Postgres autovacuum, again

Every few years I convince myself I finally understand autovacuum, and every few years a queue table proves otherwise. This one held maybe two thousand live rows and roughly four hundred megabytes on disk.

High-churn tables are their own category. The default thresholds are written for tables that grow, not for tables where the same rows are updated and deleted thousands of times an hour. For those, per-table settings are not premature optimisation, they are the minimum.

We ended up lowering the scale factor sharply for that one table and leaving every other table alone. Tuning globally would have solved the same problem while creating three new ones.