[denysvitali]

At KubeCon Europe a very good chunk of booths were observability stacks. Everyone was claiming they're better than the competitors (with some of the just justifying themselves by saying "it's written in Rust).

Having dealt with Prometheus (+Thanos) / Grafana / OTEL and other stacks (e.g: custom solution on ClickHouse, Victoria{Metrics,Logs}, Jaeger/Tempo, Loki, ...) and even cloud ones (Google's Monarch rebranded as Prometheus)... what's your selling point? This to me seems like yet another way to re-invent the wheel.

If it's just for running locally, okay, fine, but when it comes to production (where the stack really matters) at scale, you end up with lots of tradeoffs and approaches.

Why is this one a winning one compared to the overwhelming "competition"? Seems like we're re-inventing the wheel for the 100th time instead of focusing on unifying the efforts in making the existing solutions better. Thankfully we now have OTEL, so at least the interoperability part is somewhat solved (or mitigated)

[yuppiepuppie]

I was thinking this might be a result of the Cheap-money (post covid) era ending and everyone scrambling to reduce their Datadog/Cloud costs. Thinking back on 2023/2024, lots of companies were leaking large amounts of capital to those vendors and I imagine lots of people saw an opportunity for creating leaner and cheaper stacks.

[dusanstanojevic]

No need to guess, I'll tell you the exact story of why I made Traceway!

Last Dec I had a customer complaint, took me 2 days to find the issue. I had to pay $800 for Sentry and a bit more for New Relic. The issue was a locking problem that happened only in very very specific cases, erroring in diff places and timing out in others, unfortunately power users were running into it often. I had two systems, no SLO to catch this and they were completely disconnected. Super annoying.

Anyhow, I spent a day looking at those and eventually went, screw this, I'm gonna just make this actually work. So I spent a few hours, hooked it up, no auth or anything nice, pulled the traces and found the issue. Turns out it was locking due to a long transaction existing in a scheduled task, it existed for years.

The big things for me is it automatically flagging issues, prioritizing them and taking into account: errors, response codes, timing. That's why I'm making it, no venture capital, funded by actual revenue from the start (not paying for Sentry or New Relic anymore). It's really a dev focused tool to help smallish teams find and fix issues before customers even have time to complain.

Anyhow, hope that explains it, kinda related to cloud costs, mostly just my personal frustration with existing tools. Also I did NOT want to host a 5 service stack (grafana, otel collector, prometheus, mimir, loki, k8s) for something that can be done in a 60mb go binary that runs on a 3$ server...

[robertlagrant]

This is my instinct too. I've had the pleasure of using DataDog and the pain of negotiating with their salespeople!

[Ocha]

Yes. Their sales people don’t even negotiate - they just tell you this is price and done. Dunno why they need sales person if prices are non-negotiable

[SOLAR_FIELDS]

It’s because you’re in a leveraged position. Why would they negotiate when they don’t need to. Tell them thanks and that you’re churning tomorrow and watch the “OH WAIT”’s come flying through the door. The insidious thing about datadog is that it snakes its way into your entire business line and so it’s really hard to extricate yourself from it down the road.

[dusanstanojevic]

Hi, creator of Traceway here. Sorry for the late response, I didn't know this got posted and then my account was rate limiting on comments.

A lot of tools in this space, most pretty good. The goals when I started Traceway were: - simple to host and reason about - cheap to host - comes pre configured for sub 15 dev teams - completely open source, no paid ad-ons

It's not aimed at teams that can afford SREs (yet), the idea was to provide a good tool for smaller teams and startups in the sub 15 dev range.

The base of Traceway is Clickhouse, nothing special there, if you want you can run it with sqlite for self hosting. Sessions are also stored in S3 so the costs are minimal.

It is opinionated, it comes with preconfigured SLOs for flagging issues with endpoints and it will never try to sell you an AI SRE, you can file your exceptions/slo issues with the git integration and run what ever AI you want on it (I was sick of observability tools trying to sell me an AI). The goal is to have a one line setup, for OpenTelemetry, that gets you everything you need in Traceway without anything needing to be additionally configured. It's Datadog/Sentry but combined and fully open sourced.

