chDB Durable Layer: flush, checkpoint, and object-storage agent memory

ClickHouse published the chDB Durable Layer on 2026-09-28 (Changshuo Chen): an addressable, single-writer, recoverable embedded analytical object—local MergeTree as the working copy, authoritative state in your object storage—via app-controlled flush() and checkpoint() (blog).

What it solves

Embedded chDB agent memory is fast on one disk but tied to a local MergeTree directory; a full ClickHouse server restores portability at the cost of remote round trips on every recall. Durable Layer sits between those shapes: no DB server, PVC, or Litestream-style sidecar on the hot path—local working copy plus authoritative bucket copy (s3://, gcs://, azure://, or local: for dev) (blog).

flush / checkpoint / lease

• flush() — Writes stay local until flush; when it returns, the writes it covers have reached object storage (the app chooses the loss window).
• checkpoint() — Between checkpoints, state is base snapshot + WAL; checkpoint writes a new base so the next open need not replay a long log.
• head.json + conditional/CAS writes — Ownership fencing: one writer per object name by design. Wrong tool for a shared multi-writer team database (blog).

Versions

Surface	As stated
Python	chDB 4.4+ — pip install "chdb[durable]" (S3/boto3); GCS/Azure via chdb[durable-gcs] / chdb[durable-azure]
Core	Durable V1 in chdb-core 26.7.3
Node	npm [email protected] (same layout/lifecycle—not the Python version number)
Go / Rust	chdb-go/[email protected] · [email protected]

Objects written in one binding are recoverable in another under the same Durable V1 contract (blog).

Vendor figures and examples

• 58× local-vs-remote queries — ClickHouse’s own local-vs-remote experiment, not an aitamer measurement.
• Compression sample: 1.45 GB / 213,721 rows Claude Code transcript → 521 MiB ZSTD(3) or 991 MiB LZ4 in MergeTree — “small local experiment” on the blog.
• ClickMem, Maple Local, ReplayHouse, vcfclick — illustrative shapes of the problem; not separate product launches. ClickMem is the clearest Durable use-case example (third shape between ~/.clickmem/data and a ClickHouse server).
• V1 WAL is statement-replay: prefer deterministic statements (avoid now() in INSERT); not OLTP / not a Postgres replacement (blog).

Who should care

Teams embedding analytical agent memory who need recoverability without a full server on the hot path should start at the ClickHouse Durable Layer post—lead with flush/checkpoint, keep the single-writer framing, and treat pip 4.4+ as separate from npm [email protected].