Postgres 19 Slips to November: AI Code Review Outran the Fixes

PostgreSQL 19 has slipped again. General availability is now around November 19, The Register reports, and a leading committer says part of the reason is LLM code review: it found defects faster than the project could fix them. Every team adding AI review to its pull requests should pay attention. Finding bugs got cheap. Fixing them didn't.

What actually happened

In September we covered the first slip, when SQL/PGQ property graphs were reverted and GA moved to late October. It has now moved again. In a blog post titled "The reverts will continue until morale improves", Peter Eisentraut, a leading PostgreSQL committer and VP at EDB, counts eight to ten significant features reverted from the 19 branch after beta started, some of them late in beta.

His diagnosis is a mix of causes. Some features had architectural defects. Senior developers lost time to security work during the spring. And LLM-assisted review now produces a long list of small issues quickly, each needing a fix with fixed overhead, so the team ran out of calendar. In his words, the "effective quality bar has been raised" since many of those features were committed. He says this is not about vibe coding, because most of the code was written before that was possible. He believes about two more months would have cleared the long tail, but the community preferred to keep the schedule.

A correction to some roundups: they reported that AI reviewers found "40 bugs." The Register says Eisentraut used "a list of forty defect reports" as a hypothetical example, not a count.

What still ships: REPACK CONCURRENTLY, logical replication of sequences, and pg_plan_advice. Several reverted features will likely return in PostgreSQL 20.

Why AI code review matters for your business

Budget for the findings, not the scan. An AI reviewer on your repo is cheap to turn on. The cost comes later: someone has to triage, fix, test and ship every finding. If one of the most disciplined open-source projects in the world ran out of time, a two-developer shop will too.

Triage by severity before anyone opens the list. Route security and data-loss findings to a human the same day. Batch style and edge-case findings into a weekly pass. Never let the reviewer block a merge on low-severity items.

Run AI review early, not at the end. Postgres found its problems during beta, when the cost of every fix was highest. Put the reviewer on every pull request so issues arrive one at a time.

Keep your upgrade on 18. Same rule as last month: plan against the current stable release and wait for 19.1.

Sources: The Register, Peter Eisentraut.