2026-01-18

Vacuum signals you should log before blaming the planner

Priya Desai · PostgreSQL · operations

Teams often interrogate join order before checking whether dead tuples skewed cost estimates. We keep a lightweight dashboard that snapshots n_dead_tup, last_autovacuum, and autovacuum_count whenever a regression ticket opens.

During the Indexing Telemetry Clinic we ask participants to attach those snapshots beside their baseline plans. The habit shortens triage because everyone sees the same environmental story.

If you already export pg_stat_all_tables to your warehouse, add a narrow view that only surfaces tables with growth beyond agreed thresholds. That keeps noise low while preserving forensic depth.

Finally, document who is allowed to trigger manual vacuum so on-call rotations stay predictable. Clear ownership beats heroic midnight runs.