there are still some remaining DaemonSet endpoints which are untested. Create a watch to track daemon set events that DaemonSet with a static label Create a daemon set with a static label. Confirm that the pods are running. Get the daemon set status. Parse the response and confirm that the daemon set status conditions can be listed. Update the daemon set status. Confirm via the watch that the status has been updated. Patch the daemon set status. Confirm via the watch that the status has been patched. Using an existing as a template for a new which validates that three new endpoints are hit. This query shows the endpoints hit within a short period of running the e2e test If a test with these calls gets merged, test coverage will go up by 3 points This test is also created with the goal of conformance promotion. /sig testing /sig architecture /area conformance\n/triage accepted\n/reopen\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-1937bc6ff097136d61bdcffb4b5f031c13c8962141a01f9eaee469752986052f","text":"Originally tracked at . After fixing \"not include test files in import-boss\" in , we have a regression here: is using . We need remove dependencies from integration tests to e2e tests again. /assign /sig testing\n/sig auth as per the required change\n/kind cleanup /area test\nIt would be helpful to outline an argument as to why this is the case. I can imagine reasons, but none are mentioned here.\nDescription updated.\n/triage accepted /priority backlog /milestone v1.21 This seems a simple enough fix to land (and thank you for doing this) but marking low enough priority to make it clear we may back out and wait for v1.22 if we don't have bandwidth for this\nSo while I think we should probably do this anyhow I want to note: This issue lists the reasons: we have no intention to stage e2e (there's been much discussion about this since then) we generally don't want anything non-informational (required jobs) running only sometimes anyhow, it's just not reliable Decoupling these seems like a good idea from a purely code-organization perspective, but the reasons in the previous issue are no longer valid.\n/milestone v1.22 The PR didn't land in time. At this point I view this as nice to have. Even if it's low impact, let's wait until next release."}
{"_id":"q-en-kubernetes-acaad09ffe9089aa1ab907f13c735e3b5b6e231b29e84b98246018036fc733fd","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --e2e test fails with ECDSA Service Account key e2e test to succeed with ECDSA Service Account key Use ECDSA key for the service account key. Logs from the started test container: library defaults to by default as the only supported method while should be explicitly: Relevant code: On the API server side: Kubernetes version (use ): 1.21.0-rc.0 Cloud provider or hardware configuration: OS (e.g: ): Talos Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others:\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig testing\nChanging the service account key type to RSA fixes the test:\nthis is a problem for us, as this test is part of test suite, and it fails with 1.21-rc.0 e2e tests\nSIG Auth own their test cases (it's tagged in the test name), SIG Architecture own conformance. Testing owns the frameworks, CI, tooling, etc. EDIT: just clarifying and making sure the right people see it. I raised this to the conformance work group.\n/assign /priority critical-urgent /milestone v1.21 because this is in the conformance suite, it must not restrict valid+supported cluster configurations\n/reopen the image needs to be promoted and bumped in the e2e test\nReopened this issue. <detailsIn response to : /reopen the image needs to be promoted and bumped in the e2e test Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\npromote image PR:\nI believe with merged this can be closed as all the fixes were done and backported to 1.21\nThanks!"}
{"_id":"q-en-kubernetes-07888485ec8613c44f8e19f56ea34081a93556565b1203f65ca7f9a0248b5ff5","text":"<!-- Please only use this template for submitting reports about flaky tests or jobs (pass or fail with no underlying change in code) in Kubernetes CI --pull-kubernetes-e2e-aks-engine-azure-file pull-kubernetes-e2e-azure-file-windows AzureFile CSI Driver End-to-End Tests: Dynamic Provisioning should create a volume on demand with mount options [] [] [Windows] AzureFile CSI Driver End-to-End Tests: Dynamic Provisioning should create multiple PV objects, bind to PVCs and attach all to different pods on the same node [] [] [Windows] AzureFile CSI Driver End-to-End Tests: Dynamic Provisioning should create a volume on demand and mount it as readOnly in a pod [] [] [Windows] If the randomly generated kubetest resource group has capital letters, the share name will contain invalid characters and fail provisioning. links to appreciated links to specific failures in spyglass appreciated <!-- Please see the deflaking doc () for more guidance! --\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-65a32798bdc2e97c37a09b152287efa93004647f02464dbec67aae5381822722","text":"With recent migration of Kubelet to structured logging introduced e2e tests that enable JSON in Kubelet. Organized migration of other components haven't yet started and is on best effort bases, yet some line were already migrated. As code-base will be slowly migrated, we need to expand the test to avoid errors like I propose to expand existing e2e test suite ( and ) to cover all control plane components (apiserver, controller-manager and scheduler). /cc\n/sig instrumentation\n/help\nIf all components have this flag do we just want to enable it on all components? We already increase the log level on all components in CI, it should be pretty straightforward to expand this to all: pick a new environment variable name for the script with a more appropriate name for all components instead of just kubelet set it on the job in addition to the old kubelet one, by sending a PR to test-infra send a PR to the kind script to set the value in all places where we set verbosity instead of just kubelet, switching to the new variable name\n/triage accepted\n/area logging\nSo currently the e2e tests do not actually read and test what the logs return the current job just makes sure that kubelet is running Would we want to write similar tests for the different control planes while we start migration or do you have something else in mind with respect to the automated testing ?\nYes, currently we just test if Kubelet is running to test that flag is working and if enabling json doesn't cause panics. This exact issue proposes to expand this to other control plane, but your idea would be a great follow up. Would be great to propose a test plan that will specify what we want to test and how. If you have ideas some ideas could you collect hem in some issue so we can discuss.\nThere's no portable way to get these logs. There are non portable ways but those cannot be used in a standard e2e test.\nOh so for the standard e2e tests it might be good enough to just test the startups of these control planes with the logging enabled.\nI was thinking that we could get started with maybe writing similar to one for the remaining components and then maybe find a way of checking the logs generated as we have bigger discussions a) kube-apiserver --Completed b) etcd --- TODO c) kube-scheduler ---TODO d) kube-controller-manager --TODO\nsound good. FYI Etcd is not a kubernetes component as so it's out of scope of this issue.\n/wg structured-logging As getting core components logs is not portable, we could test a container with same logging library and using Kubenetes logs API to access logs. Created to track it\nit still might not be a terrible idea to write a simple test script to probe the exported logs from a cluster with json logging enabled for any non-json lines once we know that everything should be. I think this could probably be hacked up with jq or something, but I wouldn't position it like a normal e2e test.\nCould you give a hint how to implement and run such script?\n/priority important-soon This issue is based on feedback from Kubelet migration retrospection, it is required before organizing migration of next component. /cc\nWould this follow the same flow as in where we check if json logging is enabled and then run jq separately or within the same flow. To be more specific: Seems to be getting the data and validating it so in that case would we be running after grabbing the logs ?\nFor now let's focus on the original tasks of having ( and ) test scenarios, also configure other components. We just need to update Kind script to also pass to other components. Example:\nWould like to take this one! /cc\nGreat, let me know if you need any help? /remove-help\njust created a PR:\nComponents to be picked: [x] kube-scheduler [ ] kube-apiserver [ ] controller-manager\nAttempted to cover the remaining components in this\nIt would be great if we had e2e setup for all components before we start structured migration next week. will you be able to work on it this week? Please let me know if there is anything I can help with.\nThanks for help with control plane components, for 1.23 we are planning to migrate kube-proxy to structured logging. Could you also try to enable JSON logging on kube-proxy?\n/reopen\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nConfirmed that in last CI run Logs from kubelet and control plane components are written in JSON: Great Job Could you take a look into changing log format of kube-proxy?\nFiled for kube-proxy. Are we good to close this issue ?\nsg"}
{"_id":"q-en-kubernetes-8bbee7e9a6860cdafba148d29efcd05cc1b83ff0ba8c5fce73f89572318cc3d6","text":"<!-- Please only use this template for submitting reports about continuously failing tests or jobs in Kubernetes CI --pull-kubernetes-integration runkubectldiff_tests It only shows failure cases in recent 2 days. The last date is I tried to reproduce it locally, but can't\n/sig testing\n/sig cli\n/cc\nI’ve only briefly taken a look at this, but my understanding is that server-side diff is flaking because the timestamp is different (by like a second) between the one already applied and the diff apply. do you think this could be related to this recent change Could delaying when we update the timestamp cause it to sometimes get misaligned with what’s being diffed? I’m not super familiar with diff, but it feels like it should ignore differences between timestamps that get updated on each call, right? Is there some sort of mechanism we have in that knows to ignore timestamps that just isn’t being done for fields?\nThis error probably shows that we have a path that causes the object to be re-written to etcd when it shouldn't.\nAs we can see, the diff only shows a difference in the timestamp but not in the fields that are managed. We are not supposed to update the timestamp if the managed fields haven't changed. That timestamp only gets updated if the two steps happen in different seconds (that's why it's flaky). That is definitely not a flake but an actual failure.\nFailed 81/236 times today in /priority critical-urgent I think, correct it if I am wrong.\n/triage accepted\ncorrect. who should be assigned as the owner of this?\nI can take it, I'm looking into it now\n/assign"}
{"_id":"q-en-kubernetes-38588f81bc6d22a43525c6f262a79282cecfd50313af364d96c67168bdad1928","text":"<!-- Please only use this template for submitting reports about flaky tests or jobs (pass or fail with no underlying change in code) in Kubernetes CI --k8s-e2e-ltsc2019-containerd-flannel-sdnoverlay-master [] Container Lifecycle Hook when create a pod with lifecycle hook should execute prestart http hook properly [NodeConformance] [Conformance] [] Container Lifecycle Hook when create a pod with lifecycle hook should execute poststart http hook properly [NodeConformance] [Conformance] Note: the tests have been excluded until they are fixed. In the case of multinode clusters, the http server pod and the test cluster can spawn on different nodes, which can be problematic for poststart / prestop hooks, as they are executed by the kubelet itself, and the cross-node lifecycle hook might fail (according to the [Kubernetes network model], it is not mandatory for kubelet to be able to access pods on a different node). However, that is not always the case, different Networking setups allow cross-node communication between kubelet and pods, which means that the above tests are passing properly (for example: ), which begs the question if this behaviour should be universal or not, especially since the tests are marked as , which would imply that it should work in all cases. [Kubernetes network model]: links to appreciated links to specific failures in spyglass appreciated Slack: /sig testing /sig networking /sig windows &lt;!-- Please see the deflaking doc () for more guidance! --\n/sig network /sig node /area conformance /assign /priority critical-urgent /cc This is important since affects conformance and defines the networking behaviour for the poststart/ prestop hooks\nThe label(s) cannot be applied, because the repository doesn't have them. &lt;detailsIn response to : /sig network /sig node /sig conformance /assign /priority critical-urgent /cc Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. &lt;/details\ni'd personally let folks have first dibs on what we need to do here. Though from the name of the test itself, it seems like the test is NOT a sig-network test :)\nIndeed, it does not have any tag in its name, but the tests rely on networking and on how CNIs work. As I've stated above, it's not a problem for flannel sdnbridge, but it is for flannel sdnoverlay; other CNIs might have a similar issue too.\nbecause the network is not a thing :) let´s tag the sig owner then /sig node I think that the key here is to decide if kubelet should be able to access pod in other nodes, and based on the link Claudiu is part of the Kubernetes networking model. I saw users that doesn't want kubelet to access even the pods on same host, so I think that is legit to say kubelet MAY reach pods in other nodes but not MUST ... but I'm not familiar with the to fully understand the context. If kubelet \"may\" reach pods in the other nodes, the test must only test lifecycle hooks for pods in the same node as the kubelet under test, right?\nThis is correct. The model also says: \"For those platforms that support Pods running in the host network (e.g. Linux): pods in the host network of a node can communicate with all pods on all nodes without NAT\" I believe that Windows host network is not supported, right?\nRight, and there are specific tests that checks that exact scenario you've mentioned: However, the flaky tests I've mentioned in the issue require kubelet to be able to dial-in into a webserver pod for a different pod's poststart / prestop HTTP lifecycle hooks. There are couple of things regarding this: of all, in order to ensure kubelet would be able to run the HTTP lifecycle hooks, we'd have to make sure both pods are on the same node, as that is a supported scenario, according to the Kubernetes Network Model. I have sent a PR for this above. Alternatively, a Cluster IP could probably work as well, but IMO, ensuring that both pods end up on the same node is simpler and easier to clean up. , it's interesting that those HTTP lifecycle hook tests work on different networking setups, without any pods having HostNetwork enabled, even though you've mentioned they should (e.g.: Linux / Windows with flannel l2bridge, and even Linux with flannel overlay). Now, my question is: is this behaviour a feature, or a bug, especially since the tests are marked as `?\noh, I missed that part, so in Linux kubelet must be able to communicate with pods on all nodes (becuase it is running in the hostnetwork) If that is correct, it seems that one solution is to keep the test for Linux and tag it as LinuxOnly, and a new test for WindowsOnly that only spawn the pods under test in the same node EDIT Discussed in sig-net meeting , no clear answer... can be kubelet considered as a hostnetwork pod?\nFor the purposes of testing, I don't mind forcing the pods onto the same node., but let's document the heck out of that at the code-site?\ncorrect but it will be supported once HostProcess containers are merged in 1.22: Only in Overlay networks (ie Flannel where these tests are flaking) does the pod have to be on the same node. It doesn't require host network mode and these tests are . I would be more inclined to disable this test for Windows Overlay test setups and have a sig-windows test for the overlay functionality. I would also question if the lifecycle hooks only running on the same node is useful and if it should just be documented that it doesn't work in the Windows Overlay setup.\nis looking at test PRs from SIG Node. I will let sig-net triage this issue\nNext steps on this?\n/triage accepted\nWe still haven't merged this: I also had this, but it has been decided to be blocked:\nrequires conformance approval, at this point I think that has to be postponed to 1.23 right, since that behavior is not required, we shouldn't test it, because is legit to fail and having a test for something that can fail can create confusion :smile:"}
{"_id":"q-en-kubernetes-bd5e4e437ff810dc05756f26f3dd878a6de1d966beaad1f34c3899f2a783903b","text":"kubelet logs misleading when volume limits not an\n/sig storage /tirage accepted /assign\nNote that it is set as info level in the log above: However, this is definitely confusing. /remove-kind bug /kind cleanup /triage accepted /sig node\nI also have this problem Excuse me, how did you solve it\nWhich version do you use .. This was fixed here Check k8s version you use please"}
{"_id":"q-en-kubernetes-0c5888dda701f0ad75368add63a64093c3931265e9811edd78389534818e2186","text":"The test only validates that the API objects are created E2e tests must verify the behaviour, i.e. that the Service and Endpoints actually work sending traffic. The file is not well organized and there is a lot of duplicate code, one possibility is to use current Services helper has a function to test the reachability that iterates over all service ports However, something that needs to be addressed is that the pods that are backing the service doesn't have any process listening on those ports\n/help wanted /sig network /priority important-soon /triage accepted\nI think that there are more tests that doesn't verify the reachability , those should be modified too /help-wanted\n/help-wanted\n/assign\n/triage accepted\n/assign /unassign I will take on this since it was more difficult than I expected, and is a Conformance test.. I assigned you to , that is similar to this, but creating a new test from scratch"}
{"_id":"q-en-kubernetes-a9abd39d55ca65ead711b529bd35608b6b03d2f91fd8b818738f1b178316e6e8","text":"<!-- Please only use this template for submitting reports about continuously failing tests or jobs in Kubernetes CI --post-kubernetes-push-e2e-agnhost-test-images Image build unhappy on s390x arch: Building image for agnhost OS/ARCH: linux/s390x... [...] [ 3/10] RUN apk --update add bind-tools curl netcat-openbsd iproute2 ipe... 0.165 fetch http://dl- 0.725 fetch http://dl- 1.719 (1/35) Installing ncurses-terminfo-base (6.2p20200523-r0) 1.758 (2/35) Installing ncurses-libs (6.2p20200523-r0) 1.811 (3/35) Installing readline (8.0.4-r0) 1.851 (4/35) Installing bash (5.0.17-r0) 1.939 Executing bash-5.0.17- 1.945 ERROR: bash-5.0.17- script exited with error 1 1.945 (5/35) Installing fstrm (0.6.0-r1) 1.987 (6/35) Installing libgcc (9.3.0-r2) 2.013 (7/35) Installing krb5-conf (1.0-r2) 2.053 (8/35) Installing libcomerr (1.45.6-r0) 2.093 (9/35) Installing keyutils-libs (1.6.1-r1) 2.133 (10/35) Installing libverto (0.3.1-r1) 2.174 (11/35) Installing krb5-libs (1.18.3-r0) 2.334 (12/35) Installing json-c (0.14-r1) 2.376 (13/35) Installing libstdc++ (9.3.0-r2) 2.493 (14/35) Installing libprotobuf (3.12.2-r0) 2.667 (15/35) Installing libprotoc (3.12.2-r0) 2.807 (16/35) Installing protobuf-c (1.3.3-r1) 2.854 (17/35) Installing libuv (1.38.1-r0) 2.906 (18/35) Installing xz-libs (5.2.5-r0) 2.954 (19/35) Installing libxml2 (2.9.10-r5) 3.071 (20/35) Installing bind-libs (9.16.15-r0) 3.287 (21/35) Installing bind-tools (9.16.15-r0) 3.363 (22/35) Installing ca-certificates (-r4) 3.426 (23/35) Installing nghttp2-libs (1.41.0-r0) 3.456 (24/35) Installing libcurl (7.76.1-r0) 3.510 (25/35) Installing curl (7.76.1-r0) 3.561 (26/35) Installing iperf (2.0.13-r2) 3.607 (27/35) Installing libbz2 (1.0.8-r1) 3.649 (28/35) Installing fts (1.2.7-r1) 3.691 (29/35) Installing libelf (0.179-r0) 3.774 (30/35) Installing libmnl (1.0.4-r0) 3.814 (31/35) Installing libnftnl-libs (1.1.6-r0) 3.865 (32/35) Installing iptables (1.8.4-r2) 4.387 (33/35) Installing iproute2 (5.6.0-r0) 4.535 Executing iproute2-5.6.0- 4.540 ERROR: iproute2-5.6.0- script exited with error 1 4.540 (34/35) Installing libbsd (0.10.0-r0) 4.584 (35/35) Installing netcat-openbsd (1.130-r1) 4.627 Executing busybox-1.31.1-r20.trigger 4.632 ERROR: busybox-1.31.1-r20.trigger: script exited with error 1 4.632 Executing ca-certificates--r4.trigger 4.686 2 errors; 29 MiB in 49 packages ERROR: executor failed running [/dev/.buildkitqemuemulator /bin/sh -c apk --update add bind-tools curl netcat-openbsd iproute2 iperf bash && rm -rf /var/cache/apk/ && ln -s /usr/bin/iperf /usr/local/bin/iperf && ls -altrh /usr/local/bin/iperf]: exit code: 2 failed to solve: rpc error: code = Unknown desc = executor failed running [/dev/.buildkitqemu_emulator /bin/sh -c apk --update add bind-tools curl netcat-openbsd iproute2 iperf bash && rm -rf /var/cache/apk/ && ln -s /usr/bin/iperf /usr/local/bin/iperf && ls -altrh /usr/local/bin/iperf]: exit code: 2 make: ** [Makefile:43: all-build-and-push] Error 1 ERROR ERROR: build step 0 \"\" failed: step exited with non-zero status: 2 ERROR: () build -78ae-4086-bc26- completed with status \"FAILURE\"\n2021/04/30 16:58:31 Failed to run some build jobs: [error running [gcloud builds submit --verbosity info --config --substitutions PULLBASEREF=master,WHAT=agnhost,GITTAG=v20210430-v1.22.0-alpha.1-58-gf039f94a6fe --project k8s-staging-e2e-test-images --gcs-log-dir gs://k8s-staging-e2e-test-images-gcb/logs --gcs-source-staging-dir gs://k8s-staging-e2e-test-images-gcb/source gs://k8s-staging-e2e-test-images-gcb/source/-1c8a-4141-9bbb-]: exit status 1] Need this fixed to finish image version dance to fix conformance tests for 1.21.1:\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThis has failed 3 times consecutively now. All because s390x issues ... I've manually re-triggered it the fourth time. /sig testing but also /sig release because this is a major problem for continuing to ship this platform. cc We've seen this before with flakes across other image builds for this platform.\nlooks like the latest run passed -\nthere's still underlying flakiness i need to get to the bottom of\nAlso what latest run? Were new changes made without bumping the version? Or did someone else manually trigger it yet again?\nI just saw a green run at and the image was pullable, so I thought it had succeeded\nI think someone must have retried it again manually. Postsubmits only run on merge, and we've had nothing else merge\nGuirish tried to see whether our local s390x environment has the same problem too, he'll update us here.\nthe issue seams to be intermittent caused while pulling respective binaries from Alpine repo. I tried to reproduce the issue locally, I am observing similar build failures while cross building the image for other ARCH too, so the issue is not specific to s390x and it is to do with alpine repo.\nthanks i've filed a PR that may help mitigate please review!\nlooks like a long pending problem\nInteresting. FWIW 4/4 consecutive failures appeared to be s390x in CI. We've had previous flakes relating to symlinks affecting this platform only."}
{"_id":"q-en-kubernetes-7798673857cea80b5648812382ad768f1113a275db3be1227df0714adb3f1ac5","text":"kubelet increased more than 100MB in memory in 4 days. kubelet to not used that amount of RAM I have reproduced it using 3 cronjobs which create hello world pods each minute and another cronjob that writes a line to a file each minute. Kubernetes version (use ): v1.18.1 Cloud provider or hardware configuration happens on both virtualbox and baremetal OS (e.g: ): centos 7 Kernel (e.g. ): 18.0-147.3.1 I have seen some existing issues like but I do not have any dangling resources associated with the cgroup.\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig node\n/assign This is a duplicate issue, there has been a fixed PR but it has not been merged."}
{"_id":"q-en-kubernetes-e2f4fe7b99012f8e7173e2f6c46ef5707967f19510033d4b6b03e3a31c7298f0","text":"<!-- Please only use this template for submitting reports about continuously failing tests or jobs in Kubernetes CI --ci-kubernetes-e2e-gci-gce-slow [sig-storage] Dynamic Provisioning DynamicProvisioner [Slow] should provision storage with different parameters 5/13 https://k8s- v1beta1.StorageClass got removed in We should remove the v1beta1 test.\n/assign can you help take a look?\n/triage accepted\n/sig storage\nI run the test with the latest from master without removing the beta test and it failed with a different error before getting to the beta test part: Do you have any idea what would've happened in my setup? I redeployed k/k using master HEAD\nFor this testcase: I saw that the provisioned PD has a size of 1.5Gi according to kubernetes: And in gcloud It's really interesting that the test was expecting 2Gi to be created, is this something that was changed recently?\nThe logs you posted are showing pd csi as the provisioner. Were you testing with the intree driver or csi driver? The CSI driver should to the nearest Gb. I believe intree does the .\nI think I did intree because I didn't send as an argument in my command: It's interesting that the PVC has this annotation too:\nmmm something interesting about my setup is that I created the cluster with CSIMigration turned on Do you think that might be forcing the usage of ?\nYes likely. Try with CSIMigration off and see if that helps. If so, then this potentially seems like a rounding bug with the csi driver\nThanks I tried with and it did indeed round up to 2Gi."}
{"_id":"q-en-kubernetes-e15f20de6681b8d4da493396450c2b5dfa66e8f08097d234d51b2316b3becc2e","text":"/sig scalability This issue is tracking process of upgrading etcd to 3.5.0 planned for 1.22 K8s release. Etcd has recently released Beta versions and is starting the process of correctness and performance testing. To reduce change of finding performance regression late, we will start qualifying Beta Etcd releases on GKE master branch and will upgrade to 3.5.0 as soon as it is available. Steps: [x] Do preliminary scale testing of etcd v3.5.0-beta.3 release [x] Release etcd v3.5.0-beta.3 image (, ) [x] Release etcd v3.5.0-rc.0 image () [x] Upgrade etcd server to v3.5.0-rc.0 () [x] Wait for 3.5.0 release and upgrade etcd client () [x] Release etcd v3.5.0 image (, ) [x] Upgrade etcd server to v3.5.0 /milestone v1.22\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/area etcd\nDoes this meant that the 3.5.0 should be promoted like\nYes, created\n/area release-eng release-team /sig release For tracking: cc:\nThe label(s) cannot be applied, because the repository doesn't have them. <detailsIn response to : /area release-eng release-team /sig release cc: Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nI do think this can be pointed in the Major Themes for the 1.22 release. I would like your opinions on this. If you do think it has a place in it, I will add it to the .\nHow can I be a contributor to this issue? what are the technical checklist that needs to be achieved to close out on the issue?\nPR to update Etcd is ready, but we found a blocker for merging . There is one test failing I will take a look into it on Monday, still more eyes on the problem could speed up resolution.\nWe'll also need a new kube-cross image w/ etcd v3.5.0. I'm working on that here:\nWe just need to upgrade the server to use new image like in\nok, looks like is busy. Let me take a look /assign\nThanks."}
