text stringlengths 0 2.02k |
|---|
**Jerod Santo:** No, it's fair. I think it's because I didn't read the docs well enough and I didn't realize how easy it is to just bypass that if you have cookies set. Because I thought -- you know, we have signed in users, signed out users... I guess I always had done it that way; I just served the dynamic parts from... |
**Gerhard Lazu:** That's right. |
**Jerod Santo:** ...maybe a lucky 3%. Maybe 1% of requests are signed in people. So a little bit of ignorance, a little bit of just old-school "This is how I do it", and then because we didn't have worldwide monitoring, we had single-point monitoring, it always seemed pretty fast. We always got good scores... |
**Adam Stacoviak:** Is it good for you? It's good for me. |
**Jerod Santo:** Yes, exactly. Is it good for us? Is it good for people in the States? Once we've set up the Grafana with around-the-world monitoring, then you start to realize "Holy cow, this is not fast for everybody." |
**Adam Stacoviak:** Yeah. |
**Jerod Santo:** So I think it was just less important because I didn't realize how bad it was out there. |
**Adam Stacoviak:** Well, that's interesting too when you talk about observability. You don't know what you don't know until you know, or something like that...? |
**Gerhard Lazu:** The unknown unknowns. |
**Adam Stacoviak:** Basically, observability provided a lot of data to understand some of the problems because either you don't have time, or you not necessarily don't care, but you don't care because you can't care. You don't have the data to really understand the full-rounded picture of the problem or the concern. An... |
I'm curious though about the average... Because you said 880 was the average. Sure, the highest -- because that says average. What was the highest? |
**Gerhard Lazu:** So this is the average latency, and we have all the different points. Can you see that? |
**Adam Stacoviak:** Yes. |
**Gerhard Lazu:** Okay, cool. So this is all probes... |
**Adam Stacoviak:** We'll put a screenshot into the show notes for sure, but... |
**Gerhard Lazu:** So let's look - for example Dallas, which is closest to where Adam is. So in Dallas, what we're seeing is the average latency is 42.20 milliseconds. And that's a pretty good latency. You can see that you have a couple of high ones... The max goes to about 200 milliseconds. |
**Adam Stacoviak:** This is now, not before. |
**Gerhard Lazu:** This is the last seven days. So we're looking across the last seven days. |
**Jerod Santo:** If your maximum response time is 200 milliseconds, then you're sitting pretty. |
**Gerhard Lazu:** 200 milliseconds, exactly. And that's where the average -- and this is Dallas. Let's take London, for example, for me. London is 87 milliseconds, and the maximum is 400 milliseconds. Now, what we need to understand is that some of this is also related to probes. Do you see the uptime? It says it's 99.... |
\[32:22\] I know that when we were looking at Bangalore - I think that was the one... Was it the Bangalore? See, for example these errors here - this was the fourth of May. The error rate was very high. But all it meant is that the probe may have been overloaded. Not necessarily the website, because I'm pretty sure Fas... |
So we have a shield in New York, and then every other point of presence basically distributes from there. It reads it from that cache and it replicates across the whole world. And we have micro-cache, so we cache every response for 60 seconds, and then if there's any cache misses, it will continue serving stale content... |
Let's see... For example, can I see one probe here that was not very healthy? For example this one. This was Tokyo. Do you see how the latency went slightly high? So Tokyo was having not a great day. The Tokyo probe. Same thing here, in Bangalore. The Bangalore probe was all the way up to five seconds, so some requests... |