I'm a huge fan of open source, here is what we've done so far for making existing solutions better:

1 - Session Replays/RUM

Session replays are usually a premium/expensive feature. With Traceway you can self host them and add them to your app in minutes. I am working on making this a standalone feature that ties into the otel sdks for mobile/js so that you can get your spans/logs/metrics/exceptions from any platform connected to your session replays in Traceway. At one point I got nerd snipped into making it work with Flutter, so we are the only solution I know of that has affordable usable session replays for Flutter.

2 - Symfony Otel

Symfony, the php framework, had no library that offered a few line setup and worked out of the box with open telemetry. We wrote one, you can use it with any tool out there.

3 - Symbolicator

We're working on a symbolicator that will be Open Telemetry Collector compatible, so that you can get your stack traces for Js/Flutter/Android/iOS resolved back. From what I can tell no good solution exists for this currently.

I will make a proper HN post at some point with more info on the project, right now I am focusing on building. If you have any ideas or things you'd like to see feel free to comment, join our discord community or open the issue in our git, we're always happy to accept PRs.

[NortySpock]

If I can ask a separate question: what scalability problems did you run into with Victoria{Metrics|Logs|Traces}, and at what scale did you hit them?

VictoriaMetrics and Logs have worked fine in my quiet homelab, and VictoriaMetrics appeared to work great for the infrastructure team of an open source online video game I contribute to (say about 10 physical nodes and 20 applications/services ) ... I was going to suggest VictoriaLogs to them next but wanted to ask what roadblocks could come up.

[dusanstanojevic]

I honestly think you are a bot. When ever I see Victoria mentioned it is always the same, always asking about hitting a scaling problem + promoting it, never responding to any comments. Hope I'm wrong, but it's been one too many. I refuse to use a product that is this dishonest.

[hagen1778]

I work at VictoriaMetrics.

Just to clarify: VictoriaMetrics doesn't use bots for HN or for any other media for promotion.

I don't know the person who you responded to. Most of the activity you see is coming from community members who genuinely use the project or from the core engineering team trying to answer user's questions or address misunderstandings.

> never responding to any comments

Could you please share examples like this? I can't say for community members, but our internal policy for engineers is very much focused on great support. You can check our slack/github to see that every question is answered and well explained.

[dusanstanojevic]

Hi, first of all, thank you for your response, I really appreciate actual Victoria team commenting.

This is probably the 5th comment (almost identical) I have seen about VictoriaMetrics, mostly on Reddit. I engaged with a few trying to learn more about your product and eventually just gave up. If you really want you can comb through my reddit comments, but be warned, I have commented on a lot of things... a lot...

You should be proud of what you have built, I've looked a bit more and your product looks incredible. I personally think that a sales person might have been testing an automation tool, but if it was actual customers that just shows how good the product is!

Traceway is not working on addressing the problems y'all are solving, it is more focused on having an out of the box experience with preconfigured dashboards, SLOs, integrations, automatic endpoint ranking, frontend session replays/RUM, symbolication etc.

Again, thank you for your comment.

[NortySpock]

Hi, I am not a bot. Also I do not work for VictoriaMetrics.

Please feel free to go through my post history and observe I comment on things I am interested in, like databases, servers, and video games.

[dusanstanojevic]

Got it, sorry, your comment just looked like a bunch of others and felt extremely out of place as nobody mentioned hitting any limits, especially with the Victoria stack (that I could see).

The comment read out of place/generic and given my previous experience I incorrectly assumed it was another generic bot - my bad.

Hopefully no hard feelings and Victoria looks great.

What do you like the most about it and how was your experience scaling it?

[NortySpock]

No worries, no hard feelings, I was just surprised that what I thought was a specific response was assumed to be a generic-ish bot response. (then again, I didn't spell out that the game I contribute to is Beyond All Reason.) I do totally sympathize with the feeling of being overwhelmed by AI slop.

After digging into Traceway documentation, it looks like you were looking to primarily use OTEL for ingestion? Or would you say that's a misreading of the documentation and you actually support metrics, logging etc easily? It looks easy to setup via docker, I might try the SQLite version just to get a taste for how it works and how easily data can be ingested.