{"_id":"q-en-kubernetes-598290730d0d1184c899554755e7243df1f7a1b5fb68af4568a1cff16b464a04","text":"Creating a service with fails to be created when trying to use a TCP health probe. Example service: Failed as follows: Health probe works correctly and is registered properly as TCP health probe. Take the mentioned service yaml and apply it, together with a corresponding pod / deployment exposing port 445. E.g. like this: The respective code seems to take the wrong path to me (it's +/- the same code on the in-tree as on the out of tree cloud provider): My reading of the code and passed values are like this: returns returns -- because returns returns therefore: So the requestPath is not at this point and the azure API server returns a 400 bad request. Kubernetes version (use ): 1.20.6 Cloud provider or hardware configuration: azure OS (e.g: ): linux Network plugin and version (if this is a network-related bug): vxlan & flannel\n/sig cloud-provider\nrelated PR introducing the feature:\n/assign /triage accepted\n/assign"}
{"_id":"q-en-kubernetes-c70f52d3dc7b7b99eededd5553e5feed635a0930aa274fff91207c03890e6f72","text":"When we introduced we an additional check during attach disk call to check for canonical path - Now this is creating a regression in many deployments because when volume is created, canonical path gets replaced - and hence when attach is attempted k8s can't find canonical path of the volume anymore. It throws following error: There are two ways of fixing this. We could decide to use old in attach code, if we can't fetch canonical path or we could decide not to replace canonical path during provisioning. cc /sig storage\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\ncc:\nThe code that replaces canonical Path during provisioning is pretty old and hence I don't remember/know why we are doing that. But since there could be many existing PVs that now have different canonical path - I think it might be safest to merge - to prevent any regressions."}
{"_id":"q-en-kubernetes-21127304aaa08a391779a653554a2537edee5444348a68e3e056e92299485585","text":"<!-- Please only use this template for submitting reports about continuously failing tests or jobs in Kubernetes CI --https://prow.ppc64le- Below two test cases started failing on ppc64le: 6th June 2021 https://k8s- ... was the delta between the passing and failing tests Started failing after\n/assign\n/sig cloud-provider /priority important-soon /triage accepted /assign\nThanks for reporting the issue. This was caused by vendor changes in , which introduced external dependency on IMDS. If IMDS endpoint is not reachable, then the new logic would report error.\nFiled to clean up the failed tests."}
{"_id":"q-en-kubernetes-e480cd539fe03293f3416a2ef2e904f0882d9b10e1de161bbf890a57f550eb85","text":"! no panic Kubernetes version (use ): k8s 1.16\n/sig Network\nCould you fix this? I think the function needs a lock while being executed.\n/triage accepted /good-first-issue todo: add a lock around the body of this function: Similar to\n/assign\nOh I see, thank you for letting me know, I'll keep that in mind going forward. :)\nAnyone actually on this? It seems like a good fix to have for 23 and maybe backport\nI'll have a try.\nThere is a reproducer here\nOops, I just peeked at the time of the pull request and not realized that you've got such deep conversation.\n/remove-good-first-issue /unassign\n/assign\nhi ! someone work on this now? I would like to try it.\nYes, working on it currently.\nare you still planning to give this a shot?\nIs there no one working on this? If so, I'd submit a PR for it this week. :)"}
{"_id":"q-en-kubernetes-f4e1a6c6f697968d9b1385c4b608da378c3871fbb1e708c25a58e14281a820d0","text":"/kind bug The logs of pods with label will be output expected without errors. create the following resources: kubectl logs -l Kubernetes version (use ): latest master.\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig cli\nI am able to pull logs from multiple containers in my test setup, this doesn't seem to be an issue with Could you check if all the containers in the specified pod are healthy?\nyes, all of these pods' containers are healthy. I the section.\nThis is working as expected, was not able to find the default container which in both deployments you explicitly said is but only the first deployment has container, the second deployment has . /close\nClosing this issue. <detailsIn response to : This is working as expected, was not able to find the default container which in both deployments you explicitly said is but only the first deployment has container, the second deployment has . /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-89d12b3e614d591d49449de1865679e8d74d20ff09a3b57edd7c4a325df85384","text":"we have one in k/k and one in k/utils, both need to be in LICENSES /assign\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/area code-organization\nI think we should include anything matching the LICENSE/COPYRIGHT/COPYING patthers under or subdirs in k/k or dependencies... a quick sweep showed these:\n/sig release ?"}
{"_id":"q-en-kubernetes-51f0fa90dfa98532fcd3c75518a568ceafa5b74ad593b3f097fa44f41abf8a67","text":"<!-- Please only use this template for submitting reports about flaky tests or jobs (pass or fail with no underlying change in code) in Kubernetes CI -- [sig-node] Pods should delete a collection of pods [Conformance] TBD links to appreciated links to specific failures in spyglass appreciated <!-- Please see the deflaking doc () for more guidance! --\ncc /sig windows\nI've looked through the logs for a bit and the test itself. What the test is doing: create 3 pods. expect to find 3 pods (it doesn't seem to wait for the to start, they only need to exist). delete all 3 of them as a collection. expect all 3 pods to dissapear. All 3 pods ended up on the same node, plus some other tests were running in parallel, which could also contribute to the failure. Looking at the kubelet logs, it seems that was the remaining pod to be deleted. Filtering the logs, we can see: The test failed at, while the other 2 pods did: Those times correspond with the number of pods observed by the test: Now, my question is, why are there such big gaps between the ` events? It's 12s for test-pod-2 and 21s for test-pod-3? Seems like a long time for pods / containers to linger after they've been killed. PS: It seems that Containers are still being created / started / stopped, even if the delete request came less than 0.2s after their creation. It doesn't seem to me like a reasonable thing to do, especially if that Pod would have to pull a huge image; I would have expected to be able to cancel that.\n/triage accepted\nLet's discuss at the 8/17 SIG-Windows community meeting\nHit again on 8/15\nI think was planning on opening up a PR to use longer timeouts in this test. We see that some Windows containers do not get scheduled until 1+ minutes into the test - especially when there are many tests running in parallel."}
{"_id":"q-en-kubernetes-80f52573d3f9c5e19a795878fe8592e9750799743d0d16b8b60fcbb7a7858fe9","text":"I am using client-go/tools/cache package NewInformer, find the StreamWatcher and seeing extremely high memory allocations. Attached is a memory profile: From this we have 2.3 GB of memory allocated by the StreamWatcher. To get to this state I had a script to create pod and delete pod over and over again, and every 10 seconds I run it again. I left this running 30 minutes and saw this memory usage. watch code as follows, cache.NewInformer return controller and storewriter is chan queue, receiving event and distribute to queue\n/sig api-machinery\nThanks for taking a look\n/help\nThis request has been marked as needing help from a contributor. Please ensure the request meets the requirements listed . If this request no longer meets these requirements, the label can be removed by commenting with the command. <detailsIn response to : /help Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/cc\ni find that infomer handler be stuck will lead to streamWatcher memory high allocationsdon't hang informer handler and you will resolve that question.\n/assign /triage accepted\ncan you check with a cpu profile if some handler is blocking?\ncpu profile is normalas followscode callback handler func (w *Watcher) AddEvent(obj interface{}) { data := w.genSyncData(obj, action.SyncDataActionAdd) if data == nil { return } w.writer.Sync(data) } writer is a buffer chan with length and write chan timeout for one secondprobably is will be stuck when produce lots of eventsfinallyI store events to fifo queue when it produce lots of eventsand consume fifo queue event to writer chanthe issue is solved\nSo, the problem is that your handlers are slower processing the objects from the delta fifo queue, than the rate at which the events are obtained from the apiserver, this is because of this objects are processed by the handlers locking the deltaFifo queue, hence the streamwatcher accumulates the object in memory until the handlers process the object and Pop() release the lock. The code comments inform about this problem // Add() and Get(), won't be blocked for too long. the only solution I see is logging some errors warning about this (processing withouth locking doesn't seem an option, right?) I've tried to use queue depth, but that is not an option, since the contention happens before events get into the queue and it nevers goes high. I've tried adding trace for slow handlers operations, however, those maybe harmless if the incoming rate is still lower than the draining rate, unless we decide that we want to trace process handlers that take more than X seconds Is adding another intermediate queue between the streamwatecher and the deltaFifo an option? that will free the streamwatchers ...\nYeah. I don't see it as a high priority, but that could be helpful. Following the standard hander to workqueue pattern, it isn't a practical problem. /priority backlog\ni don't hang informer hander by fifo queue, i find that memory will come down when lots of create and delete podsbut i running code onlinethe streamwatcher still exist high memory usage as follows pprof Please take a look at it\ni am also facing same issue any guidance for the watcher memory increase in high pod churn"}
{"_id":"q-en-kubernetes-6a2a5fc85d2ca69a1acd133413bb132a47ae1875a7bbd20ed79fc4ab60a0201a","text":"I can't find a single instance of this test passing in testgrid. Run via and potentially other suites. /sig node /kind failing-test\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign\nThe test fails because there is no GPUs reported by kubelet [1]. Looking at the node info right before the failure, the node does not have any gpu capacity [2]. The artifacts only have kubelet log, but not device plugin or driver installer log. So I am not sure if there's any error on them. Is there any way to see the device plugin or driver installer log? [1] [2]\n/assign I have a partial fix here - The actual feature seems broken too tho - hidden by the fact that it's been failing to provision properly for a really long time :(\nSince merged this test has been relatively stable. Going to close this issue now (kept it around while I was on PTO to take a look when I got back). Seems like there are some infrequent flakes that we should follow up on after we remove DKK. /close\nClosing this issue. <detailsIn response to : Since merged this test has been relatively stable. Going to close this issue now (kept it around while I was on PTO to take a look when I got back). Seems like there are some infrequent flakes that we should follow up on after we remove DKK. Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/reopen This is failing again, I'm not sure what regressed but it was 2wks ago.\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nI keep seeing errors like But our serial suite doesn't have GPUs attached. I'd expect this to be skipped: do you have any idea why sometimes these tests pass, sometimes they get skipped, and sometimes they time out? It makes no sense to me:\nWrote to try to dump what's matching and not causing these tests to get skipped\nThe tests get skipped when they run on nodes without GPUs attached, which is 2 of 3 runs every time, but cos-stable2 has 2 GPUs (). The failures happen because something in the GPU setup process broke again (I assume a bad image bump bc we don't pin), and so the test fails waiting for the GPUs to show up on the node after installing the plugin, but they are attached to the node.\nI think I have a fix here -"}
{"_id":"q-en-kubernetes-c919c01290a8407c507502b9e4a455203a30df2d5123aee80e357bf8c1b37ac8","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --On cgroup v1 when using systemd cgroup driver, kubelet regularly freeze and thaw its control groups, and that is causing issues further down. No freeze Running the command below should never print anything, since should always be . However, when using master/v1.22.0, this returns regularly. When using runc v1.0.0 or earlier, this can cause containers to end in a permanently frozen state. This state is tricky to debug, since no log output is given. To see if an environment is having problems, one can use: More info in and Kubernetes version (use ): : introduced in Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: Network plugin and version (if this is a network-related bug): Others:\n/sig node /priority critical-urgent\nThanks for opening this. IIUC, this issue is limited to: cgroup v1 systemd cgroup driver runc 1.0.1+\nsystemd cgroup driver runc 1.0.1+ Well, yeah, kinda. It is a bit of a matrix As long as cgroup version = v1 cgroup driver = systemd the freezing will occur. The freezing is unnecessary, and creates a big overhead on machines with huge number of processes. However, if runc (the binary used by the container runtime) is v1.0.0 or before, the permanent freeze of containers can occur. If using runc (the binary) v1.0.1, the permanent freeze is mitigated and does not occur. So in one way, as long as the user is using runc binary v1.0.1 or later, this is not completely critical.\nI am seeing a lot of flakiness lately in systemd node e2e jobs using crio, Almost all of them fail with container not being able to come up with following error, That seems to be coming from I am wondering if that flakiness is due to frozen cgroups issue reported in this issue. Unfortunately I do no have access to the system to query sysfs to see if cgroups is in FROZEN state.\nI think is right and the flakes we're seeing in the CRI-O jobs seems to be related. runc considers a frozen cgroup state as paused container: This means that in our test, the container creation works as intended, but the fails sometimes which causes test to flake and container to get stuck in the container creation error. The code in runc seems intentional and exists since a long time, are we aware of any recent changes in k/k to this kind of cgroup handling?\ncc\nYes that looks like a result of this issue. Thanks for the detailed information! This was issue in k8s (the freezing) was introduced in As disucssed here: , freezing would cause issues, and the runc bump was merged before it was properly tested. If you are talking about runc as in the binary, there was previously an issue with the definition of a \"paused/frozen\" container; should give some more context. That issue is fixed in runc v1.0.1 (the binary).\n/triage accepted\nThis issue will be fixed by or any similar way to avoid the freeze.\ncc:"}
{"_id":"q-en-kubernetes-8b27ed3ea13cd578cc7da2993f0c4dc00551de50346597cfa85e2c063e948d9e","text":"<!-- Please only use this template for submitting reports about continuously failing tests or jobs in Kubernetes CI -- - At least 08/08 sequentially performs operations that take ~20s on average. Example: For a 5000 nodes test this causes a major timeout. excludes tests with tag which SCTP had until was merged. This is why there is a shift in red cells in the testgrid after 08/15. We will add a pattern excluding until this issue is resolved. /assign /cc\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig network\n/assign"}
{"_id":"q-en-kubernetes-3df4fc7426c9edb6203d9dbddb373b0b84cb8c7ea70cfee5c9908abf2aaafe95","text":"<!-- Please use this template while reporting a bug and provide as much info as possible. Not doing so may result in your bug not being addressed in a timely manner. Thanks! If the matter is security related, please disclose it privately via --E2E test fails. The test is implemented at: I would expect the test pass with these objects in place: AFAIK this fulfills the general requirement of the test that there's a corresponding endpointslice object with same addressing information as is on the endpoint object. IMO the test should rather look for epslice which is owned by the endpoint rather than relying on a hardcoded name. Add flag to kube-apiserver Manually create the endpoint with proper addressing for the api-server epslice \"mirroring\" create new endpointslice with name like --The test fails as it's hardcoded to look for endpointslice named I'm working on a PR to provide a fix proposal Kubernetes version (use ): 1.22.1 Cloud provider or hardware configuration: OS (e.g: ): Kernel (e.g. ): Install tools: k0s Network plugin and version (if this is a network-related bug): kube-router 1.3.1 Others:\n/sig testing\n/remove-sig testing /sig network EndpointSlices are owned by sig network.\n/kind failing-test\n/remove kind-bug /remove kind-failing-test The test is correct, apiserver must generates both endpoint and enpointslices with the special name \"kubernetes\" , that are the backends for the \"kubernetes.default\" service with the IP addresses of the apiservers. This special service \"kubernetes.default\" and its endpoints are used by clients with In-Cluster auto discovery. I think that whatever cluster that is not generating the endpoint and endpointslices for the \"kubernetes.default\" service should not meet Conformance, at least I can't see how automatic in-cluster configuration will work without that ... I only see option used for webhook tests, the other 2 reconcilers always create the endpoints and slices for \"kubernetes.default\" should we clarify the goal of this \"none\" option?\nYes, I do understand this. And in our case this is true, there is a svc called as there is also endpoints object called . This makes the in-cluster discovery work. Otherwise we'd get much more errors in the conformance instead of this only one... :) In our case we're relying on the endpointslice mirroring to take care of creating the endpointslice object, which makes this test fail as it uses generated \"suffix\" on the object name. I have not dived too deep (yet) into the internals of endpointslices but I fail to see how the endpointslice object name really matters when both svc and endpoint object does have proper name.\noh, this seems interesting then ... who is creating the corresponding endpoint? can you paste a kubectl get endpoints -o yaml?\nthis is how it looks like in my cluster\nIn our case (k0s distro) it's our custom controller that is triggered when the controlplane is configured in HA mode. So this basically sets the api to mode . When our controller runs it has a configured LB address which it resolves to IP address(es) and creates the corresponding EP object. We do NOT create the endpointslice as we rely on the mirroring functionality to let it create and sync it. Here's the object details: Reading the code for API built-in endpoint reconcilers I see they set and create the endpointslice object by themselves. Yes, we could do that too in k0s but I fail to see why the name of the endpointslice object really matters at all.\nI agree the conformance test should not be looking up endpointslice by name. It should be using a label selector of to locate one or more endpointslices and verifying the ports and endpoints across them.\nFair enough, on the good side this also will simplify the KEP from to dual-stack the apiserver you are the expert in this area, can you PTAL at this discussion? /kind failing-test\nhmm the test description explicitly says the name of the object has to be \"kubernetes\", if the name is mandatory the test is correct, otherwise we need to rewrite the test description and talk with conformance people /area conformance\nthe description should be rewritten... the test is enforcing stricter requirements on endpointslices than required/expected\n/triage accepted"}
{"_id":"q-en-kubernetes-2b197cbb727a9bbc0a905a9e9fedea145198ea12133cbfb726dba6d0aba5de31","text":"TestRecordOperation TestRecordOperation TestGetHistogramVecFromGatherer /kind failing-test <!-- If this is a flake, please add: /kind flake --<!-- Please assign a SIG using: /sig SIG-NAME --/sig testing This may be related with \n/assign I will do some initial researches on this. is similar\n/sig instrument\nThe label(s) cannot be applied, because the repository doesn't have them. <detailsIn response to : /sig instrument Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig instrumentation\n/triage accepted /assign cc\nTo fix it gracefully, we may refactor the code in like The refactor would change a lot and need some confirmations or demo.(I prefer) A temporary workaround before refactoring it: is another try to ignore the dup error in but it is odd as should panic as designed if one register failed.\nThree testings failed due to . Haven't found package promauto () being used in these testings. It might be other issues. These testings are: /filterwithonelabel\nWhat are you guys doing here? What test is this? You cannot change the component-base metrics code base it is doing what it is supposed to be doing. Does this test suite involve multiple sequential runs of the same test?\nI think this is odd to change the component-base metrics code base. But the refactor to use is another approach. You can try to reproduce it with is with\nOh, yes. This involves multiple sequential runs of the same test to . It uses \nYou can use to register the registry only once. It is important to reset the metrics, or next iteration in the test will fail\nThanks I other fixes in"}
{"_id":"q-en-kubernetes-bca617990d37a6b01f8197b681d6cb5b71d56ca9865bda160f154ac0d000a5c8","text":"<!-- Please only use this template for submitting reports about continuously failing tests or jobs in Kubernetes CI --- post-kubernetes-push-image-etcd n/a 2020-08-11 Started failing when merged No new versions of etcd are getting pushed until this gets fixed, when means the following PRs haven't fully taken effect: /sig api-machinery /priority critical-urgent /milestone v1.23 /assign FYI\n/sig release /area release-eng\nI've opened to rollback the offending commit, would like to keep this open until we can figure out the correct fix-forward\n/triage accepted\nwas triggered by the merge of , which pushed So that's now current / includes whatever PRs linked in the description may have tried to change (excluding )\ndoes this need promotion at all or are we good with the existing image?\n/remove-priority critical-urgent /priority important-soon The job is back to passing\nRecent updates are related to Debian base image update and golang upgrade. Nice to get a promotion. Good enough to use .\nYeah, looks like that the env in the build job does not search for docker CLI plugins in the path, which makes it unavailable during the run.\nI think we have to set the environment variable like we already do in other build jobs, proposing the fix in\nLatest test run is green:"}
{"_id":"q-en-kubernetes-454f87a4fe9732af2c94213d170ad3df9bf34854cac77e18869bf1bcd14a2783","text":"I am trying out For that I enabled kubelet memory manager, cpu manager and topology manager. This is tried on a dual socket NUMA server KUBELETKUBEADMARGS=\"--container-runtime=remote --container-runtime- --cpu-manager-policy=static --feature-gates=RotateKubeletServerCertificate=true,AllAlpha=false,RemoveSelfLink=false,EphemeralContainers=true,BoundServiceAccountTokenVolume=false,KubeletPodResourcesGetAllocatable=true,TopologyManager=true,MemoryManager=true --kube-api-burst=40 --kube-api-qps=20 --kube-reserved=cpu=4,memory=4Gi --memory-manager-policy=static --reserved-cpus=0,1,2,3 --reserved-memory 0:memory=4Gi,1:memory=4196Mi --system-reserved=cpu=4,memory=4Gi --topology-manager-policy=single-numa-node\" topologyManagerPolicy: single-numa-node is set on This throws an error 2021-10-20T11:28:49.+00:00 localhost kubelet[]: E1020 11:28:49. ] \"Failed to parse kubelet flag\" err=\"invalid argument \"0:memory=4Gi,1:memory=4196Mi\" for \"--reserved-memory\" flag: the reserved memory has incorrect format, expected numaNodeID:type=quantity[,type=quantity...], got 0:memory=4Gi,1:memory=4196Mi\" This is on kubernetes 1.21.1 Client Version: version.Info{Major:\"1\", Minor:\"21\", GitVersion:\"v1.21.1\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2021-09-29T11:36:35Z\", GoVersion:\"go1.16.4\", Compiler:\"gc\", Platform:\"linux/amd64\"} Any hints on specifying \"reserved-memory\" for multi-socket NUMA system ???? One way is specifying multiple \"reserved-memory\" like But problem with this is is map, so it will get over-written. we should be able to specify reserved memory in a multi socket system. Enable above flags and run kubelet No response Client Version: version.Info{Major:\"1\", Minor:\"21\", GitVersion:\"v1.21.1\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2021-09-29T11:36:35Z\", GoVersion:\"go1.16.4\", Compiler:\"gc\", Platform:\"linux/amd64\"} none SLES 15 sp2 <details</detailscontainerd No response\n/sig node\nthis is the right way to config\nProblem with that is KubeletExtraArgs map[string]string is a map, so multiple \"reserved-memory\" will get over-written.\nthere is a tracking issue for kubeadm's api to support multiple arguments with the same name: but it is a backwards incompatible change. you should beusing the KubeletConfiguration for this instead:\n/remove-kind support /kind bug /assign Looks like a bug in a parsing logic + better error would be beneficial\nIs this a parsing bug?, if yes can you just point me where I need to fix this. I will try to fix this in the downstream release(our internal release).\nHere the parsing logic - and here it is failing - Give me know if you need some additional help.\nAre we planning to add support for multiple numa nodes in single \"--reserved-memory\" flag, or just displaying a better error msg?\nWe should support multiple NUMA nodes under the same --reserved-memory flag.\n/triage accepted\nCan you provide a sample configuration for 2 numa nodes? \".\" comma seperator is already there for single numa configuration.\nYes, it is the problem, initially we needed to split it by comma, but probably it will not work as expected, so maybe we can split it by and update documentation. Something like:\ncan you please review the PR and provide your comments"}
{"_id":"q-en-kubernetes-aab3f40b9cf67cd6fce004e1473a0bb34135092c8f2b45880ecec649bab66086","text":"configmap, like edit configmap test, and make some error and write-save, will see below when the content after edited is invalid, should not remove data , thus mean expect content below configmap 2 edit configmap make invalid No response version kubeadm <details</details<details</details<details</details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig cli\nPTAL\nHi, I am new to this project and would love to contribute. Can I work on this issue?\n/sig cli"}
