Postgres 19

Postgres 19

Michael, Nik, and Kirk discuss Postgres 19: why the release came in November and not in September, the features we like most, and what to check before you upgrade.
Tip: watch on YouTube. Since episode 193, the video shows charts and illustrations as we talk. https://youtu.be/VcDU-NRE3hI
 
Michael, Nik, and Kirk discuss Postgres 19: why the release came in November and not in September, the features we like most, and what to check before you upgrade. We cover REPACK (CONCURRENTLY), LISTEN/NOTIFY with many listeners, autovacuum prioritization, plan advice, WAIT FOR LSN, new defaults, and how to test an upgrade now that AI makes testing cheaper.
 
Here are some links to things we mentioned: 
Times are podcast times of the first mention.

Main link

The release

Episodes we referred to

REPACK

LISTEN/NOTIFY and autovacuum

Plan advice

Other features

Upgrade

About the numbers on screen
They come from the study brief above. Most are synthetic tests of setups we built; they are not failure or speed rates of real systems. All are for 19 RC2, not 19.0.
  • CVEs per release: https://www.postgresql.org/support/security/ (snapshot 2026-11-18, 229 CVEs), grouped by release day from 2019 to 18.7; releases without a CVE count as 0. 18.7 fixed 52 and 18.6 fixed 28; 2019-2025 releases fixed 1.5 on average (median 1).
  • Release credits: names in the release-notes acknowledgments of 12 to 19; commits in each development cycle (PostgreSQL git, data of 2026-11-11).
  • REPACK (CONCURRENTLY): 19.0, 16 dedicated vCPU, 20,000,000-row table, 16 writer sessions; median of 3 runs per point (2 valid runs at the first point).
  • LISTEN/NOTIFY: 1,000 listeners on their own channels, 4 notifiers, 8-vCPU dedicated host, median of 3 trials; our client was the limit, so 19 can go higher.
  • Autovacuum order: one old table (age 1.65 billion) among 120 bloated tables, 16-vCPU dedicated host, median of 3 trials.
  • Plan advice: advice generated for every query (always_store_advice_details), trivial query, small shared test host, 4 clients, 20 s, 2 trials; 11% against the modules off, 14% against them loaded and idle.
  • JIT: 22 TPC-H-derived queries, scale factor 10, 32-vCPU dedicated host, 3 trials; per-query numbers are exploratory. JIT itself did not change; the default did.
  • lz4 TOAST: insert of 200,000 jsonb documents of 16 KiB, 8-vCPU dedicated host, 3 trials. The table is 11.3% larger with lz4; if disk matters more, set default_toast_compression = pglz.
  • Plan changes: 4,273 benchmark queries (263 templates; 220 stable on 18.7) planned on 18.7 and 19 RC2 with the same data and statistics; 61 of the 83 changed templates come from one planner commit (a tie-break), shown by a build without it.

Notes (checked against the Postgres 19 RC2 source after recording)
  • [00:21:29] Parallel autovacuum is off by default: set autovacuum_max_parallel_workers. Autovacuum prioritization needs no change.
  • [00:27:12] Our test measured 11% to 14% lower throughput with advice generated for every query, on a trivial query (small shared test host, 2 trials).
  • [00:35:41] The lower timing overhead uses the CPU time-stamp counter on x86-64.
  • [00:38:18] Extended statistics in pg_dump are in the release notes ("Allow pg_dump to include restorable extended statistics"), and pg_restore restores them; dump with pg_dump --statistics.
  • [00:40:40] Correction: the removed settings cannot be set on a table. The upgrade trap is escape_string_warning, or standard_conforming_strings = off, set on a role, a database or a function: pg_upgrade --check passes, then the upgrade fails.

~~~

What did you like or not like? What should we discuss next time? Let us know via a YouTube comment, on social media, or by commenting on our Google doc!

~~~

Postgres FM is produced by:

With credit to:
  • Jessie Draws for the elephant artwork
  • Intro and outro music: our own arrangement of Apple Loops from GarageBand (royalty-free under the GarageBand licence)

Some kind things our listeners have said