For myself, I was initially interested in the Loki/Prometheus/Grafana stack but it wasn't going to fit in the 4GB of RAM I had available on a Raspberry Pi that was already hosting two services that consumed a GB of RAM each. So when I found VictoriaMetrics (a) happily ran in 200MB of RAM (b) was used by CERN (c) had excellent, comprehensive documentation with plenty of examples (d) supported so many different ingestion and export/reporting APIs that I would be able to set up everything I wanted for my homelab without any shim scripts or one-off API converters and (e) offered a basic reporting UI with sane defaults (auto-detecting rate vs sum for a graph) even without having to set up Grafana, I was blown away and grateful that such a useful thing existed. Same for VictoriaLogs, it was just easy to set up once I put my mind to it, because the documentation for everything was very clear, and they clearly had "sane defaults + configurable options" once you needed something slightly different. Having sane support for backfills and tolerating duplicates was also nice. "Throw us your data in one of these shapes , we'll sort it out" was just nice to finally see rather than digging through pages of Prometheus documentation for what the edge cases could be if I sent duplicates or the data was from a month ago.

I just have a homelab of random docker container across a few nodes thrown together with underpowered hardware, but VictoriaMetrics met me where I was and made it trivial to experiment using the nodes I had rather than have to migrate to bigger nodes, and it was very well behaved at idle, steady-state, and "I want to trickle-feed a million data points via http calls" loads. I don't yet need OTEL, I don't have cattle, I have homelab pets and very little time to play with them. I just want to either scrape metrics or fire metrics at some sort of endpoint that can figure out what I meant if I get close enough.

But VictoriaMetrics was so easy to get working because the documentation was laid out as "here's the starter command line options, here's how you ingest data in a variety of input URLs, here's how you retrieve your data via a variety of output URLs, if you want specialty stuff that's described farther down the page..." it was about as hard as falling off a log. It just became the obvious place to base anything else around because it had so many connectors and sane defaults.

So when the Beyond All Reason infrastructure team asked "is there a infrastructure and application metrics solution for a handful of nodes that is self-hosted, easy to set up and won't break the bank or require babysitting?" I had one recommendation: VictoriaMetrics (+ Grafana)

Admittedly I do sort of wish for unified metrics and logs and traces, but that's merely a platonic ideal dream state for me. In reality I can see that both I and an organization generally sets up metrics, or logs, or traces, in a piecemeal fashion. An organization (in my limited experience) generally doesn't think about all three at once, and so the "do one thing and do it well" becomes a nice simplification of scope rather than a mark against VictoriaMetrics or VictoriaLogs not having the whole enchilada under one common roof.

I have not personally worked on scaling it horizontally yet, and I didn't set it up myself, but (a) I observe the Beyond All Reason VictoriaMetrics server has 8 GB of RAM, 3 vCPU and appears to serve 75k active time series (14.5 billion data points, ingest about 5 thousand data points per second) without complaint. The resource usage graphs are flat, humming quietly and (b) I did appreciate that the vmagent and vlagent do send to multiple targets easily (tested this with vlagent) , making "active -> standby fail-over" easy to setup -- all ingestion agents would multiplex to all sinks and you were done, any sink "should" have the same data.

[yard2010]

I have tried to self host grafana (loki prom and alloy) as o11y stack for prepbook.app. This is hard. I have a bsc in cs not that it says something. I managed to do it eventually, after some research. It was not plug and play in any way. The docs kept saying this solution is not production ready even. I couldn't find the production guide, only the "forget about self hosting and simply pay for us hosting this". After I deployed it the UX was so abrasive my partner won't even try to go into it to figure out a problem. It was a few months ago. Since then new solutions have arrived and I'm waiting to have the time to migrate. I saw PostHog have a solution but I prefer something I could self host and completely own.

I thought how come no one is trying to solve this problem. It looks like it's just a matter of time.

With that being said, my experience can be very skewed since prepbook is a passion project running on a VPS with essentially 0 scale. All I care about is the UX of the stack, not scale. Just for context.