{"_id":"q-en-kubernetes-62d56f25b8a5d63e3bd63af623b0c410a88680e0f672a94810c48534f01d3874","text":"doesn't do the same embedded resource and x-kubernetes-list-type \"map\" and \"set\" invariant validation that happens in custom resource create/update validation. Does that mean embedded objects and list-type fields under custom resource status can contain invalid data? if so, that needs to be fixed in a way that tolerates existing invalid data (ratcheting validation) Originally posted by in /sig api-machinery /kind bug /priority important-longterm\n/triage accepted\nI would like to collaborate for fixing the validation issue for custom resources and tolerate the existing invalid data. Could you please guide me with some code pointers or files and elaborate on how can I proceed for fixing this bug.\nI haven't tested this, but something along the lines of this: we'd need tests to exercise the following scenarios on a CRD that specifies a status subresource with a list property with x-kubernetes-list-type=map and x-kubernetes-list-type=set a custom resource instance that already has duplicate items in the status property list can be updated successfully while still having duplicate items in the list (validation doesn't block an invalid → invalid status update) a custom resource instance that already has duplicate items in the status property list can be updated successfully to set unique items in the list (validation doesn't block an invalid → valid status update) a custom resource instance that does not have duplicate items in the status property list cannot be updated to set duplicate items in the list (validation blocks a valid → invalid status update)\nsince it's been a couple of months, i'm going to try and take a stab at this /assign\n/assign\nHi k8s 1.25 Bug Triage Shadow here, just checking in to see if this is still on track for k8s 1.25? Thank you.\nhas stalled needing tests and a rebase\nThanks for the update.\nHello with the code freeze approaching, is this still on track for v1.25? Thanks!\nnot really, the status on is the same (needs tests)\n/milestone clear"}
{"_id":"q-en-kubernetes-1fef79d7f3d7f993d523588b0a756a73084a361bc5470b14efa020a39af7d433","text":"pull-kubernetes-unit TestDifferentWidths As far back as triage board history shows No response /sig api-machinery\n/cc /triage accepted\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\nThe Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle rotten [1]:\nThe Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Reopen this issue or PR with Mark this issue or PR as fresh with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /close [1]:\nClosing this issue. <detailsIn response to : The Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closedReopen this issue or PR with Mark this issue or PR as fresh with Offer to help out with [Issue Triage][1] Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nRunning into this one a fair bit in my PRs, opened as a possible fix\n/reopen\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThe Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. This bot triages issues according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Reopen this issue with Mark this issue as fresh with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /close not-planned [1]:\nClosing this issue, marking it as \"Not Planned\". <detailsIn response to : The Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closedReopen this issue with Mark this issue as fresh with Offer to help out with [Issue Triage][1] Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/reopen I got this in --\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThe Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. This bot triages issues according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Reopen this issue with Mark this issue as fresh with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /close not-planned [1]:\nClosing this issue, marking it as \"Not Planned\". <detailsIn response to : The Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closedReopen this issue with Mark this issue as fresh with Offer to help out with [Issue Triage][1] Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/reopen Got this in\nReopened this issue. <detailsIn response to : /reopen Got this in </details\n/remove-lifecycle rotten"}
{"_id":"q-en-kubernetes-2aa0bd68651ec5eb919e1fbaa13d3d2a79722d86491f6bccc89317cd11d15c7a","text":"A/D controller tries to detach volumes that are still mounted after 6 minutes. The conditions to trigger the detach are: I.e. all containers are finished. But that does not guarantee that volumes are unmounted. When a CSI driver can't unmount a volume for 6 minutes, the volume is force detached by A/D controller, corrupting the volume and leaving volume mounts on the node in a bad state. Note that the node is perfectly healthy, it just can't unmount a volume from whatever reason. A/D controller detaches volumes after a Pod is deleted from the API server, i.e. the volumes are unmounted or someone force-deleted the pods. Very artificial reproducer: a Pod with an attachable CSI volume. the node, make the volume mount busy, e.g.: Every unmount of the volume fails and kubelet/the CSI driver periodically retry. the pod 6 minutes, see the volume is detached without unmounting first. KCM logs: Note that the Pod is still in the API server: And the node is healthy: Note that some CSI drivers (e.g. the cloud ones) do not allow detaching a volume while it's still mounted. Not all CSI driver have such possibility and they rely on Kubernetes not to call ControllerUnpublish before NodeUnpublish / NodeUnstage succeeds. <detailsToday's master </details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig storage /priority important-longterm\nRight now, a volume is detached from a node when: (all pods that use the volume are deleted from the API server OR all their container are stopped) AND (the volume is unmounted OR 6 minutes passed). Potential solution 1: Issue detach of a mounted volume only when all pods are force-deleted from the API server. More formally: Detach when [(All pods that use the volume are deleted from the API server OR all their container are stopped) AND the volume is unmounted] OR [all pods that use the volume are deleted from the API server AND 6 minutes passed]. Is this change subtle enough that it does not break existing behavior (e.g. tooling on top of Kubernetes)?\nPotential solution 2: Add a new flag to CSIDriver, say . We would need to extend internal Detach() call to include a flag if the volume to detach is mounted or not (A/D controller knows that) and the CSI volume plugin would check CSIDriver and potentially reject detach of a mounted volume. This way, it would not be possible to ever force detach a mounted volume, even when a pod is force-deleted! Potential solution 3: As a variant, we could add a cmdline option to the external-attacher, like . When that cmdline option is specified, the external attacher would check , if the volume to detach is still mounted, and refuse such detach (with a reasonable error message). CSI driver vendors could opt-in into this behavior. This way, it would not be possible to ever force detach a mounted volume, even when a pod is force-deleted! Potential solution 4: As a variant, we could add to CSIDriver, where each driver could put their own timeout before force-detaching a volume, with an option to use infinity.\nNote to self: check , it might help on the CSI side.\nSolution 3 will not work well with ungraceful node shutdown - attach/detach controller does not watch pods and force-deleting pods is the main point of ungraceful node shutdown KEP. /me will try if solution 1 would work when NodeStage times out (is Pod deleted from the API server before or after NodeUnstage is called?)\nwas right, a pod is deleted (without force) from the API server after NodeUnpublish succeeds, it does not wait for NodeUnstage to finish. Therefore we cannot distinguish if a pod was force-deleted (and we should start 6 minute timer) or pod was deleted without any force and only NodeUnstage is taking more time than expected. Solution 1 is then not possible.\nNow we are down to solution 2 and 4. For solution 2, if driver does not have a way to check whether volume is mounted or not, the behavior will not change and we still force detach after 6 mins. And it might happen in the middle of unstage ..\nSumming up yesterday's call: A/D controller should check if a node is unhealthy before force detaching volumes from it. It should never force detach a volume from a healthy node, not even after 6 minutes. Question what healthy means is open and we should talk to sig-node. is a working assumption.\nI agree with both of these. I'd just emphasize that that particular definition of unhealthy definitely includes perfectly healthy nodes in a network partition. This change is no worse than the status quo ante for those nodes, and a definite improvement on the healthy nodes, so I wholeheartedly support it.\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\nThe Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle rotten [1]:\n/remove-lifecycle rotten\n/assign"}
{"_id":"q-en-kubernetes-8e10ff4e574cb7fe88a073310bd5305e388b18103a1b5863b3d5a524a66d25f7","text":"It is unclear why the test needs the dynamic kubelet config. Need to clean it up. /sig node /priority important-soon /help Test is not a part of any testgrid. Need to understand why and return it back to the grid. Also it is marked as flaky. Need to be deflaked. unknown No response No response No response_ /sig node\nThis request has been marked as needing help from a contributor. Please ensure that the issue body includes answers to the following questions: Why are we solving this issue? To address this issue, are there any code changes? If there are code changes, what needs to be done in the code and what places can the assignee treat as reference points? Does this issue have zero to low barrier of entry? How can the assignee reach out to you for help? For more details on the requirements of such an issue, please see and ensure that they are met. If this request no longer meets these requirements, the label can be removed by commenting with the command. <detailsIn response to : ### Which jobs are failing? /priority important-soon /help ### Which tests are failing? Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nHi, I am new to this project and would really love to start contributing and thought this might be a good start. Could you tell me a little more about this problem?\nglad to have you, welcome! Some step by step you may start with: Find the test mentioned, understand what it is doing If dependency on DynamicKubeletConfig is found, try to work around them and remove that line. Try running the test locally. If you can repro flakiness - try to fix it If test passes: Try removing the and see if passes in serial tests on presubmits\nThank you so much! /assign\n/cc\nit had a non-op config change in the past - - I potentially missed this when removing that\nso, if I'd like to fix it, what should I do?\nsee Since we understood item 2 - it is not needed, just remove the skipper lines. Ideally test needs to be de-flaked as well, but it it not necessarily in the scope of this change\nthis is a simple issue, so I tried to fix it since I've found it some days."}
{"_id":"q-en-kubernetes-69a1f590530faecad27a74254209f31c80bb63b4a1bada6b4e285c65d544480e","text":"I have a cronjob that runs a Python script every 5 minutes. After the cronjob finishes running, I start seeing a bunch of log messages in my systemd journal like the following: Those directories in /var/log don't exist as they belong to old pods that have finished running. These messages keep increasing in volume as time passes (almost received a 1000 of those messages at one time, after which I had to reboot the node to fix the issue). It almost looks like the kubelet is not cleaning up the old pods for some reason. I can't see these old pods when I run , so I'm assuming their is some issue with the kubelet's cleanup process. kubelet (or cadvisor) shouldn't try to read logs for pods that don't exist. Setup a short lived cronjob running every 5 minutes and observe the logs (Kubernetes v1.22.4, CRI-O v1.22.1, Ubuntu 20.04) No response <details</detailsBaremetal cluster <details</details<detailskubeadm (baremetal cluster) </details<detailsCRI-O v1.22.1 </details<details</details\n/sig node\n/assign for triage\nI have something similar on a fresh cluster I use for educational purposes using kubernetes 22.5 with cri-o. Here's an excerpt: This is from the master node. Nodes run Debian bulsseye in KVM provisioned with Vagrant. I have not yet tested it out with different versions. If I can assist in testing, I'd be happy to do so!\n/triage accepted /priority important-longterm\nI've produced the same issue on my clusters. I'm running Debian 11 on bare metal (homelab) and in VM (test). My clusters are installed using kubeadm, and I'm using multus and calico as CNI. The issue is triggered as soon as a pod is removed. I've triggered it by deleting a simple Pod, scaling down a Deployment, or running a CronJob. On the node where the pod was removed, the logs are then flooded by messages, at the rate of 1 line per pod removed every 10 seconds. I've seen it both on aregular worker nodes, and untainted master nodes. I've tested under Vagrant with various K8S / CRI-O versions. With CRI-O , I never triggered the issue with Kubernetes : , and . However, with CRI-O = , all my tests triggered the issue : Kubernetes , CRI-O Kubernetes , CRI-O Kubernetes , CRI-O Let me know if I can help by running other tests or capturing specific logs.\nIs this only issue with CRI-O? What about containerd? The reports above seem to all be on CRI-O? /cc\nopenshift seems to have hit this case a couple of times and digging into some bugs leads me to think it will be fixed with and kube 1.24. can you try one of the kube alpha releases? cri-o 1.23.1 should do, or you can try the main branch\nit's possible it's unrelated, but is why I am making this guess\nI'm used to installing a cluster using Debian packages, so it took me a few tries to have a working v1.24 cluster. I followed the \"Without a package manager\" to install from the binaries. I used Kubernetes and CRI-O . I'm not certain everything is working fine, but I can schedule a Pod / Deployment / Job, so I assume it's good enough to test this issue. And I still reproduce the issue as soon as I remove a Pod.\nI'm also hitting the issue in my lab environment(Vagrant). Running Rocky Linux 8.5 Kubernetes 1.22.6 and CRI-O 1.22.2\nah hah! I found a fix (and found out why it's only being hit in cri-o and not cadvisor)\nWe're also having this issue on a cluster hosting Gitlab CI jobs, this is quite a trouble since many jobs are spawned and removed daily. Kubernetes version: v1.22.7 CRI-O version: 1.22.2 Is there any news about the fix ? I looked into 's PR, it seemed ok to me but I'm not really experienced in Go... Nevertheless I'd be happy to help\nup\nI can confirm on a bare metal cluster that I'm running on containerd that I am also seeing these log messages.\nObserving similar issues on (k8s and crio ) and (k8s and crio ). The strange thing is that one cluster just some nodes affected while the other, all of them. Below an overview of 30m occurrences of Cluster running version : Cluster running version : No idea where the discrepancy is coming from.\nTested with and on a fresh cluster bootstraped with kubeadm, all machines run on Debian 11. Ran a job 3 times and there's a log line for every execution, so it spams 3 log lines per kubelet loop.\nIt looks like cgroups are left behind. Removing them fixes the logspam (temporarily). Can use something like this to find all of them, depending on actual error in journal: Then iterate over the and remove them, while checking that running pods are not considered: It's quick and dirty, but does the job. I've deployed a garbage collector daemonset with the above and until this is properly addressed, I'm quite happy with the fix. The problem is, for a medium to large size cluster with pods coming and going, it's very easy to bust your elasticsearch storage and performance with this log spam.\nYesterday we tested with the newly released version of and the problem was still here. After reading answer I digged a little bit because we use systemd to manage cgroups and it seemed the error was on our side. I updated the kernel boot command in the grub file because in order to have cgroupsv2 unified behavior in systemd you have to set both options: and and we only had : Used the following command for the changes to take effect: And then rebooted the node. At first it seemed like this did the trick as I popped a job on the node I changed and the kubelet didn't spam the logfile anymore. As soon as I deleted the pod from the cluster the kubelet started spamming the logs again with: So this did not resolve the issue. This was tested on a node running , , and .\nLooks like the cgroup slices not being cleaned up by systemd is an old issue:\nwould you mind giving a try to see if it helps?\nwhile your fix does make sense, the root cause (which atm is somewhat unknown) is not really addressed. Something's leaking cgroup slices... and have both been closed off.\nI am not sure if they're leaking as much as have yet to be cleaned up. Have those pods been deleted by the kubelet yet? and by deleted, I mean fully removed (no traces if searched by ). The kubelet will clean up the pod scope when GC kicks in, but will keep them around until then to allow for forensics and inspection to be performed. Since your script uses to find the scopes, I am guessing the pod themselves have yet to be GC'd, and should still exist. Which is why I think that is the correct solution: we expect those scopes to be there, but shouldn't unnecessarily log an error if a pod that's stopped but not yet removed is found\nto verify my suspicion, I would be interested in seeing for all the pods that are showing up in your script, as well as to cross reference pods to containers, if you wouldn't mind uploading that somewhere.\nI'm sure they're not in crio anymore. I'm only using to ensure I'm never going to touch a scope that's reported under a pod still seen by crio. I'll scale down the DaemonSet on my test cluster, run some workload to generate the leaking pods and then provide you some log snippets and requested output from crictl.\nI took one of my clusters that's suffering heavily on this issue (hosting gitlab CI jobs), reactivated the logs on it for a second, doesn't take long for the kubelet to log things like this: Trying to inspect the pod with returns: I tested with a few pods and the result is always the same. This seems logic as gives me the following: But the slices are still here tho: So in my opinion crictl return the right things, it's either the kubelet that doesn't update correctly the list of pods to watch or the systemd problem with the slices.\nI've attached the logs and requested outputs here (sanitized hostnames). Let me know if you'd need anything else; for now I'll leave them around and not scale up the DS on the test cluster.\nI've posted a quick gist with all the manifests I'm using to deploy a garbage collector daemon set. Nothing fancy, just a bit of patchwork to address the issues I was having.\nHere's another really interesting bit of information. In my case, this issue seems to be only affecting VMs and not bare-metal nodes. It's completely puzzling, but I have 12 bare-metal nodes and 1 VM node in one cluster (the VM is just for testing) and it manifests only on the VM node. As well, my test cluster is made up fully of VMs (ie no bare-metal) and there it happens across the board. The bare metal nodes where the VMs are running are the same are same hardware as the bare-metal nodes of my main cluster (just more NVMe drives). Edit: amended to clarify this is observed in my case running OKD.\nSorry but I have the same issue on bare-metal nodes (x86 ArchLinux kernel 5.17.5 with K8S 1.23.6/Crio 1.23.2)\nThe issue in title happens on bare metal nodes too.\nI've got the same issue on bare-metal nodes(k8s v1.23.0/cri-o v1.22.3) and I'm not sure if it's for the same reason, but I have this problem as well.\nI figured out the reason for the discrepancy. The VM node was using cgroups v2, while the bare-metal, older nodes, had the kernel parameter . Removing that and rebooting, triggers the problem.\nDid some testing and I can confirm that switching from CRI-O to containerd has stopped these log messages from popping up all the time. For now, this seems like a reasonable solution, at least for me. Using Cgroups v2 Kubernetes Version: v1.23.7 OS: Ubuntu 22.04 Containerd Version: v1.6.4 Runc Version: 1.1.1 EDIT: Kubernetes Version was v1.23, not v1.24\nAnother option is one could switch to the CRI stats provider by setting kubelet's containerruntimeendpoint to instead of\nIs this also an option for OKD/OCP? Is there any progress on a fix or should we look at the above workaround instead?\nOut of curiosity, why does this work? /cc\nThis would work for OKD, but wouldn't be supported for OCP OOTB. should fix, but it needs a reviewer/approver... Fun kubernetes fact! The CRI-O team found performance issues with the CRI stats provider, so there's a hardcoded check to see if the socket is . However, since is symlinked to , and the check is a string comparison, one can use and not trigger that toggle, turning on CRI stats provider. We're working on fixing the perf troubles with CRI stats KEP 2371\nUsing doesn't solve the problem on my environment, we're still seeing the same error logs. We're running: Flatcar 3033.2.4 Kubelet 1.22.4 CRI-O 1.22.5 cgroup v2 enabled.\nI had this error on Kubernetes version 1.24, with CRI-O 1.24. Today I did a fresh install with Kubernetes 1.25 + CRI-O 1.25 (with the same machine types and OS versions) and the error went away. I tried deployment scale-down, pod deletion, but I no longer see the mentioned \"Unable to fetch log stats\" messages in the log. VirtualBox machines OS: RHEL 9 with kernel 5.14.0-70.22.1.el90.x8664 Kubernetes version: 1.25.0 CRI-O version: 1.25.0 Runc version 1.1.3 Using cgroups v2\nWe are running kubernetes 1.24.9 with cri-o 1.24.4 and runc 1.1.4 on Ubuntu 22.04 with cgroups v2 and experienced the same issue. For some time I was running a deployment with a dozen of pods and rotated them every 10 seconds to create a lot of garbage in cgroups. This caused rather slow increase in both mem & cpu usage and a lot of error messages like : Kubelet mem & cpu usage: ! ! After I stopped generating those pods the resource usage stopped. After running script from the CPU&mem usage dropped."}
{"_id":"q-en-kubernetes-b1e788704854242162ef75c028f95b1731d4922663decd69ec4a147fa76a7a9e","text":"PIDPressure is not set to true when the system has insufficient pid resource. PIDPressure is set to true correctly when the system is insufficient in pid resource. See cause below The max pid is not solely controlled by which is referred by . But it is also controlled by (whichever is smaller). However, they have different semantics resulting in the same constraint to how many processes/threads can be running on the system. <details</details<details</details<details</details<details</details<details</details<details</details\n/sig node\nI asked the same issue before,\nMy opinion: the minimum of and is the right number to count, for Linux. That will provide better instrumentation. If you're reserving a number of task IDs (PIDs) for system use, that minimum is the number that you're reserving from.\n/triage accepted /priority important-longterm\n/retitle System maximum PID inconsiderate of sysctl 'threads-max'\nBTW I made a thing that might help folks test a fix: Originally I used it to create PIDs, to test whether apps cope with long PIDs. This version keeps the threads around so that you see them in ."}
{"_id":"q-en-kubernetes-a9a4da968c099cbb8e97f58bc3ec7d8c320196c52b6c958ad307c5036b23408a","text":"Add more message for no PodSandbox container when kubelet start. When error return, We need more message to solve it, contains container name and container id.\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig node\n/assign"}
{"_id":"q-en-kubernetes-1e01fd7a18bf0262657c5f79bd6c99559287e592ed4eeab4a0c742d833179a6a","text":"I've been following the official docs for kubeadm and how to set up a new cluster. The cluster is up and running but I noticed warning events when creating node port services: However, the node port does work (I see the output when accessing node ip:nodeport). Might it be related to something else than Kubernetes? There is no such warning event. Set up a fresh cluster with kubeadm and then create a service of nodeport type. No response <details</details<detailsHetzner cloud </details<details</details<detailscloud-init </details<detailscontainerd </details<detailscni=cilium csi=hetzner </details\nThere are no sig labels on this issue. Please add an appropriate label by using one of the following commands: - Please see the for a listing of the SIGs, working groups, and committees available. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\ncheck with netstat or ss if there is a process listening on that port\nno, I specified exact random port e.g. and checked it before with ss - it was not being used. Then I created the service, the warning occured, and after that the port was taken:\nkube-proxy binds the NodePort to warn users that they may be shadowing a process, but it logs the errors if it can not bind the port\nI met the same error and I confirmed the node port was not used by any other process. After looking into the code, I think this is a bug in kube-proxy that it would try to open a node port twice with same (empty) IP and port when is empty. contains and in below code.: I have created to fix it.\nAwesome find. Did you figure out a work around? I am encountering this issue with this version of the image:\nThe error looked scaring but the nodePort should be working, as it was just caused by a duplicate port opening attempt. v1.23.5 and later releases have the fix."}
{"_id":"q-en-kubernetes-6a9e3833522bcc990c0dcf3596b97119de8d0d04e732a719d9ef25a6adaf3478","text":"kube-controller-manager yields whole stack trace when it loses leaderelection. kube-controller-manager to only log and to exit (without logging 17K lines of stack trace that is not useful for anyone). kube-controller-manager yields around 17K lines of stack trace (omitted in the above example). No response <detailsv1.20.12 </details<details</details<details</details<details</details<details</details<details</details\n/sig api-machinery\n/assign /triage accepted\nhas Fatalf always triggered a stack dump of all goroutines?\n/cc\nSeems like does a dump of all goroutines Could we and then maybe? /cc in case you have thoughts about this\nIf I read the right, it only printed the current goroutine to stderr. Dumping everything seems to have been later. make sense to me.\nDumping of all goroutines was in\nI have similar doubts. I want to fix it. /assign\nkube-scheduler seems to have a similar issue (but a shorter stack trace)\nWhen replacing (useful because it doesn't do structured logging, not just because of the stack trace problem), remember to use + the new . Updated migration instructions are pending in There's also - reverting the klog change might be a faster way to address this."}
{"_id":"q-en-kubernetes-1d77bd2649871c6ff6c830115417b6f0e711b8271e807b6e87ccd28d09c3b787","text":"/sig api-machinery /kind bug /priority important-soon\nIt seems something is racing writhing the logs ? cc:\n/cc /triage accepted\nFailed with this: pull-kubernetes-unit\nthis looks like a test-only race where the test overwriting the global os.Stderr in order to capture error output races with an async attempt to log to it from another test in the same package\nwhich still needs to be fixed\nIt seems to me that this bug can not be fixed with as it is, namely used without a logger object or even a . The smallest tweak that occurs to me would be to add an option to to add an output destination, and later remove that addition. Thus, a test could add logging to its own buffer for the duration of running the code under test.\nThis may be similar to . How can I reproduce this issue?"}
