text stringlengths 0 2.02k |
|---|
**Jerod Santo:** Gotcha. |
**Gerhard Lazu:** And Kubernetes right now I would say it's the late majority. It's not laggards. You can still not do Kubernetes. But I think it's the late majority now. So we waited for it long enough before we went into Kubernetes... I would say we were towards the end of the early majority that adopted it. That's w... |
Our hosting provider, Linode - because that's how it all started... "Let's get some VPSes" - remember those days? And then VMs, and then cloud instances... |
**Adam Stacoviak:** Yeah. |
**Gerhard Lazu:** So they offer a managed Kubernetes service. And that was the thing which we were waiting for, so that we wouldn't need to worry about the control plane, about Etcd, and certificates, and the integration with \[unintelligible 00:10:37.28\] So all that stuff was abstracted away from us. Once we had that... |
So the components were fairly mature. There were so many blog posts, and use cases, and mistakes that have already been made before us. And what we wanted to do was fairly standard. So there's nothing crazy. The documentation was written... We weren't those early adopters. We were like towards the late early adopters; ... |
So a lot of the stuff made sense, and it was easy. Now, having said that, we still hit a couple of interesting things. |
**Jerod Santo:** It doesn't mean there wasn't any pain, right? |
**Gerhard Lazu:** Yeah, we still hit a couple of interesting things. Shall we go into that? What do you think? Some interesting things that we've hit? So some interesting things that we've hit were around the PostgreSQL operators. We chose PostgreSQL Crunchy first, and it was fairly hard to work with it because of how ... |
**Jerod Santo:** Right. |
**Gerhard Lazu:** So we had a replicated PostgreSQL, and we had downtime because it was replicated. You wouldn't expect that to happen. |
**Jerod Santo:** \[12:12\] Because it wasn't replicated, right? |
**Gerhard Lazu:** Because it was replicated. We had downtime because it WAS replicated. |
**Jerod Santo:** I thought it stopped replicating. |
**Gerhard Lazu:** Exactly, it stopped replicating. |
**Jerod Santo:** Okay, so it wasn't replicating... \[laughs\] |
**Adam Stacoviak:** Which one was it? \[laughs\] |
**Gerhard Lazu:** No, no, no. Hang on. So we had the replication in place. The replication stopped working... |
**Jerod Santo:** And it took down our primary system. |
**Gerhard Lazu:** It filled up the writer headlog, it filled up the disk... |
**Jerod Santo:** Right. |
**Gerhard Lazu:** It went down. The secondary was way, way behind, so it couldn't be promoted to primary, and we had downtime. |
**Jerod Santo:** Right. And we had data loss. |
**Gerhard Lazu:** And we had data loss. |
**Adam Stacoviak:** Yeah, we did. |
**Gerhard Lazu:** Oh, yes. |
**Jerod Santo:** That's way worse than downtime, in my opinion. |
**Gerhard Lazu:** Oh, yes. We had the backup from like six hours ago, or...? |
**Jerod Santo:** It was like six hours. |
**Gerhard Lazu:** Nine hours ago, or something... It was like a bunch of hours and we've lost some data, yes. |
**Adam Stacoviak:** Thankfully it wasn't a ton of data, but it was definitely a data loss. |
**Gerhard Lazu:** Because we had backups... |
**Adam Stacoviak:** That's the lucky part on our side. |
**Gerhard Lazu:** Because we had backups, yeah. |
**Jerod Santo:** We had good backups, but yeah, six hours back... Thankfully -- was there any podcast episodes that were published during that time? I don't think there was an episode. That would have been a bigger problem. |
**Adam Stacoviak:** Something was... |
**Jerod Santo:** There was just news items and comments... |
**Adam Stacoviak:** Mm-hm... |
**Jerod Santo:** ...and a few things where I had edited a thing and I had to go back and edit it again. Thankfully, we caught it fast enough that I remembered, and we're a small team, so we remembered our data loss. We were like "I know what I did yesterday", or the last six hours... So we fixed it up. |
**Adam Stacoviak:** In a larger team that would have been catastrophic. |
**Gerhard Lazu:** Yeah. That was not cool. That was really not cool... And you go through the documentation, and it's not like "Do this" or "Do that." You don't have a list of simple steps to follow. And then you're scrambling, and it's like "I just need to get this thing back up." That's all we cared about. And what w... |
I think this goes to show that it has not matured that much. I mean, it's getting there, but it hasn't matured that much. And if you need that type of redundancy from PostgreSQL, then - well, you either have some DBA chops, especially when it comes to PostgreSQL and you know what you have to do, or you're just paying f... |
But the interesting thing is, I always thought that maybe Crunchy was too complicated... And then we tried the operator, the Zalando one, and the same thing happened. So it wasn't an operator thing. And here's the thing - we still don't fully understand where the latency is in the Kubernetes networking stack, but we kn... |
It was a surprise to me - and I remember looking at this for a really long time, and thinking "Could it be Linode's private networking?" And it wasn't. That wasn't the problem, even though it indicated there's some network latency. So we went down to a single Kubernetes node, everything was running on the same node, an... |
So how many metrics would we need to enable in the different layers of the stack, and how well would we need to know that stack to debug this issue? I think that's where a lot of people that hit issues with Kubernetes - that's where they're coming from. You wouldn't expect these -- these aren't normal problems; these a... |
\[16:16\] But I do hope that some technologies -- I think they've been around for a while, but again, it goes back to how do you pick and choose your components. So what I'm wondering is would Linkerd have helped with this? Could Linkerd show us the latency between the different services, and to see when it spikes? |
**Jerod Santo:** What is Linkerd and how would it do that? |
**Gerhard Lazu:** It basically intercepts all the traffic between -- imagine Ingress NGINX when it talks to the app, Linkerd would place itself between Ingress NGINX and in this case the app. So we'd see all the latency between the two components. |
The same way it would intercept all the traffic between the app and the database, the PostgreSQL service, to show us when there's any sort of weird latency between the two services. |
Now, we could enable all the metrics for PostgreSQL, but you need to find the dashboards, then you need to understand those dashboards, if you have Grafana or something else... Then you're literally becoming a DBA. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.