text stringlengths 0 1.8k |
|---|
**Gerhard Lazu:** \[51:47\] It exists. The OpenTelemetry integration exists in Erlang. It's not that mature, but it's improving. Every month it's getting better. And I think it's more around the queries that go all the way to PostgreSQL, so how does the request map to that. I know that the database has some impact on t... |
Now, maybe if PostgreSQL was like a managed service, we wouldn't have this problem. Maybe. But regardless what the case would be, you'd want to know what is the problem, and if I change this, does it actually improve it? And by how much? If you have the trace, it's really easy to understand "Well, I should not qproxy, ... |
**Tom Wilkie:** You've hit on one of the many problems with distributed tracing. You have to have the entire stack instrumented to get a lot of value. And if you have holes in the middle, or black blind spots from a kind of tracing perspective, then the value is greatly diminished. |
**Gerhard Lazu:** Yeah. |
**Tom Wilkie:** You can get tracing information out of load balancers, and I've never actually done it myself though. I've always kind of stopped there. I'm hoping that things like OpenTelemetry -- and I know Amazon are heavily investing in OpenTelemetry, so I'm hoping that it'll be possible (if it isn't already) to ge... |
I'm hoping that things like the W3C Trace Context makes this easier, and maybe this even allows things like the CDN, Fastly, to also emit a span. That would be kind of cool, being able to see a CDN and an ALB and your application. |
When it comes to Postgres and MySQL, I don't know. I'd love to see spans coming out of those systems, but I don't really know the status, and I'm not really an expert on this side of things. |
A common misconception is that every service emits one and only one span. It doesn't have to; you can emit as many spans as you like. You probably shouldn't emit too many, but you can do whatever you like. So one of the things we do a lot of is client-side spans. Whenever we do a request to a database in Cortex in pret... |
In your situation, because it's a monolith, I would instrument the Elixir server and client going out to Postgres, and that would probably give you enough information to know if it's Postgres, to know if it's the qproxy or the ELB. You wanna get a span from something further up the chain, and then start to look at the ... |
**Gerhard Lazu:** Ingress NGINX? Does Ingress NGINX and NGINX support spans, do you know? |
**Tom Wilkie:** I don't know off the top of my head. |
**Gerhard Lazu:** Okay. |
**Tom Wilkie:** One of the things I've definitely seen engineers go down this rat hole of trying to get complete traces and spans from everywhere, and there's just kind of a -- there's an effort/reward trade-off to be made. It might take a lot of effort to get a complete span from every single service. If you're on a m... |
**Gerhard Lazu:** I understand what you're saying. I think on the client-side that is less of an issue because this span, which is the longest one, happens server-side, where it's waiting - or processing, whatever the name may be - and that tends to sometimes be really long... So what happens inside of that span? |
We know that it goes to (let's say) Fastly. Great. We can remove that, and we can go directly to the load balancer... I don't think there's much we can do about the load balancer, so let's say we ignore that... So our span really starts at possibly the Ingress NGINX. So that's the first start point. |
**Tom Wilkie:** \[56:05\] Mm-hm. |
**Gerhard Lazu:** Excellent. What happens inside Ingress NGINX maybe would be interesting. I mean, this is NGINX specifically. Maybe it would be interesting. But the next hop will be into - as far as I know, this will be the entrypoint into Kubernetes. So that will be the service that is responsible for routing the tra... |
So having these first 2-3 steps in the span would be already helpful, but realistically, I think we can only start from the Kubernetes side, and that's okay. So from NGINX, the next hop would be really the application. So how does that span vary, and regardless what happens inside, it doesn't matter. How does that dura... |
So it's just like, okay, together, maybe seven steps, and which is the step which is more variable. That's the way I think about it. Is that right? Does this sound right to you? |
**Tom Wilkie:** With distributed tracing you've always gotta kind of see -- the great thing about it is being able to visualize the actual flow of the requests. So yes, I'm agreeing with you. |
One of the things I will say is it's probably not Kube-proxy. My understanding in most deployments is that it's not a layer seven thing; it's done at the TCP level, where it doesn't intercept any traffic, so it's not worth putting a -- or it's not even technically possible, I guess, to do a request-level span there, be... |
**Gerhard Lazu:** Right. |
**Tom Wilkie:** You know, one of the promises of OpenTelemetry, because it's so vendor-netural and because it's so open as a standard is that we might even be able to get spans into more established open source projects who don't wanna pick favorites. Maybe one day we will be able to get spans into Postgres and into My... |
**Gerhard Lazu:** Neither do I, but that's really fascinating. So this is what I'm thinking... First step is let's pair up on what it looks like to do Grafana dashboards Tom-style. I'll call it Tom-style; I know it isn't, but... Grizzly-style, or whatever. The point being is the way you develop them. Big fan of GitOps,... |
And then, this specific problem once we have that iteration set up really nicely, and those feedback loops that operate nicely, so we can experiment, which goes back to what you were saying, being able to ask interesting questions, being able to figure things out, like explore, which I'm a big fan of... Figure out, lik... |
