[1] https://grafana.com/docs/mimir/latest/references/architectur...
[2] https://github.com/grafana/mimir/blob/9a52522a05f25d6d6bc2e8...
[1] https://grafana.com/docs/mimir/latest/references/architectur...
[2] https://github.com/grafana/mimir/blob/9a52522a05f25d6d6bc2e8...
Mimir will consume multiple Prometheus with remote_write's and de-dup those series. Because it puts older (default 2h) data out to block storage, costs plummet without impacting performance too savagely - the vast majority of queries are going to be hitting fresh data.
We've been running monolithic (Nomad, docker) but are experimenting with their 'in-between model' which has 3 components - read, write, backend - which should help us scale more easily than juggling the full microservices deployment method you mentioned above.
> If Prometheus supported horizontal scaling ...
Well, yes, precisely so.
Would appreciate if you refer to link where this explained. But from what I was able to find, both Mimir and Prometheus store indexes within data blocks. It doesn't matter where data blocks are: locally or on S3. Because you still need to read the index and fit it in the memory in order to locate the needed data. So I don't understand how S3 helps here, unless you run an additional store-gateway which does this job instead.
> but Mimir was designed to handle much more data than Prometheus was
Do you have any numbers to refer? Isn't Mimir based on Prometheus source code? The storage, indexes, query engine - that should be the same or mostly the same. I'd be really surprised if Mimir could outperform significantly Prometheus in monolithic mode.
If I understood correctly, the whole "Mimir was designed to handle" is related to microservice architecture, to its horizontal scaling capabilities and ability to shard read queries across multiple components.
https://grafana.com/docs/mimir/latest/get-started/about-graf...
Perhaps indexes wasn't the optimum word. There's some references to Prometheus' memory usage if you go looking, typically if you're searching around 'why does prometheus use so much memory'.
I'm struggling to find the comparative metrics on capacity that you're seeking, though I do recall seeing them previously, and having run up multiple prometheus instances (in vanilla and agent-mode) and fed those into mimir, I'm aware that prom has some non-negotiable limits on time series.