{"_id":"q-en-kubernetes-7601f9dc12de055cd1b4cb3ef8d76a31219db8f7e91e8e590024fe83def0efae","text":"As suggested in , some portions of the code contains the map checking for nil values in map, and others don't have. Even though we're never assigning nil values to this map, we should standardize the portions that does not have this check in module. Example: Should be like this: I can take care of this issue.\n/sig network\nThere have been discussions of deprecating the user-space proxier. Maybe transfer it to . A question is if anybody still uses it. Seems like you do, so if possible please explain if you have a use-case where the user-space proxier is wanted/needed.\nHey Actually, I was casually reading the source code and I saw a over there regarding the nil checks. Regarding your question, we should check if anyone actually uses it though, or maybe discuss its usage or deprecation a little further.\nQuick question though: I saw in the news that the user-space proxy porting was near completion, maybe it's worth talking to them in order to check upon the status.\nI don't object to making it clearer or more consistent, if you want to spend time on it, but it's not a high priority. That said, I think a nil value in that map would be really bad (and should be impossible) and I don't know how we would raise a signal about it.\nlet us know if you want to try to fix it, otherwise we can just close this.\nI already opened a PR with the clearer nil checks for this specific module in , so you can review it. I think whether this would have nil values into the service map we should discuss this apart and try to reproduce this case (if any would happen).\nPR submitted, so let's review and discuss further there. /assign dcbw /triage accepted"}
{"_id":"q-en-kubernetes-a878fa21d0f2204f54eadcddf0eb2e2b9c70477037849758e5d849eca5a66082","text":"As discussed in , Endpoint(Slice) controllers can get stuck if an Endpoints or EndpointSlice resource is stuck in a terminating state. I don't think there's much we can do for Endpoints here since the same name is reused, but EndpointSlice controllers can ignore EndpointSlices in a terminating state. See No response All supported versions Any Any Any Any Any\n/sig network\n/triage accepted\ndeleting a service with a finalizer doesn't affect the endpointslice, it doesn't change its status to termination or unready Repro: finalizer to the service and delete it, it keeps waiting until the finalizer is removed with the deletion timestamp set. the endpoint slice associated Do you mean that when the endpoint slice controller detects that the Service is being deleted should switch the endpoint slice status to Terminating? cc:\nI don't think we need to do anything special if the Service is waiting to be deleted. There are many legitimate reasons for this to happen, including waiting for infrastructure to be removed. My main concern with a finalizer was that one on an EndpointSlice could cause the controller to get stuck on a Service. We don't have the same risk with a finalizer on a Service - I don't think that would cause the controller to get stuck in any way.\ni was unable to repro the issue on my machine: I was monitoring the logs of , and the endpointslice controller never complained about , when I deleted the service. Though, I was able to get the controller to log this error under a different scenario. If we don't explicitly remove the finalizer from the endpointslice, and instead re-create the pod and service like: From what I can understand, because the endpointslice with the finalizer is owned by a service with a different UID than the new endpointslice, it is marked as expected for deletion . The endpointslice controllers calls , which returns true for the endpointslice with the finalizer, because of the above reason. I might be wrong anywhere here, so please let me know if that is the case. As for further action, I'm not sure what can be done here? We could check the endpointslice for finalizers and then skip them . Keen to hear your thoughts about this."}
{"_id":"q-en-kubernetes-494218afe3da5d8310439245c2c94cf516be0eccdf416bb2ba6e4f388dd85472","text":"Cluster api provides a framework to run the e2e tests which uses the conformance image: The gmsa tests use bash to install the webhook: The test will fail when run in the conformance image because of the bash: we should install the component another way run this e2e conformance image with /sig windows fyi <details</details<details</details<details</details<details</details<details</details<details</details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nConformance image:\nHi does the \"work e2e\" test case pass? Deploying GMSA webhook is the same for both \"work e2e\" and this test case.\nno, they both fail due to bash not being present in the conformance image. The gmsa tests themselves pass if I run locally\n/sig testing"}
{"_id":"q-en-kubernetes-0cad02d163e61aff3cc2c1257b68701a42fb1145e0b51f6816fd9a85adca78e6","text":"I create a ResourceQuota with two pods limit in my namespace. Then, I create a statefuleSet with 3 replicas. As expected, two pods are created successfully, and the third pod failed to create. But, the ReadyReplicas is not 2 ,but 1. The ReadyReplicas is correct 1Create a ResourceQuota with two pods limit in a new namespace. 2Create a statefulSet with 3 replicas 3Check if the ReadyReplicas of statefulSet matchs the number of stateful running pods. 1According to the logic of the statefulSet controller, new pod created in current sync is not be immediately counted into the ReadyReplicas. 2Failure to create pod will cause to return error, then skip updating status. <details</details<details</details<detailskind v1.23.3 </details<details</details<details</details<details</details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig apps\n/assign"}
{"_id":"q-en-kubernetes-69811eb8caf7f5e819a148a0e37972ed630718e9e7bfae82935203b5eb94aac5","text":"A lot of different routines failing just when setting up an apiserver. Example: <details</detailsNo errors The test still passes, but those errors seem very worrisome <detailsmaster </details<details</details<details</details<details</details<details</details<details</details\n/sig api-machinery /sig testing cc\nyeah, the framework has some long standing tech debt /cc\nall the panics I see are:\nlooks like again, due to\nLooks like some integration tests must not be enabling SSA.\nyeah, a lot don't, because it requires openapi, which also isn't universally enabled in integration. it is really expensive at API server startup, and when we tried enabling it universally in integration it blew up runtimes and integration timed out super frequently\nthat's what is still open for\n/triage accepted"}
{"_id":"q-en-kubernetes-2764b793b8c8439788da87475c0c83ab3708bc7e770cd0931b8f8933f3ffe095","text":"A pod uses a PVC with . The block volume is backed by a block device on the node (which is created by some CSI storage providers). If the underlying block device corresponding to the volume is accidentally disconnected from the node (due to network, hardware problem, etc...), the user cannot delete the pod. The pod would be stuck in terminating forever since Kubelet refuses to unmount the block volume. This prevents the user from cleaning up the pod and spinning up a new replacement pod thus leading to a long service degradation. For example, if the pod is part of a statefulset, the replacement pod cannot come up due to the old pod being stuck terminating. Kubelet should be able to finish terminating a pod when the corresponding block device is already removed from the node. a CSI storage provider that supports . For example, a statefulset that has volumeClaimTemplates with . For example: <details close<summaryclick to open yaml manifest</summary</detailsa network/hardware problem in which the backing block device is removed from the host unexpectedly. In our example, we delete a Longhorn engine instance manager which leads to the underlying block device being removed from the host. that the pod cannot access the block volume, try to delete the pod. pod is stuck in terminating state forever. Logs from kubelet: <details close<summaryClick to open logs</summary<br</detailsPlease pay attention to the log . Kubelet is refusing to terminate the pod because the underlying block device is removed from the host. If we run on the host, we can see that the backing file of the corresponding loop device is marked as : In our humble opinion, we think that Kubelet should process to remove the loop device corresponding to the deleted block device so that pod can finish terminating as detailed in this PR: We have tested that PR and can confirm that it fixes the issue. We have prepared a PR at : We have tested that PR and can confirm that it fixes the issue. <details</details<detailsRancher RKE2 </details<details</details<detailsRancher RKE2 </details<details</details<details</details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig storage\nRef:\n/cc\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\nThe Kubernetes project currently lacks enough active contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle rotten [1]:\n/remove-lifecycle rotten\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\n/remove-lifecycle rotten\n/remove-lifecycle stale\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\n/remove-lifecycle stale\nkubelet log 13:31:57 wanguo-72 kubelet[1131]: E0428 13:31:57. 1131 ] Operation for \"{^-54fa-4068-bbad- podName:-9521-4aed-9bfe- nodeName:}\" failed. No retries permitted until 2023-04-28 13:33:59. +0800 CST m=+8838. (durationBeforeRetry 2m2s). Error: UnmapVolume.UnmapBlockVolume failed for volume \"pvc--a4e4-48a7-9057-\" (UniqueName: \"^-54fa-4068-bbad-\") pod \"-9521-4aed-9bfe-\" (UID: \"-9521-4aed-9bfe-\") : blkUtil.DetachFileDevice failed. globalUnmapPath:, podUID: -9521-4aed-9bfe-, bindMount: true: failed to unmap device from map path. mapPath is empty node poweroff restart , pod use block mode volume, and pod stuck in terminating\nso it is a new issue, and I encounter it as well with XSKY CSI after node restart. My kubernetes version is 1.25.8\nwhich Kubernetes version is it?\nHi, I also see the issue reported by I'm using v1.25.10. I assume it happens when using containerized kubelet. In my environment, losetup reports weird backing file paths. This can be reproduced as follows. Until this moment, the correct path is visible from the host. But, after terminating the container, I get the wrong results. I consider this behavior causes failing unmap. Any thoughts?"}
{"_id":"q-en-kubernetes-7e37b82aa223af68f9956082602568d7ce4656261f4826eb78fdf65b850e6afc","text":"17:42 PDT 3/29 Failed with the same error for last 4 runs. Beta release cut is scheduled for tomorrow and this is currently a blocker. /sig testing\nLooks like this is also being discussed and x-ref this test was merged 4 hours ago\n/sig api-machinery\nMore x-ref: SIG api machinery is tracking it themselves in There is a PR to increase the timeout in\nMore x-ref: Same post in the .\nPer this should be resolved and not be a blocker, testgrid is still red but I think the expectation is that it should clear up. Leaving this open for now but currently expecting this not to block release cut tomorrow."}
{"_id":"q-en-kubernetes-0e2b3f8fb6f67adcac712517004ad25cf90d5fb207868cfe4db75e7dec831aa1","text":"Schedulerperf sometimes causes \"json: unsupported value: NaN\" panic . Not panic. in my PC, when running schedulerperf with the below config, there is a high probability of the panic. No response_ latest master <details</details<details</details<details</details<details</details<details</details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig scheduling /assign\n/kind failing-test\nNo. This problem is expected to exist for quite some time. Also, this test is just for measuring performance and does not block merging. Correct labeling requires background knowledge of the issue. For example, if, hypothetically, this test is built into ci and blocks the merge when it fails, then you are right, it would be a high priority."}
{"_id":"q-en-kubernetes-03bcb4aac54faf8c21c66b3aef9eb383b1d48edc1c3895ff4293f85ba573947a","text":"As described in , When the pod start or stop, negative metrics generated. generate the correct metrics pod restart No response <details</details<detailsnone </details<details</details<details</details<details</details<details</details\n/sig node /sig instrumentation\ncan you please give a little bit more information on steps to repro. Is it happens all the time? Is the root cause known (I see a linked PR). /triage needs-information\nThis is a feedback from contributors in kubernetes-sigs/metrics-server. It is riggered on specific conditions. From the description of several issues, the way to reproduce should be pod start or stop. such as:\nDo you know if this is specific to ?\nI'm not sure. However, obtains data from the kubelet endpoint. If there is a negative value, it will cause statistical problems.\n/triage accepted /assign\nFYI I encounter this with . seems to be an uninitialized time type. The in is create from the response of cri RPC, this maybe cause either by cri-o or the metrics collect part of Kubernetes. I have not dig into it yet."}
{"_id":"q-en-kubernetes-6dbd0dcd2518d2794a3d8c0ac32676a9023d5ea15d877323b04943ef993b6e1b","text":"once the we hit a beta release, we can reset the default value for strict handling to false to prep for 1.25. It will likely be introduced in /kind bug /priority important-soon /milestone v1.25 ignore template No response this time <details</details<details</details<details</details<details</details\n/sig api-machinery\n/assign /triage accepted\nHello my name is Justin Chizer and I am one of the k8s 1.25 bug triage shadow assigned to track this body of work. Is this still on track for 1.25?\nThis needs to be done now that master is on 1.26 until 1.25.0 is cut, since verify jobs are failing on removal of 1.26 deprecated APIs /priority critical-urgent"}
{"_id":"q-en-kubernetes-66c295a42b3a78c97ddac4723221b560767cd59a6b0c5994796fe27222f09242","text":"Follow-up to Discussed in As part of kubelet tracing beta, we should add tracing support to the kubelet's read-only endpoint, which currently does not have tracing. It should use WithPublicEndpoint to prevent callers from influencing sampling decisions. Part of\n/sig node /sig instrumentation\n/triage accepted /assign\ncan i help you with this issue?\nI don't think we'll be able to start with this until merges\nOk, I'm waiting to see if I can help you. Thank you\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\n/remove-lifecycle stale is merged now, so this can be worked on.\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\n/remove-lifecycle stale\nCould I help with this issue?\nSure. See We pass a NoOp Tracer provider to disable tracing for the endpoint. Instead, we need to pass the tracerprovider, but make sure that the eventual uses withPublicEndpoint. Looking closer, i'm not sure its currently possible to use WithPublicEndpoint with the otelrestful package... doesn't contain that option (unlike otelhttp: ), so we may need to add that upstream first. Another thing to keep in mind is that updating otel libraries is currently blocked by (see my attempt in )\nI gave the whole thing another try. At least locally, the dependencies seem now to be able to be updated. :raccoon: Could you maybe have a look at it :)\nHe's on leave for the time being.\noh, can you provide some feedback or point me to the right people?\nI'm back now. Feel free to assign me to PRs related to this, and I'll review\nWelcome back :) Could you re-run the CI ? Its just the dependencies. The PR itself shouldn't be that much afterwards.\notel is now up-to-date enough. here:\nHi, I would like to ask about one trouble that caused by WithPublicEndpoint. Aggregated APIServer that I customized uses kubernetes's RecommendedConfig. Due to the fact that WithPublicEndpoint is enabled, it can't be tracked by the full trace for its span-link model. And I think child-parent span model is more appropriate. So, could you please give me some advice? Looking forward to your reply.\nthanks. We know about this, and are tracking it in"}
{"_id":"q-en-kubernetes-0e3d2b8fc44edebfc78ccc4286d681ffe69ed5a33416e7378e32d1bbbf6eb91e","text":"/kind failing-test <!-- If this is a flake, please add: /kind flake --/sig node\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nSame failure in\nI think this is related to the change @ could you please help take a look?\nWe already discussed this in The change from correctly points out that the test isn't working as expected. It needs to be fixed. /unassign\n/assign\nAfter , the tests are still failing. Could you help check it?\nThe skip function is used in another case as well. We may remove that like .\nwill fix this.\n/close the test grid is green now.\nClosing this issue. <detailsIn response to : /close the test grid is green now. Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-ee5435918a052cfc901642b3a50e1d285bd128a74b2aad63415d671dbd3daf8c","text":"In our custom plugin, we want to move pods that is unscheduled by the plugin to activeQ/backoffQ when a new Pod is created. So, we define like this: But, even when a new Pod is created, the Pod that was unscheduled by our custom plugin isn't moved to backoffQ/activeQ. when a new Pod is created, the Pod unscheduled by our custom plugin is moved to backoffQ/activeQ. a custom plugin that has pod events in EventsToRegister. a new Pod that the custom plugin will reject. another new Pod that makes the Pod created in (1) schedulable. can see the Pod created in (1) isn't scheduled immediately (it should've been scheduled immediately after (3) because (3) should move the Pod created in (1) to activeQ/backoffQ.) If I understand correctly and if it's not expected behavior, this bug has a non-small impact, especially in v1.24 because we change to 5 min. v1.23.4 but I think latest master has the same issue. <details</details<details</details<details</details<details</details<details</details\n/assign /sig scheduling\nMaybe it's only on a newly created unscheduled Pod. We've moved correctly for events related to assigned Pod.\nAnd deleting Node events has the same issue...? I guess we need to run when some Nodes are deleted. In in-tree plugins, only pod topology spread uses the deleting Node event and I guess this bug affects its performance. (Even when some Nodes are deleted, the Pods, that were unscheduled by Pod topology spread, aren't moved to activeQ/backoffQ by that deleting events.)\nHm, it may be a problem that events related to Node and Pod are not always called when an event occurs. (Even if it's intended behavior, at least we need to write this on the comment of .) /retitle scheduler: events in related to Pod and Node don't always work\nTo summarize: The problem is \"not all events related to Pod or Node don't always make move from unschedulableQ to active/backoffQ and it may not be intended for custom-plugin developers because this behavior isn't even documented.\" I think we should move all unschedulable Pods with all Pod and Node events. If concerned about performance degradation, we can extend or , for example, to have the function like ? This problem has a non-small impact, especially in v1.24+ because we change PodMaxUnschedulableQDuration to 5 min. Let's see the details... ↓ We move the unscheduled Pods (rejected by the plugin that has Pod add/update events in ) in or But, the target Pods against this moving are only unschedulable pods which have any affinity term that matches \"pod\". So,,,, users cannot move from unschedulableQ to activeQ/backoffQ with all assignedPod events? For example, for Pod Topology Spread (it uses Pod All event in ), the new assigned Pod creation may make some Pods, that are rejected by Pod Topology Spread, schedulable. Regardless of whether that unschedulable Pods have any affinity term that matches a newly created assigned pod or not. Of course this is not just a problem for Pod Topology Spread, it could be a problem for user's custom-plugins. I guess we should pass all unscheduled Pods to . Or at least we should document this behavior. No issue on this. We don't move the unschedulable Pods with those events at all. That's a problem... Even when a plugin uses Pod add/update/delete event in and wants to move the rejected Pods with non-scheduled Pod's add/update/delete events, scheduling queue doesn't move the unschedulable Pods with those events. (In our company, we have the custom plugin for and the plugin wants the scheduler queye to move the Pods rejected by the plugin from unschedulableQ to active/backoffQ when a new non-scheduled Pod is created.) We move only unscheduled Pods that pass . checks the Node can run that unscheduled Pod from the points of view of some in-tree filter plugins like noderesources/nodeport/nodeAffinity/nodename. The problem that arises here is similar to \"Assigned Pod's add/update events\". For example, for Pod Topology Spread (it uses Node Add event in ), the new Node creation may make some Pods, that are rejected by Pod Topology Spread, schedulable. Regardless of the results of against that Node. (In , we change the skew calculation of Pod Topology Spread to exclude Nodes that don't match the node affinity. But, after is implemented, the calculation can be configured to include those Nodes.) We don't move the unschedulable Pods with Node delete events... That's a problem. Even when the plugin use Node delete event in , scheduling queue doesn't move the unschedulable Pods with any Node delete events.)\n/cc Could you please take a look\nThese are all intended optimizations. These basic checks prevent the scheduler from retrying unschedulable pods over and and over. So far we haven't had the need to make these checks configurable as they check well established scheduling features. +1 on documenting the behavior, but I'm not sure about changing it or making it configurable (that adds complexity that might not be justified). What is it that you are trying to do?\nWhat is it that you are trying to do? We are trying to remove flushing in queue in , right? Without improving like I suggest, some Pods in unschedulableQ cannot move back to activeQ forever in some cases if we remove flushing. (Actually, our company's custom plugin needs to wait flush because non-assigned Pod's events don't let scheduling queue move unschedulable Pods rejected by that plugin.) So, I think we should simply inform all events to queue and let queue move unschedulable Pods with that events. This can happen in some custom plugins like our company's one and also even in in-tree plugins. For example, Pod Topology Spread (after is implemented). When a user uses and , the creation of Node that doesn't match nodeAffinity/nodeSelector may change the scheduling result.\ncc who implemented NodeInclusionPolicy. Can you provide a concrete example? The policy just refers to the topology spread calculations. It doesn't mean that the node doesn't have to match the nodeaffinity to be able to accept the pod. I'm asking what is your custom plugin trying to do. Depending on whether it makes sense, we need to evaluate if it's worth the complexity.\nFor example, for Pod Topology Spread (it uses Pod All event in EventsToRegister), the new assigned Pod creation may make some Pods, that are rejected by Pod Topology Spread, schedulable This is actually a problem here, 5 mins would be too long for background flush. But it doesn't mean that we should simply shorten the interval time. For example, Pod Topology Spread (after is implemented). When a user uses NodeAffinityPolicy: Ignore and WhenUnsatisfiable: DoNotSchedule, the creation of Node that doesn't match nodeAffinity/nodeSelector may change the scheduling result. The skew of a new created node is always 0, so it's still unfit I think.\nOur company's custom plugin is like plugin. And like coscheduling's prefilter, our plugin rejects Pods in PreFilter when the number of Pods is less than the defined number. Yes, so, maybe coscheduling plugin has the same issue as our plugin. The coscheduling plugin also uses Pod Add event and I guess it's for non-scheduled Pod's creation. (I don't know much about the coscheduling plugin though.) Oh, that's true Please forget about that example. But, how about like this scenario: user uses and Pod is rejected by Pod Topology Spread because of the skew. e.g. a user uses and TopologyKey is hostname, NodeA/NodeB has 10 Pods each, and NodeC has 8 Pods. adds un-tolerated taint to NodeC skew calculation result should be changed and the Pod rejected in (2) will be allowed by Pod Topology Spread. But the Pod isn't moved from unschedulableQ without flushing because of this issue.\nMaybe we can find some pointcuts by the in , e.g. if pod was rejected by PodTopologySpread, we'll when Assigned Pod's add/update events arrived. For customize plugins, we may have to do some refactoring to register the filter func together with the event.\nSo now NodeC is excluded from calculations. Hence the pod fits in either NodeA or NodeB, which were unchanged. Tricky... I'll have to think about it. We really need those quick checks to pass for performance in most scenarios. /triage accepted\nThanks for bringing this up. Some comments below: You're right. Pods failed by PodTopologySpread may not get re-queued immediately. However, an assigned Pod's add/update event would only impact unschedulable Pods with cross-node scheduling directives. As of today, it's InterPodAffinity and PodTopologySpread. So instead of \"pass all unscheduled Pods to movePodsToActiveOrBackoffQueue\", I'd like to come up with a solution like: It's pretty rare to move pods depending on Unschedulable Pods' events. In your particular case, maybe a proactive call would behave more efficiently. It doesn't sound correlated, is it? As the check will just pre-verify the basic checks which are orthogonal with PodTopologySpread. In other words, it's impossible that a Pod failed can become schedulable eventually. Updated: The example is a valid point... Can you share some use-cases to move Pods upon node's deletion? In the before, I don't think we have any, until the last release the introduction of feature (as in that case deleting a node would impact the number of domains).\nto come up with a solution like: If we do so, also, we need to provide the way to pass some functions to decide which Pods should pass to like for out-of-tree plugins? call would behave more efficiently. Ah~, that's a good idea :) Maybe like this (similar to ) user uses WhenUnsatisfiable: DoNotSchedule. Pod is rejected by Pod Topology Spread because of the skew. e.g. a user uses maxSkew: 1 and TopologyKey is hostname, NodeA/NodeB has 10 Pods each, and NodeC has 8 Pods. deletes NodeC skew calculation result should be changed and the Pod rejected in (2) will be allowed by Pod Topology Spread. But the Pod isn't moved from unschedulableQ without flushing because of this issue. So,,, ignoring some events for performance in scheduler side (without providing a way to change this) could be inconvenient for some plugins. What I thought in the above previous comment () is to define \"which kind of resource's events should be ignored\" in plugin side. (similar to 's idea? ) We can add pre filter functions in like: Then we can set any pre-filter functions for each plugin. And, we can define some default behavior for PreFilterFn and use them when is empty. So, for example, if most of the plugins are considered to prefer to ignore the event for Node that doesn't match nodeAffinity/nodeSelector, we can define this behavior as the default behavior. What do you think about this idea?