[embedding-shape]

FWIW, I have no CS degree and barely attended school at all, and found Grafana + Prometheus + Loki fairly easy to setup, at least compared to what we used to use before those tools were available. Maybe it's because I used NixOS for the setup, but besides learning some new domain-specific things I didn't know since before, I don't recall hitting any particular bumps or roadblocks, I also went the 100% self-hosted route (spread across two hosts at home).

What exactly were you struggling with when it came to the setup? Just a ton of new concepts to learn which took time, or something specific to Grafana/Prometheus/Loki?

[dijit]

"Getting it running" is the easy part.

"Getting it ready for production" is a different game.

I've fallen on my sword many times by trying to explain that prometheus fails every metric of production ready; in fact Google themselves replaced borgmon (prometheus) for Monarch because the "tiny unreliable time series databases everywhere" was in fact, not the successful and reliable deployment strategy that they had claimed.

But, it is very easy to set up. Just don't go looking for failure modes, because they're everywhere and every single one of them is catastrophic.

[denysvitali]

There are ways to scale Prometheus (look at Thanos), but none of the solutions is really bug free.

See this PR for example (https://github.com/prometheus/prometheus/pull/18364) - this used to impact a production deployment I worked on. Prometheus, Thanos and even OpenTelemetry are full of those kind of problems - but at the same time it's the best we have and we should be grateful they're free and open source.

I'd still choose an open source stack (and contribute to it) rather than go for a proprietary solution - we've all seen what happens with DataDog & co.

Please don't take my words lightly, I worked with the rest of my team in a large scale observability platform and scalability should not be underestimated - at the same time DataDog / Splunk prices are simply unjustified. It's ironically cheaper to build a team of engineers that will maintain a sane observability stack instead of feeding the monster(s).

[otterley]

> It's ironically cheaper to build a team of engineers that will maintain a sane observability stack instead of feeding the monster(s).

Can you show the math here? This is a very bold claim, and I’m super curious. A shared Google Sheet would work well.

[embedding-shape]

Well, I am running the stack in production right now, but everyone has a different understanding of what that actually means...

Do you have concrete examples of these catastrophic failures? I've personally havent experienced any myself during these years, but I'm doing very boring and typical stuff, so wouldn't surprise me there was hard edges still.

[dijit]

There's a difficult distinction here, you're right.

Technically even a single server running LAMP as root but taking frontend traffic meets the definition of in production but I think we all recognise that it's not the right idea.

What I'm referring to is: should the disk start to have issues: what does prometheus do? If the scrapers start to stall due to connection timeouts: what does prometheus do? If you are doing linear interpolation of data and you have massive gaps because you're polling opportunisitically: what does prometheus do.

I'm all about boring technology, but prometheus assumes too much happy path. It assumes that a single node is enough for time series data that is used for alerting.

Which, it is: at very small scale and with best effort reliability.

It's not acceptable as soon as lost data could be critically important in diagnosing major issues in billing systems, or actually billing users, or to infer issues that need to be correlated across multiple systems.

[jmalicki]

What is the disk? You've already failed by not running distributed. The problem isn't Prometheus, it's "the cloud is too expensive I'll just run on a single VPS"

[embedding-shape]

> should the disk start to have issues

If that happens, is prometheus really the biggest of your worries here? Software breaks left and right when disks disappear from under them, I'm not sure this is neither unexpected or unique to prometheus.

> If the scrapers start to stall due to connection timeouts: what does prometheus do?

I'm having this "issue" all the time, as some of my WiFi connected (less important) cameras are just within the WiFi range, and I'm using prometheus to scrape metrics from them. It seems like the requests times out, then the next time it doesn't, and everything just works? What's the issue you're experiencing with this exactly?

> It's not acceptable as soon as lost data could be critically important in diagnosing major issues in billing systems, or actually billing users, or

Wait what? Billing systems? That stuff would go into your proper database, wouldn't it? Sure, if prometheus/node_exporter fails or whatever, you won't get metrics out of the host, but again, if those things start failing on that host, the host is having bigger issues than "prometheus suck at scale".

