Root cause: SQLite's default PASSIVE checkpoint mode cannot complete
under sustained write load (4 nodes, hundreds of events/min) because
active readers always block it. The WAL grows unbounded and eventually
corrupts the main DB file.
- Add dedicated checkpoint_wal() using PRAGMA wal_checkpoint(RESTART),
which waits for readers to finish then checkpoints fully and resets
the WAL back to the start.
- Add background checkpointer task running every 30 minutes, separate
from the 6-hour prune cycle.
- Remove PASSIVE checkpoint from prune_old_events(), replace with call
to checkpoint_wal() after pruning completes.
- Set wal_autocheckpoint=500 on every connection (~2MB threshold) so
SQLite's own auto-checkpoint also triggers more frequently.
- Reduce HOT_DAYS from 30 to 7: at current data rate (~500k beacons/day)
30-day retention would grow to several GB before first prune runs.
7 days keeps the DB under ~200MB and queries fast permanently.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
bot
9fd97e6925