\nThanks. It makes sense to me. To summarize, the following cases doesn't properly re-queue relevant pods: Pod's add/update events -scheduled Pod's add/update/delete events 's add/update/delete events For 1 & 2, I've commented above with solutions. (1 needs core changes similar like podAffinity, 2 may leverage the Activate hook) For 3, with 's examples, I'm convinced this is some area we need to improve. One simple and stratightforward solution is to plumb a hook to let the plugin claim \"alright, upon Node events, move the pod failed by me anyways\". Then once this knob is registered, we move pods failed by that plugin back unconditionally; instead of proceeding with . This hook should be available to in-tree and out-of-tree plugins. In terms of in-tree plugins, it's just PodTopologySpread (we discussed this in today's sig meeting, even InterPodAffinity doesn't fall into this case) So, as a first step, let's ensure we're on the same page and agree on the ^^ summary. Next, we can better evaluate trade-offs and come up with an optimal solution.\nThanks for discussing this topic in the sig meeting. But, I am still believing that all logic related to a plugin should be controllable from the plugin side and we should not create something that can be achieved in in-tree plugins but not in out-of-tree plugins. So... the benefit of the \"customizable PreCheck approach\" () isn't only to solve this issue. We can use this feature for other optimization: like improving the whole performance of scheduler by moving only Pods that are more likely to be able to be scheduled by defining more detailed events on with customizable PreCheck. Many plugins have a particular scenario that makes rejected Pods schedulable and the current doesn't have an abundant enough ability of expression for this. (So, today many plugins need to say \"register this hoge event on foo resource. This event may make rejected Pods schedulable\" on comments in . ) Yes, I watched the sig-meeting recording as well and the opinion that the \"customizable PreCheck approach\" increases some implementation complexity surely makes sense. But,,, isn't this complexity increasing acceptable even if the above is taken into account? EDIT: It is a bad habit of mine to always write long sentences even though I am not very good at English. Summarizing the above: I think it's not a good idea to create something that can be achieved in in-tree plugins but not in out-of-tree plugins. If we need a solid advantage to extend like this, then it has the advantage of optimizing the queue per plugin like described above.\nI didn't jump in the thread since it's actively on the discussion. But, the issue is the similar one. (This time, we can solve that issue by changing scheduling queue logic since It's the k/k feature though.)\nI'm not sure I fully understand your intention. The slack thread was discussing the AssignedPod update case, so it fallls into case 1 in , which we already agreed to improve in upstream.\nHi Are you working on this? I think this is quite important and if you're not working on this, I can follow.\nSorry, leave here for a while. I wanted to wait for (and then forgot to come back here. ) That proposal is quite similar to what I want to see in the conclusion of this issue; every plugin can judge if each Pod in unschedQ can be moved to schedQ/backoffQ. Do you think it makes sense to add \"By which event the pod will be moved back to schedQ/backoffQ\" on 's argument? I know and isn't needed in DefaultEnqueue plugin for API, but that's the last missing puzzle piece : If has , enqueue plugin can judge if the event makes it possible to schedule the Pod. For example, I believe we can move ↓ this logic to Pod Affinity plugin by implementing in Pod Affinity plugin. Pod Affinity plugin will remember Pods rejected by itself in . And in , if the given Pod was the one rejected by Pod Affinity, is AssignedPod/Created|Updated and is Pod affinity with the rejected Pod, then move the Pod back to activeQ/backoffQ. Otherwise reject the Pod moving. We can resolve (3) by implementing as well. (by exactly the way you said.↓) My original concern in this issue is that having logic for specific plugins in the scheduler core may prevent out-of-tree plugins' extendability. The scheduler itself should be pure, and shouldn't do anything special for any plugins implementation ideally. If we can move all such logic to each plugin, that'd be perfect.\n^ is more like a new feature than just a solution to this problem. I created another issue and describe my proposal in detail. Let's have a discussion there.\nIt can be resolved by moving all preCheck into each plugin's QueueingHint. Let me retitle this issue so that the current status is easy to understand. /retitle scheduler: move all preCheck to QueueingHint\nThis issue has not been updated in over 1 year, and should be re-triaged. You can: Confirm that this issue is still relevant with (org members only) Close this issue with For more details on the triage process, see /remove-triage accepted\n/triage accepted"}
{"_id":"q-en-kubernetes-dbc237a3a86890818f5842ea7d977bf1078e775a28c86f5dbe70ccabacff3484","text":"I'm writing a scheduler plugin using the scheduler framework, and in my plugin I'd like to add a using the provided by the . Adding such a Indexer lead to a runtime crash like The reason for this is that I'm not sure the scheduler framework wants to explicitely disallow adding Indexers using the facilities. If this is the case, it should be more explicit somehow because the current behaviour is surprising. If the scheduler framework doesn't want to add this limitation, then adding indexers should not lead to runtime crashes add a new scheduler plugin trying to add indexers. Sample code (relevant minimal portions only) I'm not sure this is a real bug and not a deliberate decision; a research was inconcusive, and a (admittedly quick) search in the past issues was not fruitful either. The fix could be as trivial as <detailsconsole $ kubectl version Client Version: version.Info{Major:\"1\", Minor:\"17\", GitVersion:\"v1.17.2\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2020-01-18T23:30:10Z\", GoVersion:\"go1.13.5\", Compiler:\"gc\", Platform:\"linux/amd64\"} Server Version: version.Info{Major:\"1\", Minor:\"23\", GitVersion:\"v1.23.5+\", GitCommit:\"\", GitTreeState:\"clean\", BuildDate:\"2022-05-10T16:30:48Z\", GoVersion:\"go1.17.5\", Compiler:\"gc\", Platform:\"linux/amd64\"} </detailsbut applies also on latest master (see above) <detailsany </details<detailsany </details<detailsnot relevant </details<details</details<details</details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig scheduling"}
{"_id":"q-en-kubernetes-edf12f67ebea6e64d761d4b44434ed531295d0918a12f54c3c58a4e7fc1edeaf","text":"Issue tracking the addition of Ephemeral Containers to the . Exercising ephemeral containers in the node conformance test is required for\n/assign /sig node /priority important-longterm /triage accepted\nThe label(s) cannot be applied, because the repository doesn't have them. <detailsIn response to : /assign /sig node /priority important/long-term /triage accepted Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/priority important-longterm"}
{"_id":"q-en-kubernetes-27a24ebb81ccbf0117f48661f365439205933241e8a2ae3206579bfe52eec1e4","text":"I think this got stale and closed because it was posted on but is still very relevant: The created by memory.NewMemCacheClient(...) returns an error when calling ServerResourcesForGroupVersion if the APIResources list returned is empty, as seen here: Our particular case came up due to the API group being registered with the API server, but having no resources. If using like to discover with zero current resources there should not be an error returned. Do a discovery with a against a Kubernetes cluster that has at least one with zero resources. The condition on the server that triggers the described issue can be expressed as e.g.: I would be good to get to know the reasoning behind the emptiness check in the , maybe there is good reason for it. <detailsClient Version: v1.20.15 Server Version: v1.22.8-gke.202 </details<detailsGoogle Cloud (GKE) </detailsNo response No response No response No response\n/sig api-machinery\n/cc /triage accepted\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\n/remove-lifecycle stale\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\nReviving this issue based on This behavior has become quite annoying for the custom/external metrics aggregated APIs since after , client-go is now producing a very spammy error log when an aggregated API is found without resource. This is a very common scenario with the metric APIs since metrics are dynamically registered as Resources to the aggregated APIs so on startup the aggregated apiserver may not expose any resource and that's completely normal. It is even more common for implementations of these APIs that register metric Resources from CRDs as CRs tend to be created later than APIServices and the aggregated apiserver. I can see a few path forward to improve the situation for us: the error if we consider that having no resources exposed by an aggregated API is acceptable the verbosity of the log to just show it in debugging scenarios somehow update to defer registration until we have at least one Metric resource. could you perhaps share your insight on that one please? /remove-lifecycle stale\nHello folks! Any update on this topic? The message is really spammy\nRemoving triage accepted manually since there was no update on this for more than 6 months. /remove-triage accepted\nFor context, I traced it back and check has been considered an error since the first iteration of the in-memory discovery client in : It is unexpected for a built-in or CRD groupversion to have no resources, but for aggregated-apiservice I can see it being possible that the object is created and set up for discovery before the aggregated-apiservice is ready.\nI think this is which introduces (but maybe I'm wrong). There is also in that repo about the same issue after upgrading the kube-client pkg. in custom-metrics repo be correct from Kubernetes pov? IMO we need to find a solution soon or the amount of noise will increase once people bump their deps and tooling (, kubectl is affected, and any tool which uses kube-client is affected)\n/triage accepted"}
{"_id":"q-en-kubernetes-3d9481520127fa0e3a70a29d58b540adab132ce25b778905ba34e4a1d73d91cf","text":"Hello, we have been using client-go to measure the latency for all client-go traffic from Cilium to kube-apiserver. We have that since v1.24.0 the URL path is no longer aggregated in a way that this cardinality would no longer happen as presented in the following image: <img width=\"330\" alt=\"Screen Shot 2022-08-08 at 10 18 24 AM\" src=\"https://user-\"It seems this functionality was removed as a part of as it seems that PR only took into account that users of the would only be using it to measure latency to a particular host and not for a particular URL path as we are doing in . Would it be reasonable to revert commit in order to fix this regression? Thanks cc\n/sig api-machinery /sig instrumentation\nBut reverting will bring back the cardinality problem, will not? The latency of each URL path is available in the apiserver metrics ...\nWe have never any cardinality problem with the previous code since we were not using the \"host\" field which is what I assume the original issue about cardinality was trying to fix. We only used the \"path\" of an URL, which was aggregated in a way that would not be facing any cardinality, like the figure attached in this GH description. Now, starting with = 1.24.0, each and is actually a different value which does cause the cardinality problem to appear.\nah, I see the disconnection, I wasn't understanding the issue sorry. The Interface uses the url as parameter to keep backwards compatibility, but the method that gathers the metrics only uses the host to avoid the cardinality problem. You have to update the method that gathers the metrics to only use the host part of the url\nFiled\nshould fix the code in kubernetes ? however, that will break consumer of the Interface that still wants to process the path, but we know using the path will cause issues with cardinality :shrug:\nI would be fine with adding a new label to mitigate that issue since originally the cardinality explosion that was reported was caused by the query string and not the path: When trimming down the whole URL to only the host we even had a discussion to keep the path, but we decided not to since we didn't have any use case for this data. If you have a use case where you need the path in order to get more granularity on the latency of your network then I would be fine with adding the label to Kubernetes, but another path forward could be to also remove the path information from cilium if deemed unnecessary. In the most common scenarios, you will only need the host to get the network latency based on different proxy routes.\nyeah we can't really do much about it, it is either we leak the unsanitized paths or we prevent the users from using the path completely. Since we want to not break the interface and keep the URL parameter, we might just have to keep sanitizing the paths for our consumers even though we are not using the path anymore.\nIIUIC the RESTClient is used in the dynamic client-go too, for these users with 1900 CRDs, the path is really going to explode for them\nThen it seems that the best option to not break users would be to revert , but at the same time I don't really like the idea of spending CPU time sanitizing something that will never be used in the RESTClient\nA suggestion would be to create a new interface similar to the that would have the following method: Thus, those users with 1900 CRDs would use the instead of the .\nafter digging a little bit the underlying issue it looks that the problem was initially caused by a kubelet bug as pointed out . The dependents of client-go don't have access to all the fields of the struct which uses heavily to generate the aggregated URL.\nwell, it is clear now that current status is wrong, we pass the whole URL to the interface and rely on the to use only the Host part of the URL :smile: I'm fine with whatever you decide, to restore or just completely strip the Path Sorry for the unfortunate consequence, at that time I didn't consider that the Interface/Method drift and thought that function was orphan\n/triage accepted /assign"}
{"_id":"q-en-kubernetes-606f0bca48d16c439586f496ff7d55d9265c2f0f7142a6c88110def925a1a6f8","text":"Add additional logging at level that associates UID with the metadata from the request. xref: /sig auth\n/triage accepted\n/milestone v1.26\nMind if I take a look at this? /assign\nHello I am wondering how's this coming along. Is there anything I can help with?\nSorry. I got a bit occupied with a different project.. Will have a PR opened for this for review tomorrow.. I am trying to find a way to inject the Verbs into log. But haven't found the right way to include that into the context.. Any suggestion on what can be done?\nany suggestion on what is the best way to get the additional metadata for verbs into this log? Move that info across using ?\nHello , bug triage team here, just wanted to touch base with you for updates on this ticket. Thanks!\nThanks! This is being worked on and is still on track for .\nThere is a Pull request opened on this already as well. But Waiting on some clarification to see if that needs additional changes.\n/reopen /assign /unassign\nReopened this issue. <detailsIn response to : /reopen /assign /unassign Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/milestone v1.27\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues and PRs. This bot triages issues and PRs according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue or PR as fresh with Mark this issue or PR as rotten with Close this issue or PR with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\n/remove-lifecycle stale\nHi Bug Triage Team for v1.27 here o/ Just wanted to touch base with you for updates on this ticket. Thanks!\nI am not working on this anymore Carlos. This was re-assigned\nThis is tracked for . I've linked the PR."}
{"_id":"q-en-kubernetes-bdbd254a609ed37e50ac4aad043795a1d3e6f151ef17dcbd1f3666120d85d217","text":"4 Oct 2022 [x] Two weeks soak end date: 18 Oct 2022 [x] Test promotion PR: According to this APIsnoop query, there are still three remaining LimitRange endpoints which are untested. listCoreV1LimitRangeForAllNamespacespatchCoreV1NamespacedLimitRangedeleteCoreV1CollectionNamespacedLimitRange Using a number of existing e2e test practices a new has been created for three LimitRange endpoints. The e2e logs for this test are listed below. This query shows which LimitRange endpoints are hit within a short period of running this e2e test example endpoint | useragent --------------------------------------------+\ncreateCoreV1NamespacedLimitRange | should ensure that a limitRange can be listed, patched and deleted by collection deleteCoreV1CollectionNamespacedLimitRange | should ensure that a limitRange can be listed, patched and deleted by collection listCoreV1LimitRangeForAllNamespaces | should ensure that a limitRange can be listed, patched and deleted by collection listCoreV1NamespacedLimitRange | should ensure that a limitRange can be listed, patched and deleted by collection patchCoreV1NamespacedLimitRange | should ensure that a limitRange can be listed, patched and deleted by collection (5 rows) If a test with these calls gets merged, test coverage will go up by 3 point This test is also created with the goal of conformance promotion. /sig testing /sig architecture /area conformance\n/triage accepted\n/reopen\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-124d4c7fb9ee696a83909f715e9bd8706a0f4465e08a35801f7cd2c1ed445555","text":"/kind failing-test <!-- If this is a flake, please add: /kind flake --/kind flake <!-- Please assign a SIG using: /sig SIG-NAME --/sig testing apps\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\ncan you PTAL, it seems this test was in It doesn't seem to have a high rate of failure , but seems flakes are legit\n/assign\nside note: it's really hard to make it fail in a fast machine :) I had to temporarily add a sleep in the reconciler.\nhave you tried Jordan's instructions compiling and running with ? for integration you should have an external etcd instance\nI haven't. Good to know for next time :)"}
{"_id":"q-en-kubernetes-60ef49109d2d49b96468323bb98999522d28f06033d70a5fcbd54b47a5206e9d","text":"There is a regression in v1.24.0 (and above) that causes stale HNS load balancer proxy rules anytime a backend pod is deleted. Each subsequent deletion will leave behind an additional external VIP load balancing rule that references endpoints which no longer exist. This can cause occasional connectivity issues and timeouts, if a stale load balancing rule is matched and it redirects traffic to an endpoint which no longer exists. We are under the impression this may have been introduced (creds to for identifying this.) I expect all load balancing rules created by winkernel proxier (aka HNS load balancer) to be referencing valid backends (aka HNS endpoints) a Kubernetes cluster v1.24 or higher with Windows nodes. a Kubernetes Service with type \"LoadBalancer\" on Windows. one or more pods of the Service. stale load balancing rule left behind referencing endpoint that was already deleted. This issue can be discovered by establishing multiple connections to the service, after pods were deleted. Some of the requests will fail. This issue can be monitored using the following script: Execute as follows: It will print something along the lines of: Another way to inspect this manually is using the following PowerShell: In the $refs, you would see multiple entries (indicating there are duplicate HNS load balancer proxy rules with the same VIP) and that some of the references are invalid (ie not showing up in the get-hnsendpoint output) can force all pods of a given service onto a single Node. This will still cause duplicated load balancer entries, but the connections should all succeed as the right rule should be matched. K8s version <v1.24.0 . This will remove all the rules and re-create them. 24.0 and above Azure Kubernetes Service, but likely impacting others. Windows Server 2019, Windows Server 2022 n.A. n.A. n.A.\n/sig windows\n/triage accepted\n/sig network\n/milestone v1.26 /priority critical-urgent\n/reopen Let's keep this open until backport PRs merge and the fix is available in v1.24 and v1.25 builds.\nReopened this issue. <detailsIn response to : /reopen Let's keep this open until backport PRs merge and the fix is available in v1.24 and v1.25 builds. Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nbackports are merged and patches were released /close\nClosing this issue. <detailsIn response to : backports are merged and patches were released /close Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-b8cca93ec17da9339f471fa9ca1e026d0a3111b9465a3f099d801f39960c5813","text":"Occasionally job completions are counted twice (or potentially multiple times?) in the metric. This happens for about 20% of jobs, but may depend on the Job specificity, which results in an overestimation of the number of finished jobs. The metric reflects the actual number of completed jobs. a kind cluster with prometheus monitoring stack (for example follow this: ). the number of completed jobs by inspecting the metric and delete a job multiple times in a loop Example job: Example repro script: . Check the number of completed jobs by inspecting the metric and the job Yes, in the kube-controller-logs we find the following log lines which are related and probably indicate repeated updated which results in double counting. <details</details<detailsReproducible with kind </details<details</details<details</details<details</details<details</details\n/assign\nFYI\n/sig apps\n/triage accepted"}
{"_id":"q-en-kubernetes-33bcccc36ad7857ca48513378b58272c994a19d465e7a3fa04c2d18941354147","text":"As per, , wire up metrics/slis for kube-proxy\n/assign\n/sig network /sig instrumentation\nThis sounds OK, but we need to translate it into real work-to-do - volunteers? /triage accepted\nis tackling this one :)"}
{"_id":"q-en-kubernetes-b6030dc3945340864f18af2fb5e72900a6b48fa3046bf55a71dccd4a22a1c61f","text":"As per, , wire up metrics/slis for kubelet\n/assign\n/sig instrumentation\n/sig node /triage accepted"}
{"_id":"q-en-kubernetes-c5d117a8f336d7baa6c7dc464d71aea674fd6790ebd6568977412da3e6c52947","text":"The code at: assumes that the transformation does not change the size of the data. Since we know it most cases that it does, we should try to account for that during the allocation.\n/sig auth\n/triage accepted\n/assign\nI think there is no problem with the memory allocation of , because needs to be executed afterwards, in order to avoid executing new memory allocation during append, so is no problem. IMO\nAfter reading your code, I understand.\n/assign"}
{"_id":"q-en-kubernetes-d04eb2bc5823f1c1a59cd0c1fa3dd349c4a46e4907b7acae4285ab5b3420fcb5","text":"I started missing error checks. When adding error checking to this call here, the test started to fail: The test should have continued to work when checking the error. No response <detailsmaster </details<details</details<details</details<details</details<details</details<details</details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig apps\n/priority important-soon can you or someone in apps PTAL?\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues. This bot triages un-triaged issues according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue as fresh with Close this issue with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\n/remove-lifecycle stale\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues. This bot triages un-triaged issues according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue as fresh with Close this issue with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\n/remove-lifecycle stale\n/help\nThis request has been marked as needing help from a contributor. Please ensure that the issue body includes answers to the following questions: Why are we solving this issue? To address this issue, are there any code changes? If there are code changes, what needs to be done in the code and what places can the assignee treat as reference points? Does this issue have zero to low barrier of entry? How can the assignee reach out to you for help? For more details on the requirements of such an issue, please see and ensure that they are met. If this request no longer meets these requirements, the label can be removed by commenting with the command. <detailsIn response to : /help Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign\n/assign"}
{"_id":"q-en-kubernetes-6a3801b014cea1f986db2b64ce4fd61f8c82ef32bc0c69ea931f9eb011400cc9","text":"pull-kubernetes-unit /cc /sig scalability\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nIn local tests, he always passes, so I'll see if the problem is caused by the environment\nIf the lru cache exceeds 30 seconds, the cache will be invalidated, which will lead to unit test errors, and the problem will run normally again after retesting.\nAfter I manually reduce an unhold <- struct{}{}, the same problem occurs, so it seems that the problem is caused by the cache being invalidated for more than 30 seconds. If the test doesn't complete in 30 seconds, then this is probably a machine performance issue\nstress will hang after about a minute in my local test.\n/cc\nThe issue is being fixed"}
{"_id":"q-en-kubernetes-a4c532b66c46d7e1c86822e565ba94897f08b23b3487bd6a9e3ab5ef461e0594","text":"The of a stale pod disruption condition remains after two minutes as the disruption controller fails to change it to to disable the condition. The of a stale pod disruption condition should be changed to after two minutes. with Alpha feature gate enabled is without associated delete operation (it may happen if preemption delete call fails, but it can be simulated by API) 2min as after this time the condition is recognized as stale by disruption controller, as the pod is not terminating The patch of disruption controller fails with the following message in the logs: <details25.x <!-- --</details<detailsReproducible on KIND. </details<details</details<details</details<details</details<details</details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig apps /assign"}
{"_id":"q-en-kubernetes-5ede7b51e13c1ea3ea2962ab6dcab9f1d1124074d4f18613be2484ed83af20a2","text":"Currently we have inconsistency in most of the API where the field name should be used in description so the generated docs match the API fields people actually set. A linter could be for checking To avoid confusing users\n/good-first-issue\n/triage accepted\n/assign\nHey can you provide some more information for this issue\nare you still working on this?\nHey feel free to take it. I am unassigning myself\n/triage needs-information /remove-good-first-issue\n/sig api-machinery"}
{"_id":"q-en-kubernetes-fe6ea65cdcc950bf656b2a6101c40b2d9fb7ef8286e4fe7c20e36c2484125d80","text":"The node selectors gets parsed by in for each node. That work can be done once in . Better performance.\n/sig node /sig scheduling /triage accepted /priority backlog\nI can try this one. /assign"}
