text stringlengths 0 2.02k |
|---|
**Adam Stacoviak:** That's the hard part though. You talked about Crunchy... What was the other one you talked about we moved to? |
**Gerhard Lazu:** The Zalando PostgreSQL. |
**Adam Stacoviak:** So you've got those two, and then you consider "Would Linkerd have helped us?" But that shows to me at least, from someone from this perspective, which is not a Kubernetes operator - I'm not an SRE - is that you have to have some sort of understanding of the different tooling available in the ecosys... |
**Gerhard Lazu:** Yes. |
**Adam Stacoviak:** And not to just know which tools are available to manage Postgres, like we need to, and replicate and whatnot, but also a high degree of understanding of those tooling and how they'll actually help you. I think that's just a very daunting, high-touch world that Kubernetes presents. It may be the fut... |
**Gerhard Lazu:** Yeah. So they do say - and let me be specific. Kelsey Hightower has been saying this for a long, long time - don't run your data services on Kubernetes, because things get complicated. And I think this is a first-hand experience of what he was referring to. Things may seem okay for a long time, but th... |
But all these things - first of all, they made us just understand the stack a little bit better; when I say "us", mostly me... And it made me realize that simple is best. So for the 2021 setup, we are running just a very simple, stateful set, single PostgreSQL instance that can restore from backup in less than one minu... |
\[20:14\] Let's say that you lose everything. If you backup frequently, which we do - every hour, by the way; and I have to change that setting. I've set it to be three hours, but I need to change it to one hour. It's super-simple. And then the database will back itself up every hour. We can lose an hour worth of data;... |
These backups, the reason why they were important is because when we did the 2021 setup, all I had to do - I had to let this system restore from backup, to pull all our media, which is 85 gigabytes right now; all the files, all the mp3's, all that stuff. So to download that from S3 is fairly fast, especially for mp3's.... |
But we can restore everything from scratch -- so let's say we delete everything. Within 27 minutes, because of all those small files, everything is restored. The database, super-fast; the media files, the whole lot. And because it's so simple, do you need to have a distributed system? You can use these local SSDs; that... |
**Jerod Santo:** Right. |
**Gerhard Lazu:** It shipped in November 2019, we started using it, it was the beta... And just when it went live I think in May, we were already starting to switch some production workloads across. And then by -- was it August or September? I can't remember. Everything was across. Something like that. |
So did we need a multi-node Kubernetes cluster? The answer is no. What we needed was proper CDA integration. And that's where the speed comes from. So by properly integrating with a CDN, in this case Fastly, the website is actually 15 times faster. The latency. |
**Adam Stacoviak:** Did you say 15 times? |
**Gerhard Lazu:** One five. Yeah. |
**Adam Stacoviak:** One five. |
**Gerhard Lazu:** Actually, let's do this... By the way, we are integrating with Grafana Cloud, so we ship all the logs, all the metrics to Grafana Cloud, and then we have synthetic monitoring set up there, we have probes running all around the world... By the way, not all probes are reliable, but we have plenty to sho... |
**Adam Stacoviak:** And we're monitoring our babies now. |
**Gerhard Lazu:** We are, yes. The feeds, and we have alerts, and reports... There's so many things we have set up, so thank you, Grafana Cloud. That's a really cool thing. |
**Adam Stacoviak:** Behind the scenes Jerod called our feeds our babies, so... |
**Gerhard Lazu:** Yes, he did... \[laughs\] |
**Adam Stacoviak:** A little joke there, but... Yes, we're monitoring our babies, which is our podcast feeds. |
**Gerhard Lazu:** Yes. And if a baby is crying, guess who gets the Telegram message? This is a Grafana Cloud integration. I do. |
**Jerod Santo:** That is the way it should be. |
**Gerhard Lazu:** Yeah, exactly. That's how you stand by your infrastructure, if you're willing to be woken up at night. And guess what - we're caching it... So caching doesn't go down anymore. All of Fastly would have to be down before Changelog would be down. So we have proper integration, which we didn't have before... |
Anyways, before we enabled caching, the Changelog website, the average latency - so we have San Francisco, Dallas, New York, London, Frankfurt, Bangalore, Sydney and Tokyo. These are all our probes. So the average latency across all probes was 818 milliseconds, before. |
**Jerod Santo:** \[24:06\] That's kind of embarrassing. |
**Gerhard Lazu:** Yup. Now it's 66 milliseconds. So how much is that? 880 by 66 - 13.3 times. Not quite 15, but not 10 either. It's more than 10. |
**Adam Stacoviak:** We can round it to 15. |
**Gerhard Lazu:** And guess what the uptime is? |
**Adam Stacoviak:** A hundred percent. |
**Gerhard Lazu:** A hundred percent, exactly. |
**Jerod Santo:** \[laughs\] |
**Adam Stacoviak:** A hundred percent. \[laughs\] |
**Gerhard Lazu:** It's a hundred percent, that's exactly right. |
**Jerod Santo:** All the nines. We want all the nines. |
**Break:** \[24:32\] |
**Jerod Santo:** So this speaks to really geographic relocation of our assets, right? We had all of our images and mp3's and CSS and JavaScript assets served via CDN, all the way back to when we've set this system up. |
**Gerhard Lazu:** That's right. |
**Jerod Santo:** But we didn't serve the entire website via that CDN. |
**Gerhard Lazu:** That's right. |
**Jerod Santo:** So even though Phoenix is really fast, even though we're set up good, and we even have in-memory caching in places where it makes sense, like the feeds... Because who wants to recalculate the Changelog's feed of 400-some-odd things every time it gets request... We cached that in the app. In addition to... |
**Gerhard Lazu:** Yup. |
**Jerod Santo:** There's not much we could do about that, while our responses were coming from a centralized, single pop, as they call it, point of presence... Which is the way it was. So now every request goes through Fastly, and... We should have done that a long time ago. |
**Gerhard Lazu:** We should have. \[laughter\] |
**Jerod Santo:** I'll take full responsibility on that one, because I kind of slept on it for years. |
**Gerhard Lazu:** Nah... |
**Adam Stacoviak:** I think you resisted it, actually. Didn't he resist it for a little bit? You were like, "No, let's not do that." |
**Jerod Santo:** \[28:15\] Yeah, I think it was -- |
**Adam Stacoviak:** I'm not trying to call you out or anything, I'm just trying to know what were the circumstances for saying no, really. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.