Looking at their diagrams it seems that the k8s cluster exists solely to handle their monitoring and logging needs which would be extreme overkill, especially since 18k metrics/samples and 7k logs per second are nothing. Plus you now suddenly need a S3-compatible storage backend for all your logs and metrics. Good thing Ceph comes 'free' with Proxmox, I guess.

Deploying an instance of Prometheus with *every host is also unusual, to say the least and I don't quite understand their comment to that. If you don't like a pull-based architecture (which is a valid point) why use one at all!? There are many more push-based setups out there that are simpler to set up and less complex.

> k8s cluster exists solely to handle their monitoring and logging

Does image processing, runs our analytics, runs our Sentry, runs our gitlab-ci runners, and quite a few other things not mentioned expressly

> which would be extreme overkill

That's an interesting argument against k8s ; if anything I find it much easier to work with -- once accustomed to its idioms, ofc -- than alternatives like dedicated VMs, Docker Swarm etc

Getting HA and auto-healing for free is possible without it, of course, but does require much more work, especially if you aim for a somewhat minimised amount of statefulness (as in deviation from the template of your system)

Also S3-compatible storage backends are really aplenty, from commercial offering to simpler ones like MinIO. Ceph just happens to be a bit higher of a deployment investment with the benefit of fantastic performance, flexibility and resiliency. Somewhat like k8s itself, it's a bit daunting at first but does actually make things simpler in the long run (imo)

> 18k metrics/samples [...] are nothing

Well yes and no, the number of metrics isn't relevant per se, but its cardinality is very relevant, and managing that in a single prometheus instance will quickly require some serious vertical scaling, especially if you want to look at data on longer ranges (which, in contrary to logs, we are interested in)

> 7k logs per second [...] are nothing

That's an interesting take . Surely this isn't a world-record-shattering amount indeed, but no one seems to have such a great non-SaaS-or-cheap solution to storing, sorting and querying this amount of logs either (at the resource efficiency of Loki anyway), so maybe we just have a different set of expectations for log management

> If you don't like a pull-based architecture [...] why use one at all!? There are many more push-based setups out there that are simpler to set up and less complex.

Are there really? That is non-SaaS and with as widespread 3rd-party software support as Prometheus does? ie great integration with essentially any database, webserver, runtime, OS, etc?

Because if we talk only about node metrics like CPU etc then yeah, sure there are plenty of options. But (maybe not so) obviously the diagram showing only node exporter doesn't mean that this is the only integration we use -- we collect prometheus metrics for MySQL, PHP-FPM, Varnish, Nginx, HAProxy, Elasticsearch, Redis, RabbitMQ etc (essentially every single piece of software we use).

Fwiw I found very little in the way of open-source solutions to that problem that ticked as many boxes as Prometheus.

As for "simpler to set up and less complex", both Cortex and Loki would be really annoying to manage outside of Kubernetes, I'll happily give you that. But... being able to easily deploy and manage such systems once you have Kubernetes is precisely one of the reasons to use it. You can't say it's complex to deploy itself but then ignore the fact that it largely outweighs this by making reliable operation of complex-but-powerful software on top it, that is precisely one of the upsides of using it in the first place :)

Thank you for your reply and clarification. This is quite an interesting topic for me as I've tested and implemented similar setups.

> Does image processing, runs our analytics, [...]

Fair enough, I was strictly going by the diagrams. From my experience with a somewhat similar setup (HA Loki, HA Prom + Thanos with a MinIO storage backend using Terraform + Ansible and docker) I have to say that the most complex and frustrating part was configuring Loki (this was way before they expanded their documentation, which still isn't great). I'd imagine this would be even more challenging under k8s at least if you stray from the vanilla deployment and/or charts. I agree with your statement regarding Ceph, we use it extensively in production (probably on a much bigger scale). However, I think Ceph, unlike MinIO, just adds unnecessary complexity to your setup.

> Well yes and no, the number of metrics isn't relevant per se, but its cardinality is very relevant [...]

Cardinality is something you should avoid when using Prometheus - for exactly that reason. There are, in my opinion, very few good reasons for dynamic labels (ignoring the baked-in cardinality from a setup like k8s). On first impulse I'd say you're doing metrics wrong but then again, I do not know enough about your use case. Maxing out a single Instance of Prometheus is no easy feat however, especially if your infra isn't that complex and/or big. I've used Thanos for so long now, how does the Cortex compactor handle range queries? Does it also compact and create additional 5m & 1h resolution metrics? These might help with your larger range queries.

Just out of curiosity, have you had any look at alternatives like Victoriametrics?

> 7k logs per second [...] are nothing

My remark was just regarding the added complexity as this depends solely on the size of your log messages. If you don't need or use the (awesome!) capabilities of Loki + Grafana and just need a place for long-term storage of your logs, a 'simple' rsyslog server will do just fine.

> we collect prometheus metrics for MySQL, PHP-FPM, Varnish [...]

Many (if not all) of these can be handled by Telegraf or Fluentd plus InfluxDB (not that I'd used that myself, I absolutely love Prometheus and its Eco-system). My tongue-in-cheek comment was mostly about the Prometheus instance you deploy on every server just to scrape metrics locally and remote-write them into Cortex. Why not the more usual setup of (one or more) Prometheus instances scraping their targets and writing to Cortex?