// the find
thanos-io/thanos
Highly available Prometheus setup with long term storage capabilities. A CNCF Incubating project.
Thanos extends Prometheus with global query views, HA deduplication, and unlimited retention by shipping metrics to object storage (S3, GCS, etc.). It's for teams running Prometheus at scale who've hit the single-instance ceiling — multiple clusters, multi-year retention, or cross-datacenter federation. CNCF Incubating with serious production adoption.
The sidecar architecture is genuinely clever: bolt it onto existing Prometheus instances without rewriting your setup, then get unlimited retention by offloading blocks to cheap object storage. The Store API is a clean abstraction — everything (local TSDB, object store, remote endpoints) speaks the same gRPC interface, which makes the query layer composable. Downsampling for long time ranges is handled automatically by the Compactor, so 5-year range queries don't kill you. The tooling around the object store (thanos tools bucket) is genuinely useful for operators who need to inspect, verify, or repair blocks.
The operational complexity is real and non-trivial — you're now running Sidecar, Store, Query, Compactor, and optionally Receive and Query Frontend as separate components, each with their own failure modes. Receive (the remote-write path) is the rough edge: hashring configuration for multi-tenant setups is fiddly, and rebalancing when nodes join or leave is still a pain point that can cause write gaps. The embedded Cortex code under internal/cortex is a historical wart — it's vendored in rather than properly depended on, which means divergence from upstream and confusion about what's actually maintained. Query latency against cold object store blocks is noticeable; the index cache helps but adds yet another dependency (Memcached or Redis) to your stack.