I was eagerly awaiting to be educated about potential gaps in my understanding of prometheus, instead it seems like you simply don't happen to like they way they do things? I was under the impression they did something wrong or something was broken, but these things just seems like the typical stuff you have to think about for any service you deploy.

[dusanstanojevic]

Hi, I'm the creator of Traceway.

I have created Traceway because I looked at that stack and decided that I'm not going to add 7 more services to my stack that could all fail that I now have to maintain as well. Here is the list: Grafana, Otel Collector (to forward metrics), Prometheus, Loki, Tempo, Mimir, K8s.

This is not maintainable in production, unless you have a person to manage it. My app had about 500-1000 req/sec, this sounds like a lot but it's extremely light from the observability perspective. Why would I add 7 more points of failure and services to monitor for proper resource allocation for something like this? To add insult to injury I would have to keep building my SLOs, they wouldn't be tracked automatically by default, I would have to keep paying for Sentry because the issue tracking is quite lacking on Grafana. Oh almost forgot, I would also have to get an alerting provider or pay for that (maybe I'm wrong, it was 6 mo ago).

Anyhow, Traceway is a 60mb binary in Go, it works with Clickhouse or Sqlite and the data is stored on S3 when not used. That means you can host it with sqlite on the 2$ server or even free tier and have it working for your side projects, you can host it with managed clickhouse and get auto scalability on the db level.

The goal is to provide full observability and tools to fix issues directly for developers. What we have so far: alerts, notifications, SSO (google & github), integrations, metrics, preconfigured SLOs, distributed tracing, RUM/session recordings (js & flutter).

Almost forgot, you'd need a symbolicator as well, or your fe/mobile exception stack traces will be messed up in Grafana, I don't even know which tool they have for that, but it's always a new service to host and maintain...

[SOLAR_FIELDS]

FWIW, if you come flying in saying you used NixOS to set something up you’re not what we would call a “casual user”

[embedding-shape]

Why not? Hardly unheard of for managing infrastructure. If we were talking about desktop environments, then maybe, and to be fair, I never said I was a casual user, just that I didn't find prometheus particularly difficult to manage in a production environment.

[SOLAR_FIELDS]

The implication is that by virtue of using NixOS, you're already a self selected power user. The people that would find setting this thing up in production difficult and the people who would use NixOS are a very small overlap, if any, on that venn diagram.

[embedding-shape]

NixOS is an additional thing on top of prometheus, not a replacement. Not sure why it'd dictate how easy/hard it is to run prometheus & co, you still have to know the same stuff as without it.

[parliament32]

FWIW we've also tried all sorts of different things, and honestly the very vanilla (prometheus -> central thanos, fluentbit -> central loki, grafana) ends up on top. The resource consumption is surprisingly minimal (for a sense of scale, we run about 200k eps for metrics and 1k eps for logs). For all these solutions, I find myself asking the same question as you.. what problem are you trying to solve? Is there anything actually different about your product other than less stability than the battle-tested stack?

[dusanstanojevic]

Hi, sorry for not responding sooner this one slipped through the cracks. I've tried to explain my reasons for starting traceway and what I've been building with it. It's not aimed to be the fastest ingestion tool out there, but it's backed by clickhouse and does minimal processing of the data, you can expect the perf to be as close to clickhouse as possible.

I'm working on a comprehensive benchmark of Traceway performance on different hardware configurations. The most I've tested with was the smallest managed ch instance with 250k traces per sec, handled it without a hiccup (but that's empirical). You can checkout the traceway git, there is an issue I've opened for benchmarking and you can subscribe/comment on it if you're interested. I'm benchmarking across sqlite, self hosted clickhouse and managed clickhouse. I am a huge fan of systematic, realistic and most of all reproducible benchmarks, so I am really excited about the progress on that.

Anyhow, you can checkout traceway and see what it offers, it's aimed at providing SLOs out of the box, session replays, alerting, configurable dashboards and great exception tracking (automatic symbolication) etc...

[ting0]

Do you think Prometheus + Grafana is the way to go?

[denysvitali]

Really depends on the use case. Home lab? Probably.

Production? As soon as you scale you need a proper solution. Prometheus (by itself) doesn't scale - you need Mimir or Thanos (or similar).