{"_id":"q-en-kubernetes-4e3771a706267d556a628e1592b3cfd4aa085c39590410edb29436f8af97f936","text":"During kubelet start up if the system reserved or kube reserved values failed to parse, the kubelet won't start up. The error logged looks like, However, it doesn't not inform the user which system reserved or kube reserved quantity has failed and what was the corresponding value of that quantity. The error log should also container quantity type and the value that failed to parse, so that user can fix them easily. Try to start kubelet with unparsable values of system reserved, something like, The error logged would look like, As you can see above, it doesn't say was not able to parse with the value No response <details</details<details</details<details</details<details</details<details</details<details</details\n/kind feature /sig node\n/assign\nThanks for Should I close in favour of yours?\nOh sorry, I didn't see your PR. Either way works for me. You can either incorporate ideas from my PR and I'll drop mine, or drop your PR. I don't feel strongly either way.\n/triage accepted /priority backlog /assign\nLet's go with yours!\nThanks /close (in favor of )\nClosing this issue. <detailsIn response to : Thanks (in favor of ) Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-4f3ad134c0be7c34161f0dd1d976c17eaf5b5451b4afb697d90058214a811ed6","text":"When submitting a windows pod using resource limits on a node with more than one processor group, the specified count is incorrect due to the usage of which does not take processor groups into account. For example: node with 96 cores cpu limit 94 returns This results in a CPU maximum of since the range is from [1, ]: Our initial goal for this was to limit a process to not use all available CPU and therefore render the node uncontrollable. A CPU maximum of instead of for that specific case. For the calculation, it should use the CPU count for all available processor groups instead of just the current one. Launch a node with a lot of CPUs, for example and schedule the following pod on it: Then SSH into the node and inspect the running container using . The value will be set to . (which just uses ) is able to correctly find out the number of CPUs available to a system. Using for calculation or incorporating that code might solve this issue. <details</details<detailsAzure in-tree </details<details</details<details</details<detailscontainerd://1.6.10 </details<details</details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig windows\nThere is already some code which uses GetActiveProcessorCount:\n/assign\n/triage accepted"}
{"_id":"q-en-kubernetes-3b55c55ac2585b9d5a1fe5fc4b521e0bc1779db765b7266913ee3a9839d24e7c","text":"an E2E test that assumes that a service named \"kube-dns\" exists in the \"kube-system\" namespace. This is a kubeadm-ism, and the test fails if the cluster DNS is not configured the way kubeadm configures it. Although this E2E test isn't a conformance test, making it functional on non-kubeadm clusters would increase its utility for distributors that may wish to run this E2E test as a conformance test. [sig-network] DNS HostNetwork should resolve DNS of partial qualified names for services on hostNetwork pods with dnsPolicy: ClusterFirstWithHostNet Since we ran it on an OpenShift cluster. No response The test makes assumptions about how the cluster DNS is configured. The test uses (not ), so it does not need to specify , as far as I can tell. Omitting obviates the need to look up the service. I believe that the service lookup was left in as an oversight during the development of the test when it was changed from specifying \"None\" to specifying \"ClusterFirst\": Cc: /sig network\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nresponded on the pr.\nIt seems kubeadm inherited the status-quo, I can't find any place that mandates the Service name for DNS Asking for double checking, but the docs seems to mention that this is more a recommendation than a preescription If it seems that the test can be done in a more portable way we should do it"}
{"_id":"q-en-kubernetes-cbf623f063f8ed961cdc26df1469f9163f1b56c0c9963dd70fb5ab18a7199110","text":"In , we discovered that our plan for eventually removing the scheduling plugin when CSI migration is GA is not possible because the plugin also applies to non-migratable volume types. However, if we leave the plugin as is, then even migratable PVs will be broken if we remove beta topology labels from Nodes in the future. This is because the plugin does not have any handling for migrated volumes. I suggest we add CSI migration handling to the VolumeZone plugin. I think we just need to detect if the volume is migrated, and if so, skip the VolumeZone processing. For migrated volumes, we translate into so they end up getting processed by the plugin. In a cluster with csi migration for a specific intree volume type enabled, an intree PV with beta topology labels the beta topology label from a node to schedule a pod using that pv No response <details</details<detailsAny that supports csi migration </details<details</details<details</details<details</details<details</details\n/sig storage /triage accepted /priority important-soon For e2e testing, this needs an environment that is configured with csi migration\nHi,I'd like to work on this issue. I previously wrote for plugin, so I'm familiar with it's logic. However I'm not familiar with the whole thing, and I may ask you some questions during the work :( Can I try this one? Here are some questions ahead: I think what you want to do is to schedule a pv with old labels to a node with only new labels. And you think we can translate into in volumezone during (maybe), and update this pv so that in stage, can , so we don't have compare Topology labels in . My question is: Can't we just translate this for all pvs? Why do we need to detect if the volume is migrated? What's the difference between a migrated pv and a non-migratable pv in this scenario?\nI think this is also another alternative solution that could also work, but from , it may not scalable to do an automatic migration for all the different APIs and controllers that use the node labels. The main difference with CSI migration, is we have a good idea of all the different components that would be impacted and also have reasonable control over those components.\nA few more questions Will the volumes be migrated before we take a pod off the scheduler queue? If so, how can I detect this. Currently I find a function in , but nothing like . Is this translation done in plugin or before scheduling? I think currently in , we consider topology label and beta label as different labels. That is ,if we have a pv with and a node with , even if they have same values, will let the pod pass filter. If we're going to remove beta labels on node,maybe we should try to consider two labels as the same in ? Here's my understanding: For migrated volumes, we translate Topology labels to pv.NodeAffinity and update the pv, then we skip volume_zone(that is,don't compare labels with nodes). What about non-migratable volumes, I think we should continue to compare with node labels for non-migratable volumes?\nFor detecting if a PV is migrated, you can use the same method that's used in the VolumeBinding plugin: The only difference is you don't actually need to translate the PV in the VolumeZone plugin since it's already handled by the VolumeBinding plugin. For non-migratable volumes, I think we can continue with the same VolumeZone logic. For now, we'll rely on users to update their PVs with the GA labels.\nThanks ,for your detailed explanation. This is the implementation plan I prepared: In , we check if a pv is migrtable by using ,if so we skip volumezone since will translate this pv and handle it later, otherwise we save the label of this pv in state,and compare it with node's label in .Is that feasible to you?\nYup that sounds good to me!\nThanks! I'll try to work on this one. I'll file a pr a few days later and we can talk about the details in there. In the meantime, I think I previously wrote a bug in plugin, I wonder if you have time to confirm it /assign\nThis issue is labeled with but has not been updated in over 90 days, and should be re-triaged. Important-soon issues must be staffed and worked on either currently, or very soon, ideally in time for the next release. You can: Confirm that this issue is still relevant with (org members only) Deprioritize it with or Close this issue with For more details on the triage process, see /remove-triage accepted\n/triage accepted"}
{"_id":"q-en-kubernetes-9d20c811a6cc0468fab29bb4f7b06e98edd17b0aaa8c8e7dfd37d8a793d3628d","text":"In a cluster with large enough numbers of objects, it seems to be impossible to clean them up with . Because the deletecollection is failing while listing the objects from etcd. The failure happens after exactly 34 seconds, due to the 34-second max timeout introduced in PR (cc for insight on that PR). The deletecollection should be able to run for longer, and succeed. I recognise that there are still risks of the API server using too much RAM, if deleting a very large collection due to issue . But, until that issue is fixed, it seems like allowing to run for longer might be a good idea. As far as I can tell, the 34 second timeout does not apply to normal LIST calls, so on that basis I feel it should not apply to the giant unchunked LIST call that happens inside . a few hundred thousand objects of some particular type. (Argo workflowtaskresults happen to be the type I'm looking at while I write this) deletecollection. (Optionally passing some kind of fieldSelector, but I don't think that's essential) that a 504 failure code is logged in kube-audit logs, and that a log entry similar to this is logged by the API Server (In practice, the total time may be a little over ms). far as I can tell, the delete collection really has failed at this point. I notice that PR says that, in general, \"The go-routine can keep running, while the timeout logic will return a timeout to the client\" but I don't think that's what's happening here. I think that, because the 34 sec timeout is used in the context of the etcd call, the etcd list call really does fail, and so the code never proceeds to actually deleting anything. No response 22 AKS <details</details<details</details<details</details<details</details\n/sig scalability\n/cc\nthe deadline has been in effect before PR was introduced If we remove this deadline from here, then the deletecollection request will have at most to complete (default). The list call that is failing in your case, is a direct call to the storage, right? on the contrary, we don't enforce the timeout for LIST thinking out loud here: I don't know if it makes sense to remove the from should we allow timeout for DELETECOLLECTION to exceed ? thoughts?\nThe list call that is failing in your case, is a direct call to the storage, right? It reads from etcd, not from API Server Cache. Is that what you were asking Looking at the bigger picture, I think your question about 60s timeout is a good one. Especially in the context of this: In particular, changing the timeouts won't solve all of the problems with DELETECOLLECTION. My assumption, when logging this issue, is that it might be a quick way to solve just some of them.\nI think we should remove this - it's effectively a regression that we introduced with the PR. That's a good question, and I'm not entirely sure. Maybe we should, but I would do that separately. In the meantime, I would much more prefer to switch the DeleteCollection implementation to internally use pagination, as it would allow us to 1) not block on full list (and as a result make some progress) 2) would be some preparation towards potentially exposing a Limit option in the API... /kind regression\n/sig api-machinery\n/triage accepted\nI suggest the following: - can you please take care of reverting the 34s timeout for DeleteCollection calls? this is done, let's close this issue and continue operating on\nsounds good, thanks! /assign"}
{"_id":"q-en-kubernetes-621e1f8c4833588c253900be62e53ed8bdebfa09aa7a800b8fd57d90f71db151","text":"handles cases when csiNode is nil. This is condition is used by cluster-autoscaler internal working. However, the log line refers to causing seg fault\n/sig scheduling /sig storage\n/sig storage /priority important-soon /triage accepted\n/sig storage /sig scheduling Note that the lister returns an error for the case when the doesn't exist. But it's ignored here: It looks like the log line that causes the panic was introduced in in 1.25+\ncc\n/assign\nAs far as I can tell this issue is only fixed in version 1.27 (by ). Is there a plan to port it back to releases 1.25 and 1.26 as well?\nit has already been picked to 1.25 and 1.26 () and will be in the next patch releases for those versions\nGreat. Thank you for the pointer."}
{"_id":"q-en-kubernetes-0fb185f5a6af9a37731bfe838c48ac020083298397605426e27161f4f985084b","text":"[X] Test approval issue: [X] Test PR: [x] Two weeks soak end date: [X] Test promotion PR: According to this APIsnoop query, there are still some remaining APIService endpoints which are untested. Using a number of existing e2e test practices the has to cover the outstanding APIService endpoints. There has been some around how the current Conformance test is using RBAC, which is optional and not part of Conformance. Those calls will now only happen if is true. The e2e logs for this test are listed below. This query shows the following APIService endpoints are hit within a short period of running this e2e test example endpoint | useragent ---------------------------------------------+\ncreateApiregistrationV1APIService | Should be able to support the 1.17 Sample API Server using the current Aggregator [Conformance] deleteApiregistrationV1APIService | Should be able to support the 1.17 Sample API Server using the current Aggregator [Conformance] deleteApiregistrationV1CollectionAPIService | Should be able to support the 1.17 Sample API Server using the current Aggregator [Conformance] listApiregistrationV1APIService | Should be able to support the 1.17 Sample API Server using the current Aggregator [Conformance] patchApiregistrationV1APIService | Should be able to support the 1.17 Sample API Server using the current Aggregator [Conformance] patchApiregistrationV1APIServiceStatus | Should be able to support the 1.17 Sample API Server using the current Aggregator [Conformance] readApiregistrationV1APIService | Should be able to support the 1.17 Sample API Server using the current Aggregator [Conformance] readApiregistrationV1APIServiceStatus | Should be able to support the 1.17 Sample API Server using the current Aggregator [Conformance] replaceApiregistrationV1APIService | Should be able to support the 1.17 Sample API Server using the current Aggregator [Conformance] replaceApiregistrationV1APIServiceStatus | Should be able to support the 1.17 Sample API Server using the current Aggregator [Conformance] (10 rows) If a test with these calls gets merged, test coverage will go up by 4 points This test is also created with the goal of conformance promotion. /sig testing /sig architecture /area conformance\n/triage accepted"}
{"_id":"q-en-kubernetes-3a4953639a67ca9704d0405f187ec716ef55e55986e733d797d98d93a6ddbbdf","text":"After the cluster was updated to enable PD CSI Driver, there was a race condition where AD Controller (kube-controller-manager) started to BEFORE the PD CSI Driver was up, thus AD Controller gave us the error Then AD Controller triggered the disk detach because the disk/volume was not populated in the desired state of world. Unfortunately the disk was in use during this detach and got corrupted. Since cluster upgrades can cause the controller manager and CSI drivers to start at different times, the AD controller needs to be more conservative when detaching We have not been able to reproduce this, it came up from a customer. /sig storage /assign <detailsk8s 1.23, CSI migration enabled </details<detailsGKE </details<details</details<details</details<details</details<details</details\n/triage accepted\nI think there's potentially two changes we should do: 1) for the initial DSW population, we should skip the check that if the csi driver is not installed. 2) in general, we should not skip volumes during dsw initial population if an error is returned. If we had some temporary api server error then that would cause volumes to not get to dsw and similar behavior happen. cc\nfor a repro, maybe you can try something like: a cluster with csi migration enabled (control plane + nodes) a pod with intree-volume and get it running the csi driver node daemonset. verify the driver is removed from CSINode kube-controller-manager\ncc\n/cc was trying to repro this, but on GKE with the managed driver which made it complicated.\nYeah maybe gce kube-up would be easier for a repro\nIf CSINode says that it has migrated GCP PD volume plugin to CSI, then IMO A/D controller should always attach GCP PD volumes to it using CSI, regardless if the CSI driver is present or not. In-tree volume plugin in the kubelet is disabled and can't mount any in-tree volumes anyway. To flip the migration feature gate / upgrade / downgrade kubelet, the cluster admin must drain the node first to make sure everything is detached. A/D controller should not see change of CSINode migration annotations with any volume attached to it.\n/priority important-soon\nAnother option could be to process these volumes anyways but add them to DSOW with migration status as \"uncertain\", so as they can be re-processed later. But surely it looks like might be easiest unless it has any downsides.\nThere wasn't a node update in this case---there was a control plane update which restarted the PD CSI driver, and it looks like the CSIDriver resource was updated or changed in some way which raced with the kube-controller-manager restart.\nIs there anyway we can know if the csi driver is up and running(like a health check ping) before populating DSW?\nWe use the CSINode registration to determine that. The discussion here I think is concluding that we should not use that as a criteria for adding the volume to DSW. We only need to check that CSI migration is enabled on the node.\n+1 we need to be very resilient against races, I think we have to assume a driver is in process of being installed at any point in the AD controller reconcilliation.\n/assign sunnylovestiramisu\nReproduced with the steps: In /var/log/kube-controller-\nTested with the and doing the reproduce steps above. The error of does not show up anymore. The error message still get logged in the kube-controller- for debugging purpose: Ran Ran On the GCP console, the disk is used by node: Ran I redeploy the pd-csi-driver and then make sure it is up and running. Ran"}
{"_id":"q-en-kubernetes-7b4edcc0e1d67f324507c5f522212f0fe89e066aa1850160b4c423b607e8bd6a","text":"At (originally ) component-base metrics exposed Registerer API for raw prometheus registration, and I took dependency in The resulting does not seem to emit the new metrics (example: ). The metrics handler should emit the registered metrics. Falling back to does work. So there is some material difference with . 24, 1.25, 1.26 GKE No response No response No response No response\nThere are no sig labels on this issue. Please add an appropriate label by using one of the following commands: - Please see the for a listing of the SIGs, working groups, and committees available. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign\n/assign jkh52"}
{"_id":"q-en-kubernetes-f554e53a2b0895170400ceeeb3f9d1f2002f8788d2419fe2262e5a5418c69b9b","text":"I am tracking a changed behavior in networking with kubernetes 1.26. If I create a service of type externalTrafficPolicy: Local, and there are no local endpoints in a node, kube-proxy does not create hns loadbalancers for the clusterIP of that service. With kubernetes 1.25, it creates a loadbalancer for clusterIP (but not externalIP). This is the relevant log message: Looks like this comes with a result from a new function isAllEndpointsTerminating introduced in kubernetes 1.26. In Linux, we can still access clusterIP even if there are no local endpoints for the service. Thoughts? The end result is that in Windows, clients cannot access the clusterIP of such a service with there are no local endpoints.\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig windows\ncc\nDoes this issue occur when there are no local endpoints for a given service? If there is a local endpoint, then you can reach the service again, right?\n, Also let us know, do you have any InternalTrafficPolicy configured ?\nDoes this issue occur when there are no local endpoints for a given service? If there is a local endpoint, then you can reach the service again, right? That is correct , Also let us know, do you have any InternalTrafficPolicy configured ? We do not have that configured"}
{"_id":"q-en-kubernetes-ca50d3a805f7d5ea4c42b514909abc5ee42a0696ae08bb02e50c996a5205edf4","text":"I am a new guy to kubernetes world, and today I am looking at the source code k8s, I found out a line of code can optimize in this line, `, cus the logic here should be like impl the startController logic and then we put it into the register to find out if the controller was registered twice, if so the program will panic. I think it's better that we change to the one that I provide, which can make code clean as well, but as I'm a new guy, I need some advice on it, can someone help me out? /help wanted /area controller-manager\n/sig api-machinery\n/sig node Would you submit a PR to fix it if you think this is a problem? /cc as the owner of DynamicResourceAllocation was as alpha level feature in v1.26.\nyou're say me or the pohly? this is my first time contributing on open source, might have some issues on communicate\nYou. Feel free to submit PR if this is an optimization. I as he is the feature owner.\nthanks, I will submit PR! so the following step should be: me fork the k8s repo to my github and git pull to my local, optimize code, then git push and submit PR, right? the editing should be written on branch master\n/triage accepted"}
{"_id":"q-en-kubernetes-264efff8d550b2b940859448579090d776ed468b0b9918f28891ec4e8656862e","text":"Serial e2e node jobs, e.g. Since 03-01 No response Seems related to No response /sig node\nFailing run -\nI think there is a typo there - should fix it\ncc\n/triage accepted\n/cc"}
{"_id":"q-en-kubernetes-b4d7bf6589ab07972af2f261b13b4e6cd41e93687d14412f9fc20bdc601a8f5e","text":"The test is flaking: /assign /sig api-machinery /priority important-soon\nOk - it's fairly easy to reproduce by just adding into the test I will validate the fix.\n/triage accepted"}
{"_id":"q-en-kubernetes-d2ad17ee33cbb9dd0cc3cd8888bfcb21e558f50039e64e1a5ba0cdc518d0afba","text":"During OPNFV test execution we noticed the folder selinux context changed. e2e test log How the folder changed after the test executed. I expected the TC won't change the selinux context. The question here, should this sigstorageserial TC set work on an selinux enabled K8S cluster or not? This question valid for the another sigstorage TC set with several test case, where selinux block a lot of TC too. See above. No response_ <details</details<detailsOpenStack </details<details</details<details</details<details</details<details</details\n/sig testing\n/sig storage\nThe Kubernetes project currently lacks enough contributors to adequately respond to all issues. This bot triages un-triaged issues according to the following rules: After 90d of inactivity, is applied After 30d of inactivity since was applied, is applied After 30d of inactivity since was applied, the issue is closed You can: Mark this issue as fresh with Close this issue with Offer to help out with [Issue Triage][1] Please send feedback to sig-contributor-experience at . /lifecycle stale [1]:\n/remove-lifecycle stale\n/assign\nReproducer with simple kubernetes/kubernetes artifacts:\n/triage accepted"}
{"_id":"q-en-kubernetes-5cb0acceb3b2c9741edb8287323f7f2a83b92cd21a2db6f5b7d0a934aaab11a1","text":"Hello, Upon visiting the . There is a link for interactive tutorial which is being inactive. Error from (k8 page)[] is ! We have to remove the page link.\nlabel:sig/docs\n/sig docs /kind documentation\nPer let's delete this reference for now and we can add something later potentially following\n/triage accepted /assign"}
{"_id":"q-en-kubernetes-f8150d0761e9abef87e3fc2514e317905a8a5efbd3ee4682d7d9e753bfb06f28","text":"In , when timeout or other error occurs, is called. This will make current routine terminated and miss the , which will lead to some leak routines. For example, this routine will block at waiting for No leak make any of these errors happen No response <details</details<details</details<details</details<details</details<details</details<details</details\nsig Testing\nI would like to take this up!\nActually, I have a fix in , but its fine if you have a better solution.\n/sig api-machinery\n/assign /triage accepted"}
{"_id":"q-en-kubernetes-40efd2553ca6dbc27af1b0d76be94bc4771f4cbb60579f09af69972f70d76763","text":"Data as of June 3rd, 9:00 PM Eastern. Latest data can be found here: Prow link to the 2 jobs: * Additional Context:\n/sig storage cc\nI'm working on , it's hard to get a job yaml right at the first time. I'll check iscsi. /assign\n/triage accepted /kind failing-test\nci-kubernetes-e2e-storage-kind-disruptive fixed: has 1 green run pull-kubernetes-e2e-storage-kind-disruptive fixed: , it fails when it should fail, it's green otherwise, I'll continue with iscsi.\n/reopen it's still red, but at least it installs now:\nReopened this issue. <detailsIn response to : /reopen it's still red, but at least it installs now: Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\ne2e-gce-iscsi-serial job passes now: e2e-gce-iscsi is getting greener: I'll let it run for few more days to see if today's failures are flakes. /reopen\nReopened this issue. <detailsIn response to : It's getting greener: /reopen </details\niscsi-serial is green:"}
{"_id":"q-en-kubernetes-0d64adb5f3b9cc963f235d0a661eff243a85aba7130eedf5f14efd6e46642f7c","text":"When a ResourceClaim: uses delayed allocation does not get created for a specific pod (i.e. not from a template) then it currently gets allocated and remains allocated until: scheduling of a pod fails because the node originally chosen for the allocation isn't suitable, in which case the scheduler triggers deallocation via the user deletes the claim What we want instead is that deallocation gets triggered when the last user of such an allocated claim gets removed from . To avoid races, deallocation must be done through . This can be implemented either as a controller or directly in the API server.\n/assign\n/triage accepted /sig node /priority important-longterm /kind feature"}
{"_id":"q-en-kubernetes-3b5310a4a2858051e0bb5a7cd0371fe9b77833aa3129c76b899286c169596ce8","text":"We ran into an issue with an old version of k8s 1.18 that we were running where subnet for cluster IP is given as . The control plane comes up just fine but then when you try to create a k8s service, it fails due to a broken behavior with the IP Allocator Since this is not really a supported version of the k8s, I went around digging up a bit more to see what is the regular behavior and found this to be very inconsistent across a few release branches. bash go test -timeout 30m -run ^TestAllocateNextFC$ -v === RUN TestAllocateNextFC ? --- PASS: TestAllocateNextFC (0.00s) PASS ok 0.566s go func TestAllocateNextFC(t *testing.T) { , cidr, err := netutils.ParseCIDRSloppy(\"fc::/112\") if err != nil { t.Fatal(err) } // Need to be replaced with the suitable handler for that version of the branch r, err := newTestAllocator(cidr) if err != nil { t.Fatal(err) } defer r.Destroy() ip, err := r.AllocateNext() if err != nil { t.Fatalf(\"%s\", err.Error()) } fmt.Println(ip.String()) } Add the above Test case under the IP allocator test case No response_ <detailsNA </details<detailsNA </details<detailsNA </details<detailsNA </details<detailsNA </details<detailsNA </details\n/sig network /assign /kind bug\nThanks for helping me with this issue.\n/triage accepted /assign"}
