Postgres 19 Slips to November: AI Code Review Found More Bugs Than Humans Could Fix

PostgreSQL 19 will miss its release window, and the cause is novel even by the standards of a project famous for caution: the AI-assisted code reviews that contributors ran on the codebase worked too well. LLM-based review surfaced more defects in the late beta stages than the team could fix before the planned release, forcing core developers to revert 8 to 10 significant features and push the general release to roughly November 19. Percona’s community blog published an apology post titled “Apologies for delaying Postgres 19” on October 15, noting that a large share of the reported issues came from AI reviews.

Forty defect reports and a November deadline

Peter Eisentraut, a leading Postgres committer and VP at EDB, explained the arithmetic that broke the schedule. Some committed features turned out to have deep architectural flaws that could not be fixed “quickly and during beta.” Others had a long tail of edge-case defects that, in a pre-LLM release cycle, would simply have shipped and been discovered gradually over five years of post-release patching. “With LLM-assisted code review, these kinds of things can get found much faster, and then you’re staring at a list of forty defect reports,” he said. Five years of bug discovery compressed into a few weeks of beta is not a review win when your fix capacity stays human-sized.

The highest-profile casualty is SQL Property Graph Queries (SQL/PGQ), the SQL:2023-standardized graph query feature that was supposed to be Postgres 19’s headline attraction. Contributor Tom Lane had warned before the revert that shipping it in v19 would produce post-release bug discoveries that are “unfixable until v20.” It now lands in PostgreSQL 20 instead. Eisentraut was careful to note this was not “vibe coding”: the problematic features were written before AI coding tools were common. What changed is that review capability improved dramatically after those features were committed, and the new reviewers caught what the old process would have absorbed silently.

What survives, and what it means for AI-assisted engineering

Not everything got pulled. REPACK CONCURRENTLY, which rewrites tables without the locking VACUUM FULL imposes, made the cut, and Eisentraut promises the November release will be “very robust” precisely because of the cull. His prediction that the situation self-corrects once AI review capabilities stabilize is plausible, but it names the real lesson: generation and verification capacity have to grow together, and right now verification is growing faster.

That inversion is spreading well beyond Postgres. We have watched AI coding agents move from autocomplete to million-line migrations at Anthropic, and benchmark races like Cursor versus Codex treat harness quality as the differentiator. The Postgres episode is the first time a flagship open-source project has publicly delayed a release because AI review outpaced human triage. Expect it to become a normal engineering constraint: the bottleneck in AI-assisted development is shifting from writing code to adjudicating what the reviewers find. Projects that plan for that triage load, rather than discovering it in a release freeze, will ship on time. The ones that do not will spend their autumn reverting features.