A poster of 2.7 kB broke the index build (a btree key is at most a third of a page) and left an invalid index
that IF NOT EXISTS would have taken for the real one: the migration drops it first and indexes md5(poster), and
the sweep asks by it.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The discard sweep of a dropped signature read the whole pending_releases table per signature (a parallel seq
scan over the poster test, four minutes on the HDD that starved the lanes' claims): a partial index on
(group_id, tool, poster) WHERE status = 0 makes it a range, and the sweep queries a poster's rows by it and a
tool's and group's by the ladder index.
The batch store's pending upsert, the record, and the sweep deadlocked on pending_releases and pending_classes
rows locked in different orders, and a deadlocked flush dropped its outcomes: lockOrder() sorts every
multi-row statement's set the same way, and a flush that deadlocks anyway keeps its outcomes for the next.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>