Reproduce SQLite WAL Checkpoint Starvation With One Forgotten Reader

Short summary

A single forgotten read transaction can prevent SQLite WAL checkpoints from succeeding, causing unbounded WAL growth while health checks stay green. A controlled drill—open one reader, write from another connection for 60 seconds, sample WAL size—reproduces the issue. Operational alerts should tie to disk budget projections, and runbooks must cover finding the oldest reader, shedding writes, and deciding when a restart is safer than waiting.

• •One open read transaction blocks WAL checkpoints, causing unbounded growth while SELECT 1 health checks pass
• •Run a controlled drill: hold one reader, write from another, sample WAL size and checkpoint results every second
• •Alert on projected WAL growth vs disk budget, not arbitrary file size; fix reader lifecycle before adding aggressive checkpoint loops

Generated with AI, which can make mistakes.