feat: wire the durable (L2) store seam — chdb-serverless[durable]

[ShawnChen-Sirius]

What

Adds chdb_serverless/durable.py — the DurableStore that the store seam already dispatches to for CHDB_STORE=durable:<url>?id= (see store.py) — and points the durable extra at chdb-durable[s3]. This lands the L2 (durable object) tier of the store seam.

How it works

Same open_store() API as L1. CHDB_STORE=durable:s3://bucket/prefix?id=user-123 runs the analyst against a chdb.durable object on object storage you own (S3 / GCS / Azure Blob / any S3-compatible / local). Reads go to the object; writes are auto-detected, logged to its WAL, and flushed immediately (a committed write survives the instance dying). chDB is single-writer per process, so this store is the process's one engine.

⚠️ Dependency status — read before releasing

chdb-durable (the DAO primitive) is not yet on PyPI — it is slated to land in the main chdb package as chdb.durable, at which point the extra becomes chdb[durable]. Until then, pip install chdb-serverless[durable] will not resolve. The base package and the L1 (local:) / L3 (memory:) paths are unaffected. Recommend not cutting a chdb-serverless release advertising [durable] until chdb-durable ships. Landing this now keeps the adapter next to the seam that already references it.

Testing

Verified locally and against MinIO: durable state persists across restart (base + WAL replay). In-region cold-open measured on AWS/Azure/GCP (~0.5–0.85s for 1M rows; control-plane-round-trip-bound, not restore-bound).

🤖 Generated with Claude Code

Note

Wire the L2 durable store seam backed by chdb.durable objects

• Adds durable.py, a new module implementing a thread-safe DurableStore that routes SQL to a chdb.durable object.
• DurableStore.__init__ parses a durable target URL, reads config overrides from environment variables (CHDB_DURABLE_ID, CHDB_DURABLE_OWNER, CHDB_DURABLE_DB), and opens the named object.
• Writes (classified via _is_write, which strips leading SQL comments before checking the first keyword) call execute then flush and return an empty string; reads call query with a configurable format (default JSONCompact).
• Exposes checkpoint and close methods, both guarded by the same mutex.
• Updates pyproject.toml with clarifying comments on the durable optional dependency group.

^{Macroscope summarized ddd9d53.}