Were PostgreSQL 19 Patches Reverted Because AI Found Bugs?

Summary

Tomas Vondra challenges the claim that PostgreSQL 19’s recent feature reverts were mainly caused by AI tools finding bugs too complex to fix before release. He reviewed 12 major reverts and classified only four as probably or clearly involving AI. The revert of support for more object types in CREATE SCHEMA appears to have originated in an Opus 5 review, while the removal of batching from referential-integrity fast-path checks may have involved AI-generated findings. In contrast, the other cases were attributed to human reviews, known concurrency or design problems, buildfarm failures, incomplete designs, excessive post-commit fixes, or features that were not mature enough. Vondra emphasizes that his classification is based on mailing-list threads and personal judgment, and that undisclosed AI use cannot be ruled out. His historical comparison also finds that PostgreSQL 19 was not obviously reverting more patches than earlier releases, although the timing differed: previous releases saw many reverts soon after feature freeze, while PostgreSQL 19 had a later cluster of large reverts and is expected to ship about a month late. He argues that the more important AI-related effect may be on security work rather than feature review. PostgreSQL had 44 CVEs by the time of the article, compared with about five the previous year, and AI-assisted reports may be consuming experienced contributors’ time. In this account, AI is part of a broader change in the project’s workload, but it is not the primary explanation for the recent feature reverts.