text stringlengths 0 1.49k |
|---|
Previous to that, we were kind of hand rolling some of our own mutual auth TLS, and while the TLS library and standard library is pretty solid, there's some kind of gotchas and rough edges, so gRPC kind of handles all that for you. |
We were kind of already using protobufs, so it was a natural fit. |
With gRPC you generate all of the clients for all the different languages that it supports, so that was like a huge win; we wouldn't have to maintain a lot of clients... They would just be generated. |
I think that's about it. There were some other reasons why we kind of adopted it... |
**Andrew Poydence:** Yeah, it's also really good at upgrade paths... Cloud Foundry has to be HA above all things, so upgrade paths are always something that we're considering. Every change we make, we'll talk about "How will this upgrade from a previous version? How long can we keep this thing before we can deprecate i... |
**Jason Keene:** That is something that when you're using gRPC - and protobufs, for that matter - you kind of have to consider your upgrade paths. It's something that you have to spend some time considering; it's not something you get for free... But it definitely provides a clear path for you, and it's something that ... |
**Erik St. Martin:** \[36:10\] Right, because when you get to the actual protocol that things are communicating over, it's extremely hard to change it, because then everything basically has to come down and back up at the same time, unless there's kind of well thought out upgrade paths. |
**Andrew Poydence:** Yeah, that's not a thing in Cloud Foundry. Loggregator has components and agents that run on every single VM in Cloud Foundry, so... There's customers that have deploys that go for days, because they're rolling so many VM's, so your VM that rolls from the beginning has to still be compatible with t... |
**Brian Ketelsen:** That's crazy... Multi-day deployments? |
**Jason Keene:** Yeah, unfortunately there are some war stories of multi-day deployments, with massive Cloud Foundry deployments. |
**Andrew Poydence:** Yeah, just enormous data center upgrades, and stuff. |
**Erik St. Martin:** I feel like I've gotten spoiled. I was trying to deploy a Kubernetes cluster with a single command the other day and I was getting frustrated it was taking minutes... \[laughter\] |
**Brian Ketelsen:** This deploy took eight minutes, it's a bug! \[laughter\] |
**Jason Keene:** I mean, some operation teams appreciate the slow, methodical approach. They're not as impatient as us developers; they prefer a stable system that upgrades cleanly and without incident, than a system that upgrades quickly, with massive incidents. |
**Brian Ketelsen:** Operations people... \[laughter\] How dare they want stability?! |
**Andrew Poydence:** It's tricky. |
**Erik St. Martin:** We could have like a crazy show if we just had an ops person on, and we just fought back and forth... \[laughter\] |
**Brian Ketelsen:** Fight me! Run fast, before I break things! Then that would be the title of our show. |
**Jason Keene:** Yeah, it's amazing that there's still kind of two camps... Like, even with DevOps being a thing, there's still people who kind of fall within two camps. The question that I always ask people is if you had a production system that has, say, a memory leak, and restarting the process will fix the problem,... |
**Erik St. Martin:** Well, I think you left out what impact it's having on the customer, too... |
**Jason Keene:** Yeah, obviously there's a lot of variables there. |
**Erik St. Martin:** Yeah, because if you could take the server out of rotation and let the other nodes handle it while you poke at it, then... |
**Jason Keene:** Yeah, this is more like a hypothetical situation. I'm sure many of the people on the call have been in situations in prod where you've gotta bring back to service... But it's kind of a hypothetical situation where you have to make that choice, either or. |
**Erik St. Martin:** Well, usually a lot of us developers - we get an email that they've already restarted it... \[laughter\] And you're like, "But, but..." |
**Andrew Poydence:** And then you're instructed to go fix the problem. |
**Erik St. Martin:** Yeah, you're just combing through code, looking for anywhere that could leak memory... |
**Jason Keene:** \[40:00\] "Something happened in prod... We don't know what. Fix it." |
**Erik St. Martin:** Yeah. I mean, I think we're still trying to find the perfect balance between ops and dev and things, right? And I think a lot of the initial motivation, at least from