{"_id":"q-en-kubernetes-81eb5015d5973208aacfdfea2b43bbb67343c93a54a7e2a7347e1d5403a2751d","text":"During implementation of KEP 3393 (PodReplacementPolicy for Jobs) we found that the unit tests of do not follow common patterns established in k/k. 1) The tests are not written using table driven design (). 2) The tests use rather than . This probably should be done as two separate PRs but wanted to highlight some areas that could use some cleanup. Not using means that we don't see diffs in the unit tests that makes this difficult. And not using table driven tests makes it difficult to add new test cases as it usually is brought up during review that we should use a table-driven test but the entire file does not obey this format.\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig apps /kind cleanup /help /good-first-issue\nThis request has been marked as suitable for new contributors. Please ensure that the issue body includes answers to the following questions: Why are we solving this issue? To address this issue, are there any code changes? If there are code changes, what needs to be done in the code and what places can the assignee treat as reference points? Does this issue have zero to low barrier of entry? How can the assignee reach out to you for help? For more details on the requirements of such an issue, please see and ensure that they are met. If this request no longer meets these requirements, the label can be removed by commenting with the command. <detailsIn response to : /sig apps /kind cleanup /help /good-first-issue Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nI'd like to work on it /assign\nI'd also like to make some contributions. /assign\nLooks like this need two PRs. tests are not written using table driven design (). tests use Differences rather than Cmp.Diff. I think I can work on the 2nd one and leave the 1st one to you. Sounds good?\nBy looking through the , I found the main function to comapre the results have already used cmp.Diff. Can you elaborate more on which part you expect to use cmp.Diff? func verifyDeletedAndPatchedPods(t testing.T, client fake.Clientset, wantDeletedPodNames, wantPatchedPodNames sets.String) { t.Helper() deletedPodNames := getDeletedPodNames(client) if diff := cmp.Diff(wantDeletedPodNames, deletedPodNames); diff != \"\" { t.Errorf(\"Deleted pod names (-want,+got):n%s\", diff) } patchedPodNames := getPatchedPodNames(client) if diff := cmp.Diff(wantPatchedPodNames, patchedPodNames); diff != \"\" { t.Errorf(\"Patched pod names (-want,+got):n%s\", diff) } }\nI am really sorry about this. Too much multitasking!! I meant not\nHey ! Sorry, I didn't see this before. No problem at all, but consider that one change depends on the other, and since my task implies a big refactor, there might be conflicts when merging your PR... That's why I see two possible scenarios happening: Your PR gets merged before I raise mine: Then fine, I'll rebase your changes and adapt my code Mine is ready before yours is merged: In this case, I propose you commit your changes to my PR. Both of us will appear as coauthors of the change Getting a PR merged in the Kubernetes project takes quite some time (trust me, I talk by experience and can confirm ) WDYT?\nLet's see how it goes. We can always rebase the code change to adapt new.\nbtw, did you already begin to work on task 1? If not, I may just go ahead and work on it right away since I have already dig into it.\nyes, I already started And If you ask me, better to keep it this way so we both have an opportunity to have a contribution\ncool. Sounds good.\n/reopen\nReopened this issue. <detailsIn response to : /reopen Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nHi if this issue needs to be worked on any further, please assign me for the same.\nhas a PR up that will solve the second part of this issue."}
{"_id":"q-en-kubernetes-b9c9e4c9435667e2a99f1ef15bb5076835ed010ba27dfcae1c797422f3eb8ae9","text":"When I put a Pod yaml file without in , started the kubelet and hoped it would create the corresponding staticPod for me, I got the following error, which also caused the staticPod to not be created. I then read the kubelet code and quickly found the problem. When kubelet tries to decode the pod yaml file, it will apply default values and validate the pod. The function to apply the defaults is in . In this function, the is set. This causes an error when is not set, because it causes the '-' character to be at the front. I think kubelet should circumvent this error, because is optional, It is possible for this field not to be set. When Pod's is not set, the '-' character at the start should be avoided, as it will cause an error. Prepare a Pod yaml file without and set it into . If the maintainers think this circumvention change is necessary, I'm happy to contribute code to the kubelet! <details</details<details</details<detailsconsole $ cat /etc/os-release NAME=\"Alibaba Cloud Linux\" VERSION=\"3 (Soaring Falcon)\" ID=\"alinux\" IDLIKE=\"rhel fedora centos anolis\" VERSIONID=\"3\" PLATFORMID=\"platform:al8\" PRETTYNAME=\"Alibaba Cloud Linux 3 (Soaring Falcon)\" ANSICOLOR=\"0;31\" HOMEURL=\"\" $ uname -a Linux cloud 5.10.134-14.al8.x8664 SMP Thu Apr 27 16:46:29 CST 2023 x8664 x8664 x8664 GNU/Linux </details<detailsusing kubeadm </details<details</details<details</details\n/sig node\nI would like to open a PR to fix this issue, can I be assigned?\nIf the maintainer thinks this should be fixed, I want it fixed by myself:)\nOK\nI do not know if a pod without name should be considered valid Please discuss with sig-node before doing anything to understand what is a valid solution for this problem /cc\nnames are needed for the Pod. I think the correct fix for this issue will be to add validation for static Pods to have a name. This validation is OK to add since Pods without names are not working anyways because of this bug.\nOkay, I will submit a PR to add validation for static Pods to have a name.\nHey, I found out that kubelet will apply default value first and then validate the pod. Do you mean to let me add a single validation for static Pods to have a name before applying default value?\nhi and If modified according to this logic, it seems that the current code is fine. Current logic: it will use the old pod.Name to generate the new pod.Name , in ValidatePodCreate, it will validate pod name. The pod.Name is already to new pod.Name this case, the original pod does not have a name and will also be recognized. It's just that it's identified with a validation prefixed with \"-\". The logic of the current validation will recognize the pod without a name. In this way, is it necessary to improve the error information?\n/triage accepted\n/priority backlog"}
{"_id":"q-en-kubernetes-05b4a5e90cae8fac52b933c207fc27fdc8e8366e1c9825afa2901a397fe18599","text":"What happened: Shortly after was merged, the \"multiple nodes reallocation works\" E2E DRA test started to fail randomly. Please provide links to example occurrences, if any: /sig node /kind bug needs to stay green. No response master n/a No response No response No response No response\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-dc4a0799b29ad3128b2133c6e877700f4afb8f875c01feda5b1cbe3a0b920c93","text":"Kubernetes does not respect the of a probe. I found this while fixing the e2e tests. Test result: /kind regression /kind bug /sig node /priority important-soon The of a container's probe should override the pod's . a pod with long pod-level terminationGracePeriodSeconds and short probe-level terminationGracePeriodSeconds. for the container not respecting the probe-level terminationGracePeriodSeconds No response <details</details<details</details<details</details<details</details<details</details<details</details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nI guess the probe's terminationGracePeriodSeconds field is dropped for some reason.\n/assign"}
{"_id":"q-en-kubernetes-1353727e47ac172d7c3c16bfb941e73d4388250de89fe1dda855bf0e466f34c4","text":"A client is allowed to add custom finalizers to Custom Resources with a finalizer name that's not fully qualified (e.g. or ). However they're prevented from doing the same on builtin resource types, returning the following error: Invalid value: \"\": name is neither a standard finalizer name nor is it fully qualified Per , custom finalizer names are expected to be in format \"consist of a domain name, a forward slash and the name of the finalizer\" (e.g. ). Therefore, I expected the enforcement to be the same for CRs and builtin types. a custom resource on the cluster it to add the edit does not fail the same on a builtin resource (e.g. Namespace or Pod), observe it fails No response Reproducible at least starting from 1.21 and still on latest versions. <details</details<detailsN/A kind cluster </details<detailsN/A - kind cluster accessed via macOS </details<details</details<details</details<details</details\n/area api /cc /sig api-machinery\n/assign /triage accepted\nThere is no validation of finalizer in and . Please unassign me if you guys have started to work on it. /assign\nit is called from here on custom resources:\nlacks validation for finalizer name. I have raised a pr for this. PTAL."}
{"_id":"q-en-kubernetes-9b2543205b56fd0334950c50eecaf2eb4a2a6255c212e6088674ae15dcfb1966","text":"/kind failing-test <!-- If this is a flake, please add: /kind flake --/sig network seems to be related.\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign\n/triage accepted"}
{"_id":"q-en-kubernetes-f87c181995b6fb7b5af1d861ef84bc8e51724c56f863e5e28b19cdeaf999d0b1","text":"When PreFilter returns UnschedulableAndUnresolvable, we don't need to run the preemption, but currently we run. This unexpected preemption could cause errors from the preemption by trying to read non-exist PreFilter data. As pointed out (), actually we didn't run the preemption in such cases in the past versions, but introduced this bug. /kind bug /kind regression /sig scheduling /triage accepted /priority important-soon\n/assign\nwill fix this."}
{"_id":"q-en-kubernetes-a2ec6849a5f6979dcd970b9c7d0ac1edb3fa3728f7d794bb7e4bbe45e9cad1c8","text":"introduced configuration of additional logging aspects (e.g. format) through command line flags and config file. This was merged for Kubernetes 1.28. Now the following invocation doesn't work as expected when the config file doesn't configure logging: It should run with verbosity 4. Instead it runs with verbosity 0, from the config. Run with this config file: No response 28 rc n/a No response No response No response No response\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThe problem is caused by this line: At that point, is 4 (from the command line flag), and then it gets overwritten with == 0. The normal approach would be: set up config with defaults modify config through flags modify config further by parsing a config file into it. Can kube-proxy do this? It has the downside that a config file can only overwrite fields if they are in the file. A config field doesn't get reset to defaults just by parsing the file. /cc\nit looks like kube-proxy special-cases the to let a flag override config... can/should it do the same thing for verbosity? are there any other flags like this?\nI don't think that's the normal approach... it's more normal to make flags and config file mutually exclusive. That's difficult for things which started as flags and moved into config\n/sig network /sig instrumentation /kind bug /kind regression /milestone v1.28\nlet's go with and warn it and comment as much as possible, I was going with something like but Patrick is working on it so I assign to him the issue\n/assign\nHey folks, v1.28 Bug Triage Lead here Just wanted to know if this is already tracked by SIG leads and someone already working actively fixing it and set a priority?\nPerhaps I should have said \"this is how component-base/logs was meant to be used\". Somehow it worked for kubelet. Here's my preliminary solution: It needs more testing, but I wanted to share it early because I am running out of time today.\nit is , I can take over Patrick's work in case is needed\nIt would be nice if the logs API could distinguish between and , how about make to a pointer? But this change is supposed to be a breaking change, I'm not sure of the scope of the impact. For side, if set and at the same time, should the command flag override the config file or warn it and be discarded? I think it should be overridden.\nxref:"}
{"_id":"q-en-kubernetes-28269bd85efc1b89fc648fca571d244061293e095713f655cd09a9d77d170752","text":"Drop the and dependencies from to keep the import lighter. /sig auth /triage accepted\n/assign Hi, , I created a PR to fix this. I am not sure If I am correct due to the repo seems generated. Would you like to provides some hints for me?\nThis will likely require a new go module in to move the client-go dep and then mark it as internal, ex: . is the only API machinery dep so we can probably come up with some way to re-implement the wait logic locally? Then we drop these lines so that these imports do not come back again:\nWhoa, yeah, I'd take a for loop and a time.Sleep in a test to drop those deps\nWhat is kms using client-go for? Tests only?\nThese were all needed by the KMS repo back when we had a much more complicated reference implementation which required a key hierarchy. Now they are just used in some tests and mock implementations, so we should be able to simplify things.\nI found a place for rate limit. #liggitt\nthat package just delegates to \"\", switch that mock provider to use that directly?"}
{"_id":"q-en-kubernetes-5c3ec69bec85c4facfe873c7db9df815adb6cd2a20925ee9b3982c8822c51399","text":"There are some cases when kube-proxy has to exit and restart, in order to do that we panic, but this has the problem that the log output shows a literal \"panic\" message and engineers, support, customers and automation gets confused, thinking that is a problem when is the expected behavior. As suggested by in , we can replace by /sig network /help /kind cleanup\nThis request has been marked as needing help from a contributor. Please ensure that the issue body includes answers to the following questions: Why are we solving this issue? To address this issue, are there any code changes? If there are code changes, what needs to be done in the code and what places can the assignee treat as reference points? Does this issue have zero to low barrier of entry? How can the assignee reach out to you for help? For more details on the requirements of such an issue, please see and ensure that they are met. If this request no longer meets these requirements, the label can be removed by commenting with the command. <detailsIn response to : There are some cases when kube-proxy has to exit and restart, in order to do that we panic, but this has the problem that the log output shows a literal \"panic\" message and engineers, support, customers and automation gets confused, thinking that is a problem when is the expected behavior. /help /kind cleanup </details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign I would like to raise a PR to fix this."}
{"_id":"q-en-kubernetes-b1abe0806b508755225ef99bbfe83fce127939ca0134e636d8618b9e9c58b24c","text":"found this by running tests on this PR: it's problematic for a couple of reasons: prevents kubeadm from moving out (eventually) there are no API guarantees in the given code/structs/functions in pkiutils the logic in cmd/kube- must be reworked to generate cert/key pair without using kubeadm. originally in pending fix PR\n/sig api-machinery cluster-lifecycle architecture /area code-organization\nThe label(s) cannot be applied, because the repository doesn't have them. <detailsIn response to : /sig api-machinery sig-cluster-lifecycle architecture /area code-organization </details\n/assign\ncc\nPlease make sure to add an import restriction to prevent a regression in the future.\n/triage accepted"}
{"_id":"q-en-kubernetes-a2cd3b868d94a163d9bd5ad5403e25ff088564b704a1d680828637bdbf135eda","text":"master-informing: ! 2023-09-07 03:13:59 +0000 UTC No response /sig network\nChecking this failure This is running ingress-gce version, logs in failure is It seems the healthceck interval is not updated correctly, by reading the test tagging the people with more expertise in this area /cc\n/cc\nit seems the test restarts the kube-controller-manager but we landed recently the switch from intree to outtree cloud providers, so this logic now runs in the cloud-controller-manager, can this be the problem, that we have to restart the cloud-controller-manager now instead of the kcm?\n/triage accepted This flake started after and kept failing after 09-07.\nyeah, the logic that handles the loadbalancer shifted to that cloud controller manager with that PR"}
{"_id":"q-en-kubernetes-8aad58699f7cf1d09fbc7e0dbeb3b8213bb2879a529ad817ea0c8582cc0d9652","text":"With the migration to podutils and adding some presubmits (from periodic jobs) we are seeing some failures in presubmits. The error seems to have issues finding the right node to create.\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig node /sig testing\nany leads on solving this one?\nsee notes on slack -\n/assign\nthis looks good now!"}
{"_id":"q-en-kubernetes-298e26f0ee91a3c94155f58998ca31c83fd7d5f5b2b25cd901c9193d3861c8da","text":"Two serial tests from sig-node are keep failing: E2eNode Suite. E2eNode Suite. pls see: [1] [2] <fmt.wrapError | 0x40005e7840: failed to update pod \"pod--65da-42b7-a6a5-\": Pod \"pod--65da-42b7-a6a5-\" is invalid: spec: Forbidden: pod updates may not change fields other than ,,, (only additions to existing tolerations), (allow it to be set to 1 if it was previously negative) core.PodSpec{ Volumes: []core.Volume{ { Name: \"cm\", Name: \"secret\", VolumeSource: core.VolumeSource{ ... // 3 identical fields AWSElasticBlockStore: nil, GitRepo: nil, Secret: nil, Secret: &core.SecretVolumeSource{SecretName: \"secret\", DefaultMode: &420}, NFS: nil, ISCSI: nil, ... // 9 identical fields FC: nil, AzureFile: nil, ConfigMap: &core.ConfigMapVolumeSource{ LocalObjectReference: core.LocalObjectReference{Name: \"does-not-exist\"}, DefaultMode: &420, }, ConfigMap: nil, VsphereVolume: nil, AzureDisk: nil, ... // 7 identical fields }, }, }, InitContainers: {{Name: \"init1\", Image: \"\", Command: {\"sh\", \"-c\", \"sleep 5s\"}, TerminationMessagePath: \"/dev/termination-log\", ...}}, Containers: []core.Container{ { ... // 9 identical fields ResizePolicy: nil, RestartPolicy: nil, VolumeMounts: []core.VolumeMount{ { Name: \"cm\", Name: \"secret\", ReadOnly: false, MountPath: \"/config\", ... // 3 identical fields }, }, VolumeDevices: nil, LivenessProbe: nil, ... // 10 identical fields }, }, EphemeralContainers: nil, RestartPolicy: \"Always\", ... // 28 identical fields } happened since 18/10/23 run the e2e testcases manually No response <details</details<detailsImpacts all </details<details</details<details</details<details</details<details</details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign /cc\ncc\nfailed on x86 as well:\n/sig node\n/unassign\nis merged today. /cc\nwill take a look at this ASAP /assign\nand I are sorry for the issue this caused. We decided to open to fix this issue. I think this test was more complicated than we thought. Sorry for this."}
{"_id":"q-en-kubernetes-36ece6e0d4d3e56e17172a578777fceddb06fa51b38c6f565eb56f0705ab0c8c","text":"CI always has error when using to replace in function in file . No error occurred Use to replace in function in file . No response <details</details<details</details<details</details<details</details<details</details<details</details\n/sig api-machinery /area test\n/triage accepted /assign /assign"}
{"_id":"q-en-kubernetes-6eb2ec72e99977fe71ddbf32d8013bca881a978a092aa71ce0789cddc080ca46","text":"The controller-manager node-lifecycle-controller watches node lease and update probeTimestamp , and it check node probeTimestamp is before . The node-monitor-grace-period default value is 40s. When the client connection lost, the lease watcher cannot receive new event. With the http2 health check, HTTP2PINGTIMEOUTSECONDS default is 30s, HTTP2READIDLETIMEOUTSECONDS default is 15s, it will delay 45s to kill http2 connection. During this period, node-lifecycle-controller will mark all nodes NotReady. Here are some logs When the http2 connection lost, the watch should stop as soon as possible. The reflector should relist resource, and node-lifecycle-controller should not mark all node NotReady. HTTP2PINGTIMEOUTSECONDS+HTTP2READIDLETIMEOUTSECONDS should be less than 40s. the controller-manager node-lifecycle-controller watch client srcIP:srcPort. Drop all tcp package send into srcIP:srcPort. for more than 40 seconds, all node will be marked NotReady node will be Ready after some seconds. No response <details</details<details</details<details</details<details</details<details</details<details</details\nHow about double-check node lease before marking node NotReady? The node-lifecycle-controller can get reference lease from apiserver with\n/sig node\n/sig api-machinery\n/cc We have discussed this timeouts values before and myself, I also think got some metrics with lower timeouts for the HTTP2 connections. In this situation is clear that the defaults of the node watcher and the timeouts are related and having good defaults will be ideal, but the question is which values do we change ?\nI think in internal network, the ping response almost return in 1 second, the HTTP2PINGTIMEOUT_SECONDS default value should be 5 seconds. And the total timeout will 35 seconds. WDYT?\n/triage accepted any pointers in other places where those timeouts were changed for http2? Any notes from those discussions?\n/priority important-longterm since it is not a regression I will keep it longterm priority and not higher. But we can reevaluate if we hear more reports about it.\n/assign\nI'm investigating problems similar to this, but I'm looking for a way to confirm this is what is happening - and if there are tips for how to recreate such a scenario, that would be very helpful. We get infrequent reports of \"worker goes NotReady for a short time\" incidents that appear to be triggered by events like failing over a LB sitting front of the kube-apiservers. We've done some tests with a simple client-go app and have recreated a scenario in which the app logs messages after 45 secs of \"context exceeded\" errors. The failures reported to us from customers don't have http2 errors in the kubelet logs, and our attempts to recreate with kubelet have never shown more more than a single error which works on subsequent retry.\nHi folks, we have a simulated recreate of the problem. This recreate involves injecting a 40 second delay in \"connectivity\" between the Kubernetes API server and etcd (we a 40 second delay to etcd defrag and defragged all etcd members at the same time). The connectivity between these two components is obviously critical to nodes successfully updating their leases in order to stay ready. We use 40 seconds because it matches . The test will usually cause one or more of the cluster nodes will flap their readiness state. Setting on the Kubernetes controller manager prevents the flap. Given this, I think that one or more of the Kubernetes default configurations needs to change in order to tolerate such \"connectivity\" problems in order to maintain healthy node state.\nWhat's the proposal then? Do we update --node-monitor-grace-period to 50s? It seems the less risky change as the client-go defaults are more sensitive and impact the whole ecosystem, beyond kubernetes itself are you open to increase this default? It seems worth it to improve the reliability of cluster\nI think we should update .\ncan someone send a PR to see how it will look like?\nI have been following this discussion as we had encountered the bug as well. I have created a PR - Please let me know if it works\nI'll add another data point from our testing. Namely, the simulated recreate noted depends on the Kubernetes controller manager (KCM) leader election configuration. Using the following leader election configuration for the KCM yields a KCM that has no leader after the delayed etcd defrag. As a result, nodes won't be marked not ready unless the KCM leadership is restored before the nodes renew their lease. Extending the KCM leader election configuration to the following, makes the problem easier to recreate.\n/assign Reassigning the issue to me"}
{"_id":"q-en-kubernetes-e42c9a050473ded8c69ccec55cd354933899b995476973c8d00cf808ad43e070","text":"The issue is to follow up discussions in : running conformance test with the commonly used tool, sonobuoy, failed because the tool's test runner is lack of permission towards . The original issue was resolved by changing the e2e (conformance) test of to an integration test. Several problems remained: privilege requirement in is outdated. discovery is one of the least-privileged permissions, and is expected to be granted to every authenticated user but is not actually. the result returned by should be part of conformance? 1st and 2nd are straightforward to fix, not sure about 3rd, should we add a separate test or add a verification step in existing one like With the fix of 2nd, e2e for should no longer fail with any test runner. cc - should be reflect the accurate permission requirement discovery permission should be granted to all authenticated users. test should check the result returned by . Inapplicable No response <details</details<details</details<details</details<details</details<details</details<details</details\n/sig api-machinery /sig testing\nI agree with this. We have conformance tests around things like webhooks and custom resource definitions. I'd recommend rephrasing that as \"limits itself to capabilities exposed via APIs (e.g., does not require root on nodes, access to raw network interfaces) and does not require write access to system namespaces (like kube-system)\" I'm not sure I agree with this. No binaries Kubernetes provides (that I'm aware of) and nothing in client-go needs to make use of root discovery information. I'd be very slow to expand endpoints included in . Unless we have compelling reasons for people to use and rely on this data, I would lean towards no.\ncan we fix 1 , that seems that everybody agree and looks like the most clear one? the other sound more contentious\nthanks for your comments. Created for 1 and closed the other one.\nThanks all. Agree on this outcome.\n/triage accepted"}
