Skip to content

Brainwire v0.20.0-beta.2

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 14 Nov 16:56
· 36 commits to main since this release

Experimental beta: evaluate with rebuildable data and a tested backup.

What's new

  • Accounted query growth now uses a sampled cgroup margin alongside RSS. Under pressure the server can refuse new query reservations; this is not a hard RSS ceiling. A local Linux A/B test verified rollback and retry in one short synthetic episode.

  • Label-qualified primary-key lookups avoid intermediate allocations. Timeline examples reject timestamp precision or timezone offsets that could be silently truncated.

  • DuckLake checkpoints now bind the source generation, mapping, graph, writer identity and endpoint policy. Existing checkpoints without the new contract are refused rather than silently reused.

  • Decimal values that cannot be represented exactly as floats are rejected; map text for exact values or explicitly choose DOUBLE for approximation.

  • A reproducible synthetic benchmark covers 1 million nodes and 2 million relationships on three engines, with raw receipts and restart checks. Brainwire results are medians of three runs of a locally built candidate; the published image is a separate build. Durability and client settings differ.

  • The README, agent quickstart, Colima guidance and documentation site are simpler. CI now exercises real DuckLake checkpoint races and incremental synchronization.

  • An authenticated Bolt regression checks that a stale ACL stamp in the same transaction conflicts with a concurrent revocation and that the revocation survives restart. Carrying a stale ACL to a new transaction can still restore access; reread it and handle conflicts.

Upgrade DuckLake synchronization

--source-identity (or source_identity=) and the ImportCheckpoint.contract_fingerprint ontology property are required. Stop and reconcile previous workers and uncertain commits before migrating. A new graph with a verified full reload is the safest starting point. Failed synchronization can leave earlier batches committed and its lease held. See parallel ingestion.

Evidence and limits

Local tests covered DuckLake concurrency, ACLs, process and quorum recovery, and short 512 MiB/2 GiB/3 GiB cgroup pressure scenarios. Separate 30-minute synthetic read and mixed read/write runs at 2 GiB stayed well below the limit and are not saturation tests. These do not establish a process-wide RSS ceiling or qualify distributed failover across machines. Cypher and search remain bounded, with no native ANN/BM25; native Bolt TLS is not implemented. See the benchmark report and current limits.