[goruck]

Part of #706. #706 stays open: this fixes the recent-caption scan, and one of the issue's three readers (the similarity search) is unchanged by design. See below.

What was wrong

The langgraph Postgres store matches a namespace as a text prefix (prefix LIKE 'video_analysis.side%'), so a read for camera.side also returns camera.side_gate's captions. _ in a name is a LIKE wildcard as well. The recent-caption scan reads a camera's notified captions from the last 30 minutes, up to 200 rows in 4 pages. Rows from other cameras were dropped in Python only after the store applied its limit. So a busy side_gate could fill all 200 rows, side's own recent notifications were never read, and a repeated scene notified again.

Changes

• Captions name their camera. Each stored caption gets a camera field, and the recent-caption scan filters on it in the database, before its page limit.
• Background backfill (core/caption_store.py) adds the field to captions stored before it existed:
  • Each start checks for captions without the field. This also catches captions an older version writes after a downgrade, or from another install on the same database.
  • One run per database, using an advisory lock. A run that loses the lock checks again every minute until nothing is left.
  • It walks the primary key in batches of 2000 and pauses as long as each batch took. updated_at is never changed.
  • Statement and lock waits are bounded. Lock timeouts, cancels, deadlocks and serialization failures are retried.
  • Each attempt closes its connection unless it finished cleanly, so no pooled connection is left holding the lock or the timeouts.
  • The task is cancelled at unload, before the pool closes.
• get_recent_camera_activity removed. It is the third reader Cameras whose names share a prefix read each other's stored captions #706 names, and nothing has called it since 1f68e4f.
• Unrelated TODOS entry included on purpose: the Show a stalled model backend and fail fast instead of waiting out long timeouts #828 case where a frozen model server holds a turn past its budget. It was promised to the reporter, and this was the next PR.

What is not fixed (why #706 stays open)

The similarity search (_search_own_captions) is the same as on main: no database filter, a wide limit of 50, and the Python namespace filter. A camera with 50 or more closer matches whose name extends this one's can still push this camera's matches out. That gives no_match, the scan is never reached, and a static scene can notify again.

The two database-side fixes I measured both trade one problem for another. Measurements are from pg17 with pgvector 0.8.7, on 100k synthetic captions with 1024-dimension vectors and an HNSW index:

Variant	Busy camera (45k)	Mid-size camera (25k)	Small camera (8k)
main (no filter)	45 ms, 7/10 own	fast, partial	1–3 own
camera filter	617 ms (exact scan: no statistics on the jsonb expression)	—	—
filter + CREATE STATISTICS	15 ms, 10/10	0.9 s (exact scan)	1–3 own (HNSW post-filter, ef_search=40)
+ hnsw.iterative_scan	—	0.9 s	10/10 rows

Prepared generic plans did not undo the statistics; timings stayed steady over 40 runs per connection. The search's fix belongs with #706 and #707. Measurements are posted on #706.

Testing

• make lint passes and make typecheck reports 0 errors. make test: 6490 passed, 1 skipped.
• test_caption_store.py covers:
  • batches walking the primary key and ending on a short batch
  • waiting while another run holds the lock, then taking over when it stops
  • retries, including on a deadlock
  • giving up after the retries
  • unlock failure
  • a cancel landing in a batch, during the pause, during the lock release, or while the lock is being granted
• test_video_analyzer_novelty.py uses a fake store that matches namespaces the way LIKE does (_ as a wildcard) and applies the filter before the limit. Covered:
  • 200 notified side_gate rows no longer hide side's notification
  • a pre-Cameras whose names share a prefix read each other's stored captions #706 caption is invisible to the scan until it gets its camera
  • wildcard siblings are excluded
  • the Caption dedup: suppress on any recent notified match, not only the closest (#704) #705 tests that exercised the removed Python filter now run on the fake; they were reviewed and kept
• test_embedding_record.py: the backfill starts with a database, is cancelled at unload, and doesn't start without a database.
• Real pg17 + pgvector, using throwaway containers:
  • 100k captions backfilled in 8.7 s, every batch walking store_pkey
  • a single run, two concurrent runs, and a cancel followed by a resume: every row has the right camera, updated_at is unchanged, unrelated namespaces are untouched, no advisory lock is left behind, and pooled connections are back on their default timeouts
• Not field-tested on the box yet.

Review

gstack /review: 3 cycles, not converged, because cycle 3 still needed small fixes, all applied. Codex ran in every cycle, alongside the Claude adversarial pass, the red team, and the testing, maintainability, security, performance, data-migration and simplification reviewers. Main design changes during review:

• dropped a permanent done-marker
• switched to a keyset walk under an advisory lock
• cancel on unload
• tried planner statistics, then removed them
• limited the camera filter to the scan (Lindo's call)

Ships in the 3.45.0 bundle.

🤖 Generated with Claude Code

https://claude.ai/code/session_019kuwyteUa5Q6rkkB8hgwwj

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.