• Sources: primary
  • Summary: Prompted by GitLab's plan to drop a 4 TB table with ten replicas, Radim Marek measured on PostgreSQL 18.6 that DROP TABLE writes about 80 KB of WAL while the standby startup process unlinks the files and holds replay. A cold 1 TB drop took 16.1 s on ext4 against 80 ms on XFS, with cost tracking extent count at about 17 ms per segment plus about 34 ms per cached GB. Batched deletes of 25.2 million rows wrote 8179 MB of WAL, mostly full page images.
  • Why it matters: Before dropping a large table, check the filesystem and page cache on every replica, set a short lock_timeout, and enable log_recovery_conflict_waits.