penca
Branchable, versioned OLTP and OLAP on one open columnar copy of your data. An experimental proof of concept.
Penca explores replacing the split production and analytics stack with one version-controlled transactional lake, with no CDC, no ETL, and no second copy of the data. It is an experimental proof of concept, and its core ideas carried forward into litelink and streamcast.
What it does
Live data forks into isolated zero-copy branches, so you can test migrations, tune queries, or train models without standing up a sandbox copy. Row-level versioning records who changed what and when, which makes debugging and rollback much easier.
How it is built
The core is Rust. Query, write, and lifecycle gRPC services sit behind an Arrow Flight SQL gateway, with a DataFusion-based planner that supports primary-key pushdown and secondary-index seeks. A Postgres hot tier sits over a columnar cold tier in object storage, drained by a persist, snapshot, and purge scheduler. Copy-on-write branching, time travel, and cross-catalog ACID transactions run over immutable logs.
Read more
The design is written up in Designing Penca: the version-controlled lakebase.