- Shipped
- October 22, 2026 at 5:06 AM UTC
- Author
- Kamo
- Commit
- 8ae0573
The sweep's fleet-wide lock is held only while it queues work, which takes milliseconds, so each pod ran its own sweep a minute or two apart and queued the same unfinished articles. The translator's record of what it was translating is per pod, so each article was translated twice at once (both pods logged 'Queued 15 article(s)' for one backlog every ten minutes). A per-article claim in Redis, held while an article is being translated, makes the second pod leave it. It fails open: with Redis unreachable the article is translated, as before.