PostgreSQL 19 Pulls 8 to 10 Features and Sets GA for November 19, and Its Core Developers Disagree on Whether AI Review Is to Blame

The schedule slipped a month

The PostgreSQL Global Development Group released PostgreSQL 19 RC1 on October 15. The announcement says the regular update release on November 12 will carry a large number of fixes, which the release team wants in 19 first, so a second release candidate will ship that same day and general availability is planned for November 19. After many years of on-time autumn releases, this one is about a month late, because a batch of already-committed features was reverted during beta, including SQL/PGQ property graph queries, online data checksum transitions and UPDATE/DELETE FOR PORTION OF.

"A list of like forty defect reports"

Committer Peter Eisentraut wrote on his blog on October 9 that 8 to 10 significant features had been reverted, where about one per cycle would be normal. Among the causes he lists: some features had architectural defects that couldn't be fixed quickly during beta, and there was also a long tail of relatively harmless edge cases. In the past, those would have surfaced over the next five years and been fixed bit by bit; with LLM-assisted code review they are found quickly, "and then you're staring at a list of like forty defect reports." Even at three lines per fix, each one carries fixed overhead, and time runs out.

He stresses that most of the reverted features were written before AI review was this useful, so the effective quality bar rose after they were committed, and that this was "not due to vibe coding." Another factor was that between April and August many senior developers were tied up with security fixes, so the usual beta-period review shrank.

A different reading, revert by revert

In an October 1 post, committer Tomas Vondra pushed back on the idea that AI had found bugs too big to fix. He went through the mailing-list threads behind 12 reverts from May 20 to September 16 and concluded that only 4 came out as "AI" or "probably AI"; the rest came from ordinary human review. Counting reverts by day of the development cycle since PG14, he found PG19 unremarkable in number; what changed was timing. In earlier cycles reverts clustered shortly after feature freeze; this time there was a flat stretch and then a burst in early September. In his view AI shows up more in the security queue: about 5 CVEs last year, 44 so far this year, some reported with AI help, and handled by the same people who review features.

Christophe Pettus added on his blog The Build on October 12 that the only revert clearly triggered by AI was the CREATE SCHEMA expansion. Noah Misch ran an Opus 5 review; most of the report was noise, one finding he recognized as serious, and the revert ultimately rested on his own design judgment.

For teams reviewing code with AI

The disagreement is about attribution, not the facts: AI review and AI-assisted security reports really do bring problems in faster, and the bottleneck becomes how much experienced people can triage and fix. The older a project and the more feature combinations it supports, the more it needs to set aside people and time for that inflow in advance.