Physical security: video evidence

Relata DB 3.0
Relata DB stores, indexes and retrieves. It does not decode video, run detectors or extract captions: that is OpenETL's job (or yours). POST /ingest/media was removed; the media SQL operators (FACE_SEARCH, MATCH_PHOTODNA and others) were removed from core. OpenETL itself is not built yet, so the handoff was exercised with a stub.

The problem

Cameras and body-worn devices produce footage and detections. An operator needs the footage stored with its access class, the detections queryable (who was where, repeat plates, zone breaches), a timeline for an incident, and a record of who processed each file.

The shape

  • Store footage. Resumable uploads: POST /v1/files/uploads, PUT the parts, then complete. Content is read back with GET/HEAD /v1/files/content/... and listed with GET /v1/files. Files appear to SQL as S3Object rows. Upload with process=true and an access object to queue the file for processing; a request without access is refused (fail-closed).
  • Hand off to OpenETL. A worker holding a least-privilege media-processor token calls POST /v1/media/work/claim, extends its lease with heartbeat, and ends with complete, fail or requeue. Derived rows (captions, detections, embeddings you computed) go back through POST /v1/ingest/bulk with X-Relata-Claim, so a stale worker cannot write after its lease lapsed. Duplicate delivery, a lapsed lease, retryable and permanent failures and a Relata restart mid-job were exercised; completion only succeeds when Relata's own durable counts agree.
  • Query detections. Detection rows are ordinary types. Repeat visitors, hotlist matches and zone breaches are GROUP BY and exact-match SQL; a similarity search over embeddings you supplied uses the vector and hybrid search routes.
  • Timeline. An incident window query joins detections to the footage file ids, and AS OF SYSTEM TIME shows what the record said at a given moment.
  • Access. Classification and the caller's role decide who can read the file or the rows; masked cells stay masked in results.
  • Retention. Legal hold and erasure apply as described in Erasure, legal hold and backup.

Honest limits

  • No built-in video analytics, face or plate recognition, perceptual hashing or WORM storage.
  • A tamper-evident, signed evidence bundle is not part of the open build (report signing was retired; tamper-evident audit is Enterprise).
  • A similarity query whose supplied vector has the wrong dimension was observed to return rows instead of an error.

Exercised by

V3-16 (body-cam footage through claim, fenced write-back, retry, dead-letter and completion, 67 checks, passed), V3-01 (resumable 200 MiB evidence upload, 67 checks, passed) and the Vision V1-V7 journeys on the 3.0 port (evidence clips through the files API and a stub OpenETL, passed), relata-test battery, 2026-11-03, debug build. The vendor conformance script for OpenETL (V3-16b) recorded stale assertions and the vector-dimension defect above.