Site unavailable in bursts for several hours
The site returned 502 errors in waves. Generations already running were unaffected and no balance was lost, but the studio and history were unreachable while it happened.
Feed queries pulled oversized model metadata stored alongside each generation, which exhausted the process memory and restarted the app in a loop. The heavy data was moved to a separate table, list queries stopped selecting it, and the feed was switched to cursor paging.