// the find
GreptimeTeam/greptimedb
The open-source Observability 2.0 database. One engine for metrics, logs, and traces — replacing Prometheus, Loki & ES.
GreptimeDB is a Rust-built OLAP/observability database that stores metrics, logs, and traces in a single columnar engine on top of object storage (S3/GCS/Azure Blob), queryable via SQL and PromQL. It targets teams running the Prometheus + Loki + Elasticsearch stack who are drowning in ops overhead and storage costs. At v1.0 GA with real production deployments (OceanBase Cloud runs 80+ clusters on it), this is past the 'interesting experiment' stage.
1. The architecture is genuinely well-thought-out: compute-storage separation with stateless frontends means you can scale reads horizontally without resharding, which is the exact failure mode that kills Prometheus at high cardinality. 2. Built on Arrow/Parquet/DataFusion — not reinventing the columnar storage wheel, just wiring proven primitives together with a good WAL and compaction layer on top. 3. Wire-compatible with Prometheus remote write, OpenTelemetry, Jaeger, and even Elasticsearch/Loki ingest APIs — you can migrate one signal at a time without a flag day. 4. Standalone mode is a single binary that starts in seconds via Docker, which makes local dev and small deployments actually viable rather than requiring a Kubernetes cluster just to try it.
1. The distributed mode dependency on etcd (or an RDS-backed KV) for Metasrv means you've traded one ops headache for another — if you're running this self-hosted, etcd HA is non-trivial. 2. Flownode (the streaming/materialized views component) is listed as 'optional' and the RFC history suggests it's still catching up to the core engine in maturity — don't lean on it for critical alerting pipelines yet. 3. Requires nightly Rust to build from source (pinned via rust-toolchain.toml), which is a contributor friction point and signals some unstable feature dependencies that could bite you if you need to patch something urgently. 4. The 50× cost savings claim is compared against Elasticsearch at full replication, which is a generous benchmark — real-world savings against a tuned Thanos or Loki setup will be much more modest.