// the find
apache/fluss
Apache Fluss is a streaming storage built for real-time analytics.
Apache Fluss is a streaming storage layer that holds Arrow-based columnar tables, serves them with low latency, and tiers the same tables into lake formats such as Paimon and Iceberg. It is aimed at teams that currently run a streaming log plus a separate lake pipeline and want one table abstraction over fresh and historical data.
- Primary-key tables push deduplication, partial updates, and aggregation merge engines into the storage layer, so the Flink job doesn't have to hold that state itself. This is the most useful design choice here, and the one to stress-test under load.
- Reads and writes go through Arrow (ArrowLogFetchCollector and ArrowLogWriteBatch in fluss-client), which lets engines prune columns and push predicates down instead of deserializing whole rows.
- CI is broad for a single repo: nightly runs, client integration, a Helm chart workflow, Docker gateway builds, Rust and Python release pipelines, and a dedicated unstable-test reporter that tracks flaky tests instead of hiding them.
- Apache 2.0 under ASF governance, which matters if you need to get it through a legal review.
- 'Sub-second data freshness' is the headline claim, and the README gives no number, benchmark, or link to one. Measure end-to-end latency on your own workload before planning around it.
- The engine list is narrow: Flink and Spark, with StarRocks marked 'coming soon'. The vector, multi-modal, and feature-store claims appear with no example or doc link in the README.
- The README is a pointer page. The build is one Maven command on Java 11, and nothing here describes what a deployment consists of or what you have to operate, so judging ops cost means reading the docs site first.
- Rust and Python clients and the gateway are in the tree, but the README doesn't say which features each client supports. Check parity for your language before assuming it.