my standpoint, for the whole DevOps movement is to understand how your code would be deployed and to help deliver your code as a ful... |
**Andrew Poydence:** We on Cloud Foundry have embraced that pretty heavily through [Google's SRE](http://www.oreilly.com/free/site-reliability-engineering.csp?cmp=tw-security-books-videos-lp-promo_srebook_lp) that O'Reilly released not too long ago... |
**Brian Ketelsen:** Good book. |
**Erik St. Martin:** Yeah, I still need to read it. |
**Andrew Poydence:** It's a very light read. |
**Erik St. Martin:** I've got some playing time this weekend, so maybe I'll get to it. |
**Andrew Poydence:** It's fantastic, though. There's a lot of things that are common sense, but it's nice to formalize it, and then as a team you can kind of use it as the decision maker - "How should we approach this?" or "How would [SRE](https://en.wikipedia.org/wiki/Site_Reliability_Engineering) have done it?" |
**Jason Keene:** Yeah, it gives us like a shared vocabulary now, so we're referencing the same concepts, versus before you kind of had to build up that shared vocabulary. |
**Andrew Poydence:** Right. An interesting thing that we have here at Pivotal working on Cloud Foundry is not only are we running in hundreds of different data centers, completely different configurations everywhere, but we also have our own massive Cloud Foundry running [run.pivotal.io](https://run.pivotal.io/) that w... |
**Jason Keene:** Speaking about operations stuff, we're all software engineers, so we don't really get into the operations stuff very heavily; we are on call for our production environment, but... Recently, as part of our partnership with Google, we went to a CRE, which is their Customer Reliability Engineering program... |
**Andrew Poydence:** [The nines of reliablity](https://en.wikipedia.org/wiki/High_availability#%22Nines%22) have been really nice. How to measure success on a distributed system is kind of a tricky concept. It can seem like it's working to you, but then you have a different consumer that they're not getting what they e... |
**Erik St. Martin:** \[44:00\] Yeah, and then I think you have the flipside of it, which is anytime you give somebody a metric on which to measure something, it's "Well, it has to be five nines. There has to be 100% test coverage...", you know? |
**Andrew Poydence:** Right... |
**Jason Keene:** Yeah, that's one of the things the CRE guys kind of pointed out really early - setting your SLA and SLI at the right level. Having four or five nines is not necessarily something you should strive for. Having a lower level of reliability gives you more flexibility on how fast you can iterate, and it ha... |
**Erik St. Martin:** Then I think when you talk about reliability or security or any of those things, it's always tradeoffs, right? You can't focus all your time and money on all the areas, there's just not enough time. You have to look at what the likelihood of this failing or being compromised is, and then you have t... |
**Brian Ketelsen:** The SRE book is such a good read for that, because they talk about having a budget. All of those nines are just digging into a budget. You've got a budget for downtime, how are you gonna use it? |
**Jason Keene:** As a streaming service, we kind of measure things a little bit differently than a typical request/response type service. We measure how many nines we have of message reliability, which is basically how many messages we drop. So we don't give hard guarantees of durability or delivery. We wanna make sure... |
The way we test that is we have a pretty aggressive black box testing of our system, so we have various environments that are running our software, and then we kind of probe it from the outside, we ask for a specific quantity of messages, and based on how many we get, that gives us our amount of messages that we've los... |
**Erik St. Martin:** I think we've got a few minutes left, so we should probably jump into projects and news and \#FreeSoftwareFriday and all of that good stuff... |
**Brian Ketelsen:** It was kind of a quiet news week, wasn't it? |
**Erik St. Martin:** Yeah, I don't know about a whole lot of news. Mostly product releases, and stuff like that. [Heptio](https://heptio.com/), if you are a Kubernetes and Envoy fan - they just released something called [Contour](https://github.com/heptio/contour), which basically allows you to use Envoy as the Ingress... |
**Brian Ketelsen:** Written by the legend, [Dave Cheney](https://twitter.com/davecheney). |
**Erik St. Martin:** Yes. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.