text stringlengths 0 1.8k |
|---|
**Gerhard Lazu:** No? You're not using that. And do you know why? I'm curious, genuinely. |
**Charity Majors:** I'm not sure, actually. |
**Gerhard Lazu:** And then why Kafka? I have to ask that. For other reasons. |
**Charity Majors:** \[27:54\] Since we are writing our own storage engine, it gives us like 18 hours' worth of backup. You know, if we need to replay some events, or if anything happened... It's also how we bootstrap and bring up new nodes... |
**Gerhard Lazu:** Why not Kinesis? |
**Charity Majors:** At the time, I was the one who made that decision; this was like five years ago... And there were some constraints Kinesis had that I think had to do with the events and some of the data types that we needed, that it just wouldn't support; it wasn't flexible enough. |
**Gerhard Lazu:** Okay, so it's a Kinesis limitation that was there in the past, and it doesn't matter whether it's there now... Obviously, you have Kafka, it's running well, I'm assuming... |
**Charity Majors:** Honestly, ideologically, while I do believe in outsourcing, making it someone else's problem whenever possible, given that Kafka is basically functioning as part of our database which is a very integral part to Honeycomb; it is one of things that I think is better for us to have in-house expertise a... |
**Gerhard Lazu:** Okay. So that answers my next question, which is if it's a managed service or if it's something that you install, you manage, you update... |
**Charity Majors:** Yeah, we install and manage and update. |
**Gerhard Lazu:** How is that experience, I'm wondering? |
**Charity Majors:** Kafka? |
**Gerhard Lazu:** Yeah. Some people say that managing Kafka, like installing Kafka clusters used to be difficult. With Zookeeper I think that is going away these days... I don't know. |
**Charity Majors:** You know, not many startups have an ops co-founder... \[laughs\] And we were fortunate enough to have me. So that stuff is not that hard if it's your bread and butter. |
**Gerhard Lazu:** Right, okay. And then it's S3 behind the scenes, it's Lambda... So I'm assuming that when the Honeycomb UI is displaying those charts from all those events, you're actually consuming those events from S3, is that right? To draw the -- |
**Charity Majors:** For the first few hours... It depends. It's dynamic based on your write throughput etc. But it gets written out to SSDs first, and then it gets aged from there into S3. So yeah, it's reading from some combination of the local SSDs and S3. |
It was interesting - when we moved from using SSDs for everything to age things out to S3, we really thought there would be a severe performance hit... It turns out no. The performance characteristics are different, but -- and speed is incredibly important to us, because we really want people to be in the zone, just tr... |
**Gerhard Lazu:** Right. Okay. |
**Charity Majors:** Quite fast. |
**Gerhard Lazu:** Yeah, that sound fast. I mean, we were mentioning 15 minutes before, and now you're telling me one second... So yes, it is very fast, if you have that reference point... Okay. And I'm assuming that Honeycomb uses Honeycomb to understand Honeycomb. Is that right? |
**Charity Majors:** Definitely. \[laughs\] We have office dogs... Honeycomb was actually first named Bloodhound, and then we shortened it to hound.sh, and then we got a cease and desist from Hound CI... So now we're named Honeycomb. But Retriever is the name of our database, and Poodle is the name of our frontend, and ... |
**Gerhard Lazu:** And what monitors the kibble cluster? |
**Charity Majors:** Nothing. \[laughter\] |
**Gerhard Lazu:** You do, right? Are you working alright?" "Yes" is the answer. Okay. |
**Charity Majors:** \[laughs\] Yeah. |
**Gerhard Lazu:** Right. So does the dogfood cluster run a different version of Honeycomb? |
**Charity Majors:** Well, it's interesting you bring this up. We are deployed from cron, like every ten minutes. And it first deploys to kibble, and then it waits some amount of time, and if everything is okay, then it promotes to dogfood, then it waits and then it eventually promotes to production. And all that happen... |
**Gerhard Lazu:** So how long does it take for that to make it to production? |
**Charity Majors:** About an hour, three tops. |
**Gerhard Lazu:** So that means you don't deploy to production first. You go to kibble first, and then dogfood, and then production. |
**Charity Majors:** We consider that production. |
**Gerhard Lazu:** Right, okay. That's what will become production once everything is okay on kibble. |
**Charity Majors:** Yeah. |
**Gerhard Lazu:** \[32:07\] Okay. And how long do you keep things on kibble before promoting to dogfood? |
**Charity Majors:** It's about an hour. It's about an hour from kibble to dogfood, and an hour from dogfood to production. |
**Gerhard Lazu:** Did you find that helping? Did you find it helping having kibble and then dogfood before production? |
**Charity Majors:** Oh, yeah. Absolutely. |
**Gerhard Lazu:** So everybody makes mistakes, even the best ops people in the world. Is that what you're telling me? |
**Charity Majors:** Absolutely. \[laughs\] |
**Gerhard Lazu:** Good. I think that makes a lot of sense. People think some ops people are demi-gods... |
**Charity Majors:** No...! No. |
**Gerhard Lazu:** Everybody makes mistakes, but we fix them so quickly you don't even know. And we don't let them -- I think this is propagated everywhere. We trust the system, and the system has all these gates built-in. |
**Charity Majors:** We never trust \[unintelligible 00:32:51.09\] |
**Gerhard Lazu:** Right. Or engineers. |
**Charity Majors:** Or engineers. No. Why would we do a thing like that? \[laughs\] |
**Gerhard Lazu:** So I think this brings us to the software development being a socio-technical problem. People are fallible, they will make mistakes, and a lot of the time it is about those mistakes, which - think of them like learning opportunities. And if you think of them like that, then you optimize for learning; ... |
**Charity Majors:** Yeah. |
**Gerhard Lazu:** And what else would you say about this? Because I know it's a term which is very dear to you. |
**Charity Majors:** Yeah. Well, I think that people have this image of like "Oh, you hire a Google engineer and suddenly your team will get better", or something. No, I think that it's pretty clear that any engineer who joins a team, within 36 months or so, will be shipping and performing at the level that that team pe... |
**Break**: \[34:56\] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.