atlas warm no longer spawns a thread per accepted save — offering a full history was a thundering herd of warm threads (one per save) all opening DB connections and contending (the SQLITE_BUSY spam, ~1.1G memory peak). Now coalesced to one in-flight warm thread per repo with a redo flag so the latest HEAD still gets warmed if saves arrive mid-warm; the heavy graph compute stays on a private read-only connection, and the cache write goes through the main connection under a brief lock (no second writer → no SQLITE_BUSY) (v1.31.86)
$ koh steal kepr.uk/kepr@0c523400dccf
·
parent: 28566d9f7714
discussion
log in to leave a comment.