mirror of
https://github.com/yogeshojha/rengine.git
synced 2026-10-20 23:31:31 +02:00
Yogesh OjhaandClaude Opus 5
2b433a4e1e
http_assets.software was never declared on the model, so the components the probe parsed were dropped at insert and the five-minute backfill became the only writer. Its pending test was an empty list beside a tech, and a versionless technology yields an empty list, so those rows were queued for ever. Each tick rebuilt the scan with a delete and a reinsert, and a tick outran its own interval, so eight ran at once, deadlocked on the same scan and held every worker slot. The column is three-state: NULL is not evaluated, [] is evaluated with nothing versioned. The probe writes it at insert. The backfill reads NULL alone under an instance lock and matches only when it wrote a component. match_scan builds the scan's rows in a temp table, applies confidence and rank there, and merges them by fingerprint under a per-scan advisory lock, so a row whose columns did not move is not written and finalize, the nightly rematch and the backfill queue rather than deadlock. Backfill beat entries expire after one interval. software_cves carries the result tables' autovacuum thresholds. The dashboard store loads its cells beside the overview rather than after it. A full rematch of 441,044 rows now writes none of them. The dashboard overview went from 13.7 s to 0.8 s and its software count from 13.6 s to 22 ms with no heap fetches. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CjhsosXbGgSkBU5z5yPYiz
The file is empty.