You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Pre-check can no longer be skipped silently on a process claim. A missing search index or SSCD model, a missing duration, a failed cut or a failed analysis now fails the claim (transient, retried) with the reason, instead of logging "skipped" and going on to frames with no record.
The pre-check decision is on the final done line of a process claim ("pc":"new|partial|canonical", read from precheck.json), PASS included; a duplicate still carries it in its trash/why fields.
Claim order on a running node is covered by a test: a pri line appended by another process while a claim runs reorders the next claims (priority, then added, then id).
The search index builds itself as a queue job (search-index): a follow-up is due 15 minutes after a build when more centers finished while it ran, a catch-up build runs when the head is behind the centers on the store, and each version records builtAt (the time it started).
Old search-index versions are kept 3 back (search.keep-versions, default 3; 0 keeps only the head); older index/search/vN/ are deleted after head.yaml moves, and a reader that gets a 404 re-reads head.yaml once and retries.
The catalog (catalog.json.gz, one per search-index version) lists every upload the pre-check has seen, with duration, title, channel, speech language, dubbed, AI score/flag and the pre-check's match (upload, share, offset, cosine). Every entry has four status fields: state (processed | metadata-only), held ("ai" when held, else null), checkedBy (index | precheck) and data (full | precheck: sample estimates, no centers in the index). A held upload that is processed later is replaced by its full entry.
AI dramas can be held: the pre-check flags an AI drama and the process claim releases it as held; indexnode control ai-dramas on|off (off: held subjects are not claimed, on: they resume at frames) and indexnode jobs list --held ai (with the score).
The matched upload's duration comes from download.zip's idx_summary metadata (HEAD only, never a hydrate), then the catalog; before, the repacked layout had no downloaded.json and every duplicate was "duration unknown: not skipped".
Uploads newer than the search index are found by brute force: the pre-check also compares the sample with the centers of uploads whose centers finished after the index's builtAt.
A zero-length ranged read no longer sends an invalid range (to=-1): reading an empty range of a blob now returns nothing instead of failing.