Clickhouse (the "clickstack") seems to be the new kid on the block. Looks very promising.

[parliament32]

Note Clickhouse is quite old (2010ish?) but they've always been a "web server access log analytics" solution. The pivot to "we do observability too" is new, we'll see how that plays out. Not terribly optimistic given how badly a similar pivot went for Elastic, but who knows.

[dusanstanojevic]

Clickhouse is just a database, it has a really neat feature that infrequently accessed data is pushed back to S3 minimizing the costs. It also heavily compresses the data when storing it.

I am the creator of Traceway and it's my all time fav database. Having said that the repositories in Traceway are completely modular, I've implemented the sqlite version so that I can skip docker containers locally and to simplify self hosting for side projects (it runs on like 2$ servers without issues). This is why it's uniquely suitable for telemetry data and why I've used it as the base of Traceway.

They've acquired HyperDX because it was a major Clickhouse user because their whole platform was telemetry on top of Clickhouse. I hope they don't fully pivot into the space as it would be quite awkward, but there are alternatives and I can always redo repositories with a diff storage engine/db.

[NortySpock]

I thought observability was shoved on Clickhouse by other stacks deciding to use Clickhouse as their recommended database for observability (SigNoz springs to mind but they were not the only one)

[valyala]

VictoriaMetrics CTO here.

We at VictoriaMetrics took another approach. We tried using ClickHouse as a database for metrics in 2017, but then decided implementing a specialized database for metrics. This database uses ClickHouse architecture ideas for achieving the best performance and the lowest resource usage. The main difference between ClickHouse and VictoriaMetrics is that VictoriaMetrics is optimized solely for typical observability tasks. It supports all the popular data ingestion protocols, it provides promql-compatible querying API, it provides Graphite-compatible querying API, it provides Prometheus-compatible service discovery and relabeling, it provides Prometheus-compatible alerting and recording rules. It provides built-in web UI for quick exploration and analysis of the ingested metrics, with the ability to investigate the source of high cardinality. It consists of a single small executable (~20MB) without external dependencies with minimum configs and minimum maintenance. See https://altinity.com/wp-content/uploads/2021/11/How-ClickHou... for more details.

We used the same approach for building VictoriaLogs - a specialized database for logs. It uses the most appropriate architecture ideas from ClickHouse for achieving high performance and low resource usage. It is schemaless and zero-config. It contains of a single small executable without external dependencies. It accepts logs via popular data ingestion protocols. It provides a specialized query language for typical queries over production logs - LogsQL. This language is much simpler to use than SQL for querying typical logs. It provides a built-in web UI for quick exploration of the ingested logs. It provides a Grafana plugin for building arbitrary complex dashboards from the stored logs. It provides the ability to build alerts and metrics from the stored logs. See https://docs.victoriametrics.com/victorialogs/faq/#what-is-t...

[dusanstanojevic]

This is refreshing to see, an actual comment about VictoriaMetrics that explains why and when it's good. Thank you!

I am the creator of Traceway, after reading that I think we're in adjacent vertices.

Traceway is aimed at providing teams with no dedicated SREs with preconfigured SLOs, a preconfigured dashboard and a really powerful exception tracking. It integrates with git to automatically open issues as well as slack and others. The idea is to get the experience of Datadog/Sentry in the open source space.

I've focused a lot on session replays, RUM and now symbolication for native, flutter and frontend applications - this might be a potential place where VictoriaMetrics could benefit from integrating with Traceway.

If any of that sounds potentially interesting let me know. Again, thank you for your comment.

[denysvitali]

I mean, the idea of using OTEL with ClickHouse is rather new, and solves the most painful part of metrics: high cardinality. Has its use-cases, but for sure comes with its own problems

[drzaiusx11]

We're on AWS Managed Prometheus + Grafana in production and it certainly scales just fine, although I'm sure under the hood it's an entirely different beast than FOSS Prometheus, likely only AWS engineers truly know..

[CyberDildonics]

Is "observability stack" the new term for logs and stats?

[denysvitali]

You have more than that nowadays. Tracing and profiling are part of O11y too