{"_id":"q-en-kubernetes-bf1b99b6343620ce13074ea27d87d38db69960e482b5a066e7dc780b2ee01e0f","text":"When installing the metrics-server to implement the API and the prometheus-adapter to implement the API, it is possible to make the OpenAPI handler fail to initialize with this error log in apiserver: This will result in kubectl discovery calls to /openapi/v2 to fail with a 503: The OpenAPI handler should not fail to initialize if both and are serving the same path for different API groups/versions. I can provide exact installation steps if needed, however, a high-level repro is: install metrics-server to serve install prometheus-adapter to serve ensure resourceRules is populated in the prometheus-adapter config so that resource metrics are also served by restarting apiserver will result in the OpenAPI failure and 503 responses to /openapi/v2 path No response <details</details<detailsAWS </details<details</details<details</details<details</details<details</details\nPlease let me know if I should reopen the issue in\n/sig api-machinery\n/sig instrumentation [metrics apis]\nWould you mind taking a look when have time? Thank you^^ /triage accepted\nI have a question. should be serving under . Why does an apiservice serving another gvk register paths under ? cc we've made the assumption that an APIService (such as metrics-server or prometheus-adapter) serves only gvks under its own registered group version. I'll do some more investigation but just wondering if prometheus doesn't follow that assumption?\nthe prometheus-adapter can also be configured to serve and replace the metrics-server by enabling these metrics, However, if the prometheus-adapter fronts and metrics-server fronts , it does seem like the correct behavior would be for prometheus-adapter to serve until an operator updates the APIService to reference prometheus-adapter rather than metrics-server. We're seeing this issue occur for upgrades between 1.27 and 1.28 so it does look this path duplication did not cause the OpenAPI handler to fail on 1.27.\nI just upgraded from 1.27 to 1.28 and running into this problem. Some kubectl apply operations fail, and ArgoCD cannot load its custom resources. Any suggested workaround? Would removing the from the prometheus configuration resolve it? Or perhaps disable metrics-server?\nSimply looking at the code diff between 1.27 and 1.28 I can't pinpoint the cause. Can I get your exact repro steps? suggests that the conflicting paths should be removed before the merge. We always call this function with\nOne other question is if this is reproducible by just applying a set of yaml/installations or does it require a cluster upgrade to trigger? Thanks!\nhit same problem and we cannot install any other new helm plugins. maybe we need to move forward\nHere are verbose repro instructions. An apiserver restart on 1.28 will induce the issue.\nHello. I'm still having some trouble reproducing although I'm not 100% sure if the configuration is correct. I'm trying to repro via kind. afaik installing via helm has prometheus adapter serving . seems to suggest that prometheus adapter is serving under . Applying the update to the configmap per the instructions didn't change the apiservice on my end. Not sure if this is my misconfiguration or intended. I've also tested via integration test that having two openapis serving the same path doesn't cause errors and one will get filtered out. A simple workaround is probably to delete the metrics-server if it is already superseded by the prometheus-adapter.\nThe is not set up in an resource for us, we only created an APIService with . When I go fetch the openAPI schema served up by the endpoint, it has schema information for in it if that is enabled using . If I remove then the schema fetched from the server doesn't have stuff in it, and the error is resolved. We have been keeping around because I think something uses it directly, or used to. Maybe ? That could have changed But in any case, this problem is quite severe when it happens so it would be nice to fix it as there may be others who do the same update and have to do some deep diving into the logs and all kinds of crazy diagnostics and guesswork to figure it out.\nHow does the get reached without an APIService? The resources would be unavailable for the apiserver to query and it'll be through something else? I see, adding resources for me doesn't populate the OpenAPI. I wonder if this could be cloud provider specific. I'm totally supportive of fixing, it's just quite difficult to reproduce on my part .\nHmm good question. In our staging cluster I tried bringing back the in the custom metrics server configuration and restarting all the masters, and the problem did not come back, everything seems OK now. So it maybe isn't that easy to reproduce\nAfter updating the configmap, restarting the prometheus-adapter pod to pickup the changes and restarting the apiserver are both required to induce the issue. I have been able to deterministically reproduce this following those steps, please follow up here if you are still having issues.\nThat's good to hear. I'll need some more help with reproducing it. I've been able to verify that restarting the prometheus adapter pod picks up the configmap changes but restarting the apiserver hasn't caused any crashes for me. I've tested this both on the latest master branch and 1.28.\nWhile I'm yet to fully understand these changes, the 2 PRs below (both of which went into 1.28) seem likely candidates for a regression here: - introduced lazy loading in openAPI v2 handler to load/compute the aggregator spec when requested by client - switched the spec merging function from to any clue if one of the above changes could be causing mischief?\nAlso from example request at the top, the impact seems to be to the v2 API: Nathan - when you get a chance, could you also confirm if v1 API doesn't run into this?\nOpenAPI only supports v2 and v3.\nhi, any clue about this issue?\nIt's possible but I haven't been able to reproduce it. Can I ask if this is happening on cloud providers outside AWS?\nHere are the apiservices and relevant pod objects with a working repro from these steps If it's useful, I can provide yamls for these objects in addition to the settings applied in the prometheus-adapter configmap. Can you confirm that the pods backing your prometheus-adapter and metrics-server are running?\nI meet this problem\nWe the problem in GCP as well.\nACK in Alibaba Cloud also meets this problem\nWe were able to fix the v2 availability by and re-installing the kube-prometheus helm chart. Unfortunately, after restarting the k8s api server, v2 availability is broken again.\nyes, I'm running on kind. I'll do some more digging, assuming it's related to those PRs linked by but still haven't pinpointed anything.\nI tried to reproduce on my local setup with based on steps + , but I could still get a successful response when I executed . Then I reverted one of the step mentioned in to and waited until the new pod came alive and then did a apiserver restart and I was able to reproduce the same thing on as well i.e; I get the 503 response when I execute following command after the fact. After this point, even if i change it back to and restart the apiserver , the error is persistent. Let me know if this works for you to reproduce ? PS: Also, i want to call out that there is no issue with , its .\nThanks, alright I found a way to reproduce it consistently without restarting the apiserver. Still investigating the root cause. Thanks for the detailed reproduction and\nQuick update: This issue was a combination of bugs both in k/k and prometheus adapter. I opened to mitigate on k/k side which should fix the issue. The issue was difficult to reproduce because it requires a very specific ordering of OpenAPI merging to trigger, which is why I suspect many people were able to reproduce after some restarts to kube-apiserver. Prometheus and metrics server serve differing OpenAPIs for the endpoint. Prometheus also doesn't seem to be registering discovery's OpenAPI properly (missing OpenAPI for but having the OpenAPI for whereas it is all present in metrics server. This inconsistency in how prometheus and metrics server handle the OpenAPI surfaced the merging issue.\na little wired question: why do we need to restart kube-apiserver to reproduce this problem?\nTried to reproduce the issue after cherry picking the fix and didn't run into below error. Fix is working as expected.\nIt can look like restarting prometheus-adapter \"fixes\" the issue? I am also curious about suggestion of removing the section from the config map works as a temporary workaround until released upstream (we use and cannot delete the apiregistration)? As I understand the fix for this issue is currently in"}
{"_id":"q-en-kubernetes-4a10fef5c556bfe966c5a6671d3a6c7531b89b8bdf8d0e3a308063720e57299f","text":"We should check the images configured in test/e2e to ensure we're not adding new images: In almost all cases, should be extended to support any missing functionality and used as the sole image for new tests. This has been the policy in SIG Testing for at least a few years now. To prevent regressing on this, we should check the list of images and permit tag updates but NOT adding any more names to the list. We can do this with an allowlist against /sig testing /kind test Others are welcome to work on this, but I'll submit something soon if nobody beats me to it. /assign Follow up to , we should aim to ship something this release (1.30).\nThe label(s) cannot be applied, because the repository doesn't have them. <detailsIn response to : We should check the images configured in test/e2e to ensure we're not adding new images: /kind test /assign Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nHello! Release Signal shadow here. This issue has not been updated for a long time, so I'd like to check what's the status. The code freeze is starting (less than a month from now), and while there is still plenty of time, we want to ensure that each PR has a chance to be merged on time. As this issue is tagged for 1.30, is it still planned for this release?\nWe still intend to do this and it's a test-only change so we're aiming for test freeze.\nHello! Release Signal shadow here. I'd like to check what's the status. The code freeze is starting in a week, and we want to ensure that each PR has a chance to be merged on time. Is it still planned for this release?\nAs has no update recently, /milestone clear we may make it in v1.31.\nYeah I think we have to revisit this later."}
{"_id":"q-en-kubernetes-d0e2e132009f7ccf51c7803031ff8956361c074983e993ee4f6b6ad076c145a0","text":"Part of From For now posted as a \"help wanted\" issue, however if we cannot hit the timeline for 1.30 code freeze () I would take it over so we don't miss. Still issue is simple enough that it can be quickly implemented. Implementation tips: Add a metric around waitUnitlFreshAndBlock method Metric should be a histogram with latency buckets similar to api latency Metric should have label for resource version. A way to monitor watch cache impact on request latency.\n/sig api-machinery /triage accepted /help\nThis request has been marked as needing help from a contributor. Please ensure that the issue body includes answers to the following questions: Why are we solving this issue? To address this issue, are there any code changes? If there are code changes, what needs to be done in the code and what places can the assignee treat as reference points? Does this issue have zero to low barrier of entry? How can the assignee reach out to you for help? For more details on the requirements of such an issue, please see and ensure that they are met. If this request no longer meets these requirements, the label can be removed by commenting with the command. <detailsIn response to : /sig api-machinery /triage accepted /help Instructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details"}
{"_id":"q-en-kubernetes-c802e9333b6135a07c5679f05136949c916f8d3805cbf17c61ae7577c1b78003","text":"07-02 and 13-02 No response No response /sig\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig scalability\nI might be related to the npd.\nThe same test failures happening again on 25.02:\n/sig node\n/sig node /cc The restart of caused the flake. I also opened to track it in NPD repo.\nI missed this log. It is caused by oom. Adding some memory may help. I submitted to fix it. Does mean killing metrics-server?\n/remove-sig node Sorry, I misread the log. The NPD found the metric server is OOM. It is about the metrics server, not the NPD. Sorry for disturb. I may add some memory for metric-server and check why metric server memory changed. I heard that metrics server recently released a new version. I will check if it is related.\n/sig node\nOOM metrics shows the metric server is OOMed. This is odd, that only metrics-server-nanny is with a memory limit. It uses whose command is . metrics-server container has no memory limit, but it got killed by OOM. Does this indicate that the node is OOM? (edited)\nThe last run failed in node-problem-detector ContainerStateTerminated. I0227 05:03:31. ] PodStartupLatency: perc50: 14m29.018886827s, perc90: 28m44.625518807s, perc99: 32m2.19303496s; threshold 1h0m0s I0227 05:03:32. ] Loaded restart count threshold overrides: map[coredns:2 fluentd-gcp:999 konnectivity-agent:2 kube-controller-manager:1 kube-proxy:2 kube-scheduler:1 l7-lb-controller:1 metadata-proxy:2 prometheus-to-sd-exporter:2] I0227 05:03:32. ] ClusterOOMsTracker: gathering cluster OOMs tracking measurement E0227 05:03:32. ] TestMetrics: [action gather failed for SystemPodMetrics measurement: restart counts violation: RestartCount(node-problem-detector-tgl8h, node-problem-detector)=1, want <= 0; RestartCount(node-problem-detector-vpsmz, node-problem-detector)=1, want <= 0] I0227 05:03:32. ] Step \"[step: 29] gathering measurements\" ended W0227 05:03:32. ] Got errors during step execution: [measurement call TestMetrics - TestMetrics error: [action gather failed for SystemPodMetrics measurement: restart counts violation: RestartCount(node-problem-detector-tgl8h, node-problem-detector)=1, want <= 0; RestartCount(node-problem-detector-vpsmz, node-problem-detector)=1, want <= 0]] I0227 05:03:32. ] Waiting for the chaos monkey subroutine to end... I0227 05:03:32. ] Chaos monkey ended. I0227 05:03:32. ] I0227 05:03:32. ] Stopping DnsLookupLatency probe... I0227 05:03:47. ] Stopping InClusterNetworkLatency probe... I0227 05:04:22. ] Resources cleanup time: 50.539700215s E0227 05:04:22. ]\nE0227 05:04:22. ] Test Finished E0227 05:04:22. ] Test: E0227 05:04:22. ] Status: Fail E0227 05:04:22. ] Errors: [measurement call TestMetrics - TestMetrics error: [action gather failed for SystemPodMetrics measurement: restart counts violation: RestartCount(node-problem-detector-tgl8h, node-problem-detector)=1, want <= 0; RestartCount(node-problem-detector-vpsmz, node-problem-detector)=1, want <= 0]] E0227 05:04:22. ]\nI0227 05:04:22. ]\nI0227 05:04:22. ] Running testing/huge- I0227 05:04:22. ]"}
{"_id":"q-en-kubernetes-6ea734899e2e6d4c6f94619e9e5de5ead043d20b171880b17d80d7fdee6414a0","text":"Kubernetes 1.27 with feature gate enabled and disabled, a rebooted node is not able to re-start pods that were running there before the reboot. The node reports in Pod events: The Pod should start just fine. On Kubernetes 1.27: NewVolumeManagerReconstruction feature gate. a StatefulSet that uses a volume. -reboot the node which runs the StatefulSet Pod. The reason is that kubelet's VolumeManager reconstructs the volume as uncertain during startup and it uses a provisional outer volume name for it, because it has not seen the API server and Pods in it yet. This provisional volume ID is then used in ActualStateOfWorld forever. While kubelet does mount the volume after reboot, it's not able to match it with the Pod that needs it. Kubelet reports , while the volume is mounted with ID in my case (= the provisional name, we use for it) and not (=). It is not reproducible with enabled. <details</details<detailsAWS + AWS EBS CSI driver, but it should not be relevant. </details<details</details<details</details<details</details<details</details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/sig storage /priority important-soon This is reproducible when SELinuxMountReadWriteOncePod is disabled and NewVolumeManagerReconstruction enabled. They both got enabled by default as beta in 1.28. So someone has to explicitly disable SELinuxMountReadWriteOncePod to hit this bug."}
{"_id":"q-en-kubernetes-a15d3bc2d753fcd2973b0177ed8ca7ba1f5b5835a5489e4aaf40994db346bde5","text":"Extension API servers do not currently receive the header when requests are proxied through kube-aggregator, so spans are not linked properly. Traces from the extension API sever should be linked to the parent in kube-aggregator. tracing for kube-apiserver tracing for sample-apiserver and aggregate it to kube-apiserver a request to kube-apiserver for one of the resources from sample-apiserver sample-apiserver traces will not be linked to the kube-apiserver traces No response <detailsThis problem currently exists on . </details<details</details<details</details<details</details<details</details<details</details\n/sig api-machinery\n/cc /triage accepted"}
{"_id":"q-en-kubernetes-fc1ee346cd39a038cd7031c22da95da4e006a404986ee7addd6bd858e6211da0","text":"We use a to nil out the : However on a re-sync, this results in a data race that was observed when running unit tests: The read had occurred previously during a callback. On re-sync, the retrieves every object from the cache, invokes the transform function, and re-queues them. However, the object instances are the same as those stored in the cache so mutating them outside the cache without lock protection is unsafe. It is documented that the \"TransformFunc sees the object before any other actor, and it is now safe to mutate the object in place instead of making a copy\" however this is not the case on a re-sync. The should make a copy of the objects retrieved from the cache store on a re-sync or perhaps elide the on a re-sync since the objects would already have been transformed. As a workaround, I changed our to only nil the if not already nil but this is a latent bug that others will likely run into.\n/assign That means that and introduced possible panic regressions into kube-controller-manager and kube-scheduler /priority critical-urgent /kind regression\n/transfer kubernetes\n/sig api-machinery\nkube-controller-manager and kube-scheduler do the same in 1.29+ and likely have the same issue\nI think delete has the same problem, and replace may have the same problem: Passing the same object to the transformer after it is already in the cache is unsafe and we should not do that at all.\n/assign /cc\nLooking into this. I may take time to come up with a proper fix. But would like to evaluate the risk for the current situation. Not entirely sure how delete and replace would be triggered in normal settings, but resync is fortunately low frequency as the default setting for resync period is between 12h and 24h for (--min-resync-period duration Default: 12h0m0s, see ) . As for the resync function is Just think out aloud, I read a thread discussion somewhere that resync was a mechanism in the past to prevent data inconsistency due to bugs but client-go has been mature these days, resync is no-longer needed in theory but for a safe guard. So maybe disabling resync is a reasonable fix if delete/replace can be fixed.\nYou're mixing resync with relist - I believe relist is no longer needed. Resync is actually used. FWIW - I opened an initial version of the fix:\nI think it's resync. The thread I mentioned is\n/triage accepted"}
{"_id":"q-en-kubernetes-12672a947a189075f869393abb114d8dd9861f80c842304ad0f60e558628a295","text":"The test is failing on Windows [0]: The idea of the test is to check that kubelet still follows and reads the Container Logs after a Log Rotation occured. However, because the followed file is open, it cannot be renamed, and thus, the container logs cannot be rotated. This issue could be prevented if the file would be opened with the [0] [1] Kubelet should not block container log rotation. Run the test ` is failing on Windows. /sig windows /kind bug /kind failing-test latest, master. This unit test has been failing for a while. N/A Windows nodes </detailsN/A N/A N/A\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/kind failing-test /sig windows"}
{"_id":"q-en-kubernetes-241517f08a7efb475af9790477ced0c9347cd6c91a1010c631b20abf4116466d","text":"in This flaking test case was newly as part of the above PR Flake was reproducible with stress tool: [root kubernetes]# go version go version go1.22.2 linux/ppc64le [root kubernetes]# stress - ^TestLog /sig testing\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\n/assign\nLooks like in this test case is a concatenation of logs from Pod0 and Pod1, hence it is failing in some cases. I will add a PR to split it."}
{"_id":"q-en-kubernetes-960ed57813038a78afde8a2b57c7a0085de88a198005a7e7fec097ee46c20dfe","text":"Currently, the scheduler honors Unschedulable(AndUnresolvable) or Pending only from PreFilter, Filter, Reserve, and Permit (WaitOnPermit). It doesn't expect other extension points to return those statuses and just regard all non-success status as errors, that is, immediately retrying scheduling of Pods. Actually returns Pending at PreBind, which is currently being ignored as described. So, we should expand UnschedulablePlugins/PendingPlugins to include PreBind plugins. (or maybe all extension points too?) /cc /sig scheduling /kind feature\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nBoth seems reasonable to me.\nBecause is not handled for , the following error is logged:\nThis patch gets rid of the error. I am uncertain whether it has other implications? Performance was the same with and without it.\nWhat I imagined is like that. It shouldn't have any other impact because volumebinding, which is the only PreBind plugin apart from DRA in in-tree, only returns success or error status. We have to mention the change in the release note for custom plugin developers though.\nI've turned this into a PR:"}
{"_id":"q-en-kubernetes-fa36b1c7fd0c09951628b1c964121d1a9d9ac4ca11720b4cb13858994caee5cb","text":"/kind failing-test <!-- If this is a flake, please add: /kind flake --/sig node\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nFailures look like started when landed"}
{"_id":"q-en-kubernetes-f4d4b7c473a266ffc48b535d9080ecbdaed52af3a9bd64395678a7e260a39c7f","text":"for a PodFailurePolicyRule. However, when trying to create a JobSet programmatically with a JobTemplate that defines a PodFailurePolicy rule for failing a job immediately if the failed pod had condition DisruptionTarget, I get a validation error indicating that cannot be null. I've tried using both and (with matching versions for package as well). Error: JobSet spec: JobTemplate passes validation. JobSet: applying example JobSet in the description of this issue. JobSet: and applying example JobSet in the description of this issue. No response 30 (tests on a kind cluster using kindest/node:v1.30.0) <details</details<details</details<details</details<details</details<details</details\nThere are no sig labels on this issue. Please add an appropriate label by using one of the following commands: - Please see the for a listing of the SIGs, working groups, and committees available. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\ncc\n/assign\nI confirmed the issue, moreover, I see that when the JobSet webhook is disabled then it works ok, I think it must be the JobSet webhook which adds the \"onExitCodes\" field during deserializion /serialization step. I can also see that when I rebuild JobSet using kubernetes field \"onExitCodes\" with \"omitempty\", then it the JobSet creates ok. I will open a PR."}
{"_id":"q-en-kubernetes-8e580bebcf4ffab3cbf419f47076a234c2ba43c46f10aadd238cf6766de367fc","text":"After merged we are seeing the new test in fail since it was turned on be default now (it passed in the PR linked because it didn't have the alpha feature enabled). <br class=\"Apple-interchange-newline\"<!--EndFragment--</body /sig windows /sig node /cc\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nptal"}
{"_id":"q-en-kubernetes-ee3812360963e95275f0bcb6ea7d2b46134d316f1573489c5f3be9de533696cb","text":"July 24th. I think there is some suble way the test is named. Test is covered by and that works. I think this is mainly an issue where this test should not be run with these jobs. /sig node\nThis issue is currently awaiting triage. If a SIG or subproject determines this is a relevant issue, they will accept it by applying the label and provide further guidance. The label can be by org members by writing in a comment. <detailsInstructions for interacting with me using PR comments are available . If you have questions or suggestions related to my behavior, please file an issue against the repository. </details\nwhat is the correct way to label a job now? This is apparently wrong and causing the job to be picked up in the unlabeled name.\nYou still need to have a corresponding feature in because, as you noticed, there are jobs which don't support skipping based on the feature gate.\nI created an issue in test-infra. once all tasks are completed, it's okay only to use framework.WithFeatureGate. There may be a lot of jobs that need to be updated. for now, you have to add a feature label or modify all the failing jobs.\n/assign\nxref"}
{"_id":"q-en-kubernetes-9f06ace95e768b9e28a2e012f6881c788c69b5cf718e21e17fbc090aed8a934a","text":"Related to and SIG storage meeting notes: From me: It is related to image volume plugin and selinux. Is it a known limitation of all the volume plugins when using selinux? I have a different opinion. If two pods don't share the same mountpoint (hostpath), they can run the same node when the selinux is enabled. the current implementation of cri-o is sharing the same hostpath when two pods use the same image volume. More details about my idea: From Sharing each volume between namespaces is kind of natural. Sharing PVs not possible but sharing image quite common. From Although for most volume plugins it is true that, they can't be mounted with 2 different selinux context on same node, we want to understand why is this not possible for Imagevolume source. For a traditional storage persistent storage plugins - this limitation is underlying limitation of storage backend/Linux kernel etc. For ImageVolume source, it seems like a simple solution which unpacks image to two different hostpath would work and not run into this limitation. can you specify, why this limitation exists for ImageVolume plugins? From we choose to keep it simple and follow the lead of other volume types. supporting multiple selinux types would mean making more complicated the way we mount the image into the container AFAIR. But agreed we can chat about this elsewhere, in the meantime for the purposes of this PR: this is consistent with behavior of other volume plugins and is expected behavior. Let's discuss the SELinux behavior for image volumes in this issue. cc /sig node /sig storage\nIn my opinion, from an end-user perspective, the image volume may act like a read-only with a pre-populated content via image. Two pods from different namespaces can share the same image and run on the same node.\nPasted from According to , the containerd doesn't have the same limitation.\nGiven that OCI/Image Volume Source support is in alpha, I'd prefer that we consider discussing this as part of the original feature. We can discuss this here but I don't know if we need a separate KEP for this yet.\nYes, so far, in our conversations at the containerd community meetings we decided to start this feature out using a single pod scope for the pod scoped image volume host path because in containerd each runtime handler may have a different snapshotter (possibly on a different fs), or at least that's the plan. Additionally, the pods may have different os versions on windows indicating a different platform can be selected on the same node and this sort of versioning may extend beyond to other os providers. We could ask Kirtana to confirm the versioning issue and status of platform version selection over the platform index manifest, if the snapshotter reason isn't enough to at least start with per pod scope. If this is insufficient support we can certainly discuss. Cheers, Mike\nNo need a separate KEP. We can update the original KEP once this issue is resolved.\nI'm working on a technical solution (just a PoC) in , not sure how feasible that is as time of writing, though.\nyeah I agree I don't think we need a different feature for this, especially because it's mostly an implementation detail in the runtime. would you open this as an issue in kubernetes/kubernetes instead of the enhancements repo, or even better in the cri-o repo?\n/transfer kubernetes\nI moved this issue into kubernetes repo.\n/triage accepted /priority important-soon\nthanks!\nCopid from kep: As merged, do we need to change the KEPs to reflect the new behavior?\nno, nothing needs to be changed from a Kubernetes any more :+1:\n/remove-sig storage"}