chDB

chgraph, a local MCP code-graph server built on chDB, can now outlive its disk: point CHGRAPH_DURABLE_URI at s3:// and the symbol graph plus git history restore from chDB checkpoints and WAL after the local data is gone, behind a single-writer lease.

github.com/jordan-simonov…

chDB chdb-rust v2.1.0 is out.

QueryResult::data_str() borrows a result as &str with no copy, and InsertStream now buffers writes, so a writeln! per row no longer costs an FFI call per row. Built on chDB core 26.11.0-rc.2.

github.com/chdb-io/chdb-r…

torch-spyre (PyTorch for IBM's Spyre AIU) tests its ClickHouse CI warehouse in two stages: pytest runs the views on in-process chDB, then CI applies the schema to a real ClickHouse 26.3 twice and the second pass must change nothing.

github.com/torch-spyre/to…

@ashokponkumar

Agent memory tip: keep episodes and lessons in separate chDB tables.

A materialized view folds each tool call into an AggregatingMergeTree on insert:

countState() n, sumState(toUInt8(1-ok)) fails
GROUP BY day, tool, err

10M calls -> 3,639 rows. Delete old calls; rates stay.

Kluard, a Go OpenTelemetry backend, checks its SQL on a real ClickHouse engine with no server: Go tests dump every DDL and query, then a Python script runs them in chDB and asserts exact answers, from tenant isolation and counter resets to TTL changes.

github.com/getkluard/klua…