Hey everyone! We're thinking about transitioning from our current setup with two independent Prometheus servers and using Promxy for a unified query layer to Grafana Mimir. Our main focus is on whether this switch could provide performance and efficiency gains.
We would love any insights or links to benchmarks that compare these two options, especially regarding:
- **Query performance:** How does Mimir (with a High Availability setup and MinIO backend) handle long-range queries (over 6 months) compared to local Prometheus TSDB?
- **Storage efficiency:** How does the storage usage of Mimir stack up against that of local Prometheus for similar retention periods?
- **Quorum and minimum footprint:** Does Mimir need at least three hosts for quorum/high availability, and what's the minimum deployment size that still supports HA?
Thanks for your help!
4 Answers
Mimir feels like Prometheus trying to morph into a Kubernetes-style startup. You'll gain better handling of long-range queries and potentially improved compression, but you could end up complicating your operations without significant benefit.
I suggest checking out VictoriaMetrics as an alternative. I've been using it and found it performs really well and is easy to set up. Plus, it allows for pushing metrics instead of traditional pulling, which is great for on-prem systems with complex firewall setups.
How does long-term storage work in VictoriaMetrics?
Choosing between Mimir and Prometheus isn’t so much about raw performance as it is about architecture. Prometheus is a monolithic application, while Mimir is built as a microservices solution, which can scale and tolerate faults. Performance can favor Mimir in some areas eventually, but generally, it’s the scalable architecture that drives the decision.
We actually use both—Prometheus for short-term metrics and Mimir for the longer term. We only send essential metrics to Mimir and ignore the rest.
Switching to Mimir generally means a drop in efficiency. While Prometheus relies heavily on in-memory caching to minimize overhead, Mimir introduces networking and object storage into the mix, which can slow things down. Essentially, with Mimir, data is scraped, sent to a Mimir receiver, which then mimics Prometheus by creating TSDB blocks and saving them in object storage before pulling them back for queries. This is mainly the trade-off for being able to distribute queries across multiple servers.
However, if you're considering long-term storage, cloud providers often make object storage cheaper than local disk storage per byte, which might make Mimir or Thanos a more economical option in the long run, despite the per-request costs.
I totally agree! Pushing metrics with vmagent is a game changer for managing multiple networks.