**Adam Stacoviak:** What does that mean, to be overloaded? Like, the Grafana probe - it's got a lot of logs it's doing for not just us, but others? |
**Gerhard Lazu:** Yeah. |
**Adam Stacoviak:** ...similar to the way a noisy neighbor is on a VPS? |
**Gerhard Lazu:** That's exactly right. Or whatever routes this is taking, the route is overloaded, the networking. We don't know what route it takes. So however this probe runs, we can see now -- we never had this. And this is a really fascinating thing. Who knows what problems we had in the past, in the 2021 setup, b... |
So now we know that, for example, users in Frankfurt, maybe -- maybe there's an interconnect that is slow. Maybe it's not just that probe, but still, we are able to serve within seconds most requests. So we monitor the NGINX logs, and we can see the response times, we can see the traffic served. This is, by the way, af... |
We can see that the 99th percentile, the average 99th percentile is 707 milliseconds. So we are under one second. This is NGINX to the app. But the time interval is ten minutes. So if we go to, let's say, five minutes - it's a lot. One minute - look at that. Whoa... What happened here? So when the time interval is one ... |
**Adam Stacoviak:** Well, now you know there's a problem though. That's the thing. Because before you didn't know there was a problem. |
**Gerhard Lazu:** Exactly. |
**Adam Stacoviak:** \[36:09\] And if we're dealing with replication of databases - and this was sort of like attached to that. As you begin to -- |
**Gerhard Lazu:** Here's the thing... All this runs on a single, massive host. We have 32 CPUs, AMD Epyc, 64 gigs of RAM, or 128 gigs of RAM, SSDs, super-fast... It's a single host. So how can the 99th percentile between Ingress NGINX running on that host, and the app, which is running on the same host, be this high? |
**Jerod Santo:** Bitcoin miner... |
**Gerhard Lazu:** It's not, but sure... \[laughter\] I assure you it's not. |
**Adam Stacoviak:** I'm glad you shared the specs of that server, because that does put it into context of-- |
**Jerod Santo:** It should never happen. |
**Adam Stacoviak:** ...its capability, and that this shouldn't happen. |
**Gerhard Lazu:** Nope. It shouldn't happen. |
**Jerod Santo:** Well, what do you surmise? What's your gut? |
**Gerhard Lazu:** Something in queue proxy. I mean, that's the only thing. It's not the database. It's not the app. It's something between all those components that make up Kubernetes. We have Calico for the CNI. Maybe it's that. Maybe it's the overlay network. But this is where that -- it's almost like you want more o... |
If you look at the external monitoring, everything looks good. Everything is fine. From a CDN perspective, things are okay. And that is the experience that we want to give our users. The website is always available, it's super-fast regardless of where you are in the world... And these are the things that we are now bec... |
**Jerod Santo:** I feel like on the detective shows is where they say "Zoom in and enhance." That's what you're doing to us here. We zoom in -- and at a certain point, you zoom and enhance, and you just can't enhance any further. You're staring at a blob, and you're like "I don't know what that is." |
**Gerhard Lazu:** Yeah. |
**Jerod Santo:** That's kind of where we're at, so you're gonna need another level. You need another zoom or another enhance in order to dive down. The smaller these problems are, the more time you spend figuring out how to get that zoom done, and probably your ROI (so to speak), or the law of diminishing returns hits ... |
**Gerhard Lazu:** That's right. |
**Jerod Santo:** I mean, ignorance I guess was bliss, except for our users it wasn't bliss. We thought it was fast everywhere, and now we know that it wasn't. It's better, and yet we still have this little thing that's like "What?! What's going on there?" |
**Gerhard Lazu:** Yeah. And it does happen fairly frequently, by the way; so there's something there. Would tracing help? I don't know. If we look at the last six hours, we have a spike here; that was 7 PM. And they're not periodic. They happen -- like, 4 PM... Could it be the database backups? I mean, they do run ever... |
But then you have all these smaller spikes. This is 1 PM, so not really... You had these spikes... And again, most of this stuff, if you look at the traffic that we serve, it's nothing. The server is not even 1% loaded. CPU is not an issue, and network is not an issue. Nothing is an issue. All the components are health... |
**Jerod Santo:** \[40:21\] Wait, wait, wait... What's this "Ship it" you just said? What's this thing? What are you talking about? |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.