Changelog

Change is the only constant. All the works here gets corrected, updated and/or extended after publication. PostgreSQL releases are major factor, readers catch things I got wrong or misleading. This page is a honest record of that. Every publication and every revision, newest first.

2026

11 October

new How TABLESAMPLE picks rows TABLESAMPLE picks addresses, not rows: SYSTEM samples come in bundles, BERNOULLI with REPEATABLE drifts after an UPDATE, and a hashed key gives you the same rows tomorrow.

9 October

new Postgres relation files, filed under base/ The files behind a PostgreSQL table: relfilenode versus OID, the fsm, vm, and init forks, 1GB segments, and what od on the file shows that pageinspect on the buffer doesn't.
updated PostgreSQL Storage Internals New chapter 3, Postgres relation files, filed under base/: where a page lives on disk, the forks that share its relfilenode, and why the file and the buffer disagree until a checkpoint.

1 October

updated Instant database clones with PostgreSQL 18 Corrected the description of the default template1 database, the WAL_LOG strategy, and which files a single UPDATE touches.

28 September

new What REPACK (CONCURRENTLY) costs while it runs PostgreSQL 19 comes with online table rewrites. Measured against pg_repack and pg_squeeze REPACK (CONCURRENTLY) is the fastest and the lightest on WAL under writes. Here's what it does to vacuum, readers, writers and old snapshots while it runs, and what to check before you start it.
updated The Bloat Busters: pg_repack vs pg_squeeze Links to What REPACK (CONCURRENTLY) costs while it runs, which measures core REPACK (CONCURRENTLY) against both extensions on PostgreSQL 19.

26 September

updated The Bloat Busters: pg_repack vs pg_squeeze PostgreSQL 19 adds REPACK (CONCURRENTLY) to core. Updated the comparison with measured WAL volume, swap stalls, repeatable-read visibility, and concurrent memory limits for both extensions, plus pg_squeeze's 18.6+ output_plugin_libraries requirement.

17 September

new Where PostgreSQL stores row locks What SELECT FOR UPDATE, FOR SHARE, and a foreign key check actually write into the tuple header. t_xmax as a locker, the infomask bits that say so, MultiXactIds when two sessions lock the same row, and the pg_multixact files that grow behind them.
updated Read your own writes, off the primary Inside a transaction block, WAIT FOR has to come first. Since f7c0929170a it fails if the session already holds a lock, even with NO_THROW. REPEATABLE READ works too, as long as the wait runs before the first query. thanks to Sami Imseih

12 September

updated Introduction to Buffers in PostgreSQL Reworked the article around a live pg_buffercache session: what the buffer pool actually holds after a query, why a SELECT can dirty a page, and how eviction plays out. Also corrected the version that introduced vacuum_buffer_usage_limit.

11 September

new The unbearable lightness of one more index Thirty PostgreSQL schemas written by coding agents, loaded and measured. The indexes are competent; the cumulative cost is not.
updated Read your own writes, off the primary One naming cleanup landed after this was written. The command tag the statement returns is WAIT, while the beta documentation called it WAIT FOR everywhere, and the tag cannot change once 19.0 ships. The -hackers agreement is to keep the tag and rename the documentation, error messages, and reference page to WAIT instead. The syntax does not change, you still write WAIT FOR LSN '...', but if you match on error text or link to the docs page, the wording and the anchor will differ from the betas.

31 August

new Read your own writes, off the primary WAIT FOR in PostgreSQL 19 lets a reader block until the replica has replayed a specific LSN. Measured against sticky routing, sleeps, and synchronous commit, on a replica that misses the write 99% of the time.