\[59:16\] And then, I think those spans, Tempo and integrating with that - super-valuable, long, long-term. I expect things to change along the way as the ecosystem matures. More libraries are getting instrumented, OpenTelemetry becomes more mature... I think that's a great vision and a great direction towards where th... |
As a listener, if I had to remember one thing from this conversation, what should that be, do you think? |
**Tom Wilkie:** I'd go all the way back to the early comments about observability and about the big tent philosophy, and about there not being one-size-fits-all tooling. I know as a vendor here I have a preference for Prometheus and Loki and Tempo, but honestly, that's just a preference; that's just an opinion. An equa... |
**Gerhard Lazu:** I love that. So if we were to pick one title for this discussion, what do you think that should be? |
**Tom Wilkie:** Observability and big tent philosophy. |
**Gerhard Lazu:** Big tent philosophy. I like that. I like that big tent philosophy. |
**Tom Wilkie:** I'm not sure where the term comes from, to be brutally honest. I should probably google it. I know how a lot of companies have internal mantras. Google's mission was to organize the world's information. The internal mantra in Grafana Labs is this big tent philosophy. We apply it everywhere, to everythin... |
**Gerhard Lazu:** Who came with the idea of the big tent, do you know? |
**Tom Wilkie:** I don't know where the term came from, but the idea was very early on in Grafana, when Torkel added support for multiple data sources. And very early on -- Grafana started life visualizing Graphite data. But very early on, support for other systems was added. And it's really that vision early on to brin... |
**Gerhard Lazu:** So the big tent, the way I understand it, is bringing all these (I wanna say) vendors' data sources? It's more than just data sources, right? |
**Tom Wilkie:** More than just data source, because it's data from anywhere, and combining it in a single place, but building experiences that span multiple systems, integrating them in ways that didn't exist before... But it's not just a concept that applies to Grafana and the visualization. We apply it on the backend... |
**Gerhard Lazu:** Tom, I didn't think this was possible, but it happened... I have more questions at the end than at the beginning... |
**Tom Wilkie:** I'm sorry about that... |
**Gerhard Lazu:** And I'm more excited to continue talking with you at the end than I was at the beginning; again, I thought that's not possible... I'm really looking forward to trying things which I've just said, and I'm really looking forward to the next time, so thank you for today. |
**Tom Wilkie:** Thank you very much, Gerhard. |
• Definition of observability and its origins in mechanical engineering |
• Observability as a way to understand internal system state without prior knowledge |
• Instrumenting code to gather context and information about system behavior |
• Importance of gathering high cardinality dimensions such as user IDs and request data |
• Distinguishing between metrics-based monitoring tools for infrastructure and observability for understanding code behavior |
• Shift from monolithic architecture to microservices and serverless, requiring more robust observability methods |
• Applying observability principles to monolithic systems and focusing on end-user experience and journey through the app |
• Development teams spend more time with code than end users |
• CI/CD pipelines are essential for observability and reducing bug fixes |
• Observability helps engineers have a tight feedback loop by quickly seeing changes in production |
• Time is key: shipping to production as soon as possible (ideally < 15 minutes) reduces bugs and pathologies |
• Shipping into production is crucial, and testing in production is inevitable; the goal is to do it well with guardrails |
• "Dead code" isn't valuable unless it's in production |
• The number of engineers required to maintain codebase scales with time spent on maintenance, with 2x as many engineers needed for each doubling of time |
• Importance of minimizing wait times in engineering, especially when trying to solve new problems or ship changes quickly |
• Goal of achieving a 15-minute deployment cycle for most companies, with some variation depending on stack and complexity |
• Relationship between testing and shipping speed, with tests ideally running within the same short timeframe |
• Concept of "shipping" as not just deploying code, but also learning from it and getting answers to questions quickly |
• Example of Changelog.com's monolithic application and how to make it more observable using Honeycomb OpenTelemetry instrumentation |
• Visualization of data in Honeycomb, including features like BubbleUp for diagnosing problems and identifying differences |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.