Repository navigation
v0.25.0 — Vector evidence chain, versioned migrations, small-team deployment
v0.25.0 — Vector evidence chain, versioned migrations, small-team deployment
First GitHub release for Datakeel. Highlights since v0.24.1:
🔍 Vector evidence (fake kernel, end to end)
- Signed vector pipelines with semantic duplicate detection, DBSCAN clusters/noise and relative coverage-gap candidates — frozen snapshots, atomic publication, replayable evidence.
- Read-only vector evidence queries with signed cursors, typed failure reports and a bilingual Evidence view; reviewed gaps can be added manually to an ObjectSet selection without executing actions.
- Development-only offline FAISS exact-comparison campaign with reproducible cold/hot/incremental measurements.
- Honest boundary: the chain runs on a controlled fake encoder. Real encoder quality and cloud latency are untested; real egress and budget remain disabled.
🗄️ SQLite versioned migrations & recovery
- Versioned migration registry with frozen baselines, transactional forward steps and resumable committed prefixes.
datakeel db status/upgradewith stable JSON/exit classifications.- Hash-verified local recovery packages (database/WAL facts, referenced objects, non-secret configuration); restores stay isolated from running services.
- PostgreSQL/S3 versioned maintenance is not yet supported.
🧱 Storage architecture
- Task and dataset persistence split into
worker_store/dataset_storeadapters with explicit typed schema composition (102 tables). - API, worker, CLI and MCP startup refuse non-current SQLite stores.
🚀 Small-team deployment guide
- New bilingual guide for the supported authenticating-proxy boundary: loopback defaults, shared-operator and registered per-user modes, Caddy/nginx + oauth2-proxy and basic-auth recipes that strip client-supplied actor headers, plus a forged-header/revocation/audit verification checklist.
- No application authentication change — deployment-level auth remains a separate layer by design.
📊 Capacity evidence refreshed
- 18 sequential runs on the current tree (B1–B8 M on both stacks plus B1 L), one identical source fingerprint.
- B5 completed with zero errors on both stacks; B4/B7 re-measured after the storage split.
- B8 team tier: PostgreSQL 50 annotators meets every target (save p95 113 ms). SQLite 50 did not meet the latency targets in this campaign — flagged for investigation; use PostgreSQL for the team tier meanwhile.
- The 0.13.0 matrix is retained as a labelled historical appendix.
Validation
Full make check green on the release commit: 16,829 Python tests, 97% coverage, 1,055 frontend tests, license audit 0 findings. CI runs both database stacks on push.
Full changelog: CHANGELOG.md · 中文说明 · Ten-minute tour · Deployment guide