3.0 evidence status and known gaps

Relata DB 3.0
This page records the evidence behind the claims made for the 3.0 open edition. Gaps listed here apply to 3.0 until the table says otherwise.

This page lists what the recorded evidence shows does not work, or was not tested, for the single-node open edition, so that no go-live, migration or readiness statement says more than the evidence behind it. It is a record of findings, not a plan or a commitment.

  • "Evidence" means a test, a commit or a recorded run in the source repository. A statement with none is marked not yet evidenced.
  • Linux runs were made on one shared Apple-silicon host in an aarch64 guest (4 vCPU, 6 GiB, no swap). Timings are indicative only and are not published numbers.
  • A fix verified by a macOS unit test is not a fix verified on Linux.

Gaps

#GapStatus
G1Ingest slows past about 115,000 rows at default index caps. About 1,500 rows/s, then the index RAM cap is reached and throughput falls to about 34 rows/s. The 400,000-row target was never reached at default caps. Reproduced twice.Open
G2Raising the index cap moves the limit to memory. 400,000 rows ingested with RSS near 5.1 GB of 6 GB; a larger pending-index limit returned 503 memory-pressure near 352,500 rows. One run returned an empty point lookup of the last id, not investigated.Open. No sizing figure is a claim.
G3Restart after a large ingest failed on Linux in three runs: WAL replay failed ... deferred index-work queue full.Open on Linux
G4The replay fix was verified on macOS only. It was checked by four targeted storage tests. No Linux restart was re-run after it.Fix present, not verified on Linux
G5Restart of a 200,000-vector store did not become ready within 36 minutes. A sampled stack points at the WAL sequence-spool merge. 10 million vectors not run.Open
G6Writes pause with 503 during generation rollover. Clients must retry with backoff. A later soak with repeating keys does not prove reclamation.Behaviour documented; pause not eliminated
G72.7.1 to 3.0 stored-data upgrade is not tested. No test upgrades a real 2.7.1 data directory.Not yet evidenced. Rehearse on a copy with a verified backup.
G8SIEM/SOC migration is rehearsed, not qualified. Only a synthetic fixture was migrated.Open
G9No workload is Qualified. All recorded workload records are "evidence recorded" or "unqualified"; independent review is pending and no non-author reproduction exists.Open
G10Standards: no external validator was run (PROV-O round trip, STIX validator, TAXII fixture, OpenAPI validator, Cedar reference evaluator, OTel conformance). Boundaries are documented only.Documented boundaries only
G11No release evidence bundle. No evidence bundle or named documentation reviewer is recorded for the release.Open
G12Not run in the Linux rounds: the kill -9 phase after round one, default-cap runs beyond about 120,000 rows, backup and restore qualification, and the watch phase.Not run

Later qualification work

After the table above was written, one further Linux round (boot memory work, ticket #6026) recorded: a 100,000-row memory-server run in a 4 GiB container passing, with restart ready in 125 s and a cgroup peak of 3.81 GiB, where the same restart had been killed for lack of memory before the change. In the same round a sustained run (900 s, 6 GiB) completed ingest, rollover and the soak, but the post-sustain restart was still killed for lack of memory. The table above has not been updated for this round; treat G1 to G4 as open until it is.

What this means

  • Go-live: do not plan on more than roughly 115,000 rows of sustained queued ingest at default caps without measuring your own shape, and rehearse restart after a large load. See Capacity planning and WAL retention.
  • Backup and restore: not qualified. See Backup and restore.
  • Migration: G7 and G8. See SIEM/SOC migration.
  • Readiness records support only their own revision, environment and workload.

SDK note: the three supported SDKs (Python, TypeScript, Go) are held in module parity by a repository check. One shared partial gap is known: nearest-neighbour search by a caller-supplied embedding emits an ORDER BY form the server parser rejects. That is a method-shape observation, not a statement about overall readiness.