Postgres 19 delayed after AI reviewers find more bugs than humans can fix
TL;DR Highlight
LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted
Who Should Read
Backend developer weighing AI code review in the release process and worried about handling the flood of issues
Core Mechanics
- Postgres 19 slipped to November
- It was slated for late September or early October and is now expected around November 19.
- Eight to 10 significant features were reverted and pushed to Postgres 20.
- LLM review is part of the reason
- EDB's Peter Eisentraut said in a blog post that LLM-assisted review at least partly explains it.
- Some features had architectural defects that were hard to fix quickly during beta.
- A long tail of small edge cases appeared
- Harmless issues once fixed over five years after launch are now found much sooner.
- Even three-line fixes carry fixed overhead each, so forty reports exhaust the time.
- The quality bar rose after features were written
- Review tooling got much better after the features were committed.
- Eisentraut said it should sort itself out unless LLM review keeps improving at this rate.
- Not blamed on vibe coding
Community Reactions
- QA is the bottleneck in the AI era, one view
- AI lays down a base of tests but tests only what comes to mind, like humans.
- It is very good at fixing problems once they are found.
- Weak test suite is the root cause, one commenter argued
- The pgrust author said Postgres relies on per-feature regression tests that miss edge cases.
- They said AI-driven simulation testing, fuzzing and formal verification found about 50 bugs in core Postgres.
- They verified 67k of 640k lines (10.5%) by translating C and Rust to Lean with LLM-written proofs.
- Skepticism about AI bug reports
- One commenter cited a Linux kernel case where only 6-8 of 82 reported bugs were real.
- Another asked that unedited AI reports not be filed as issues.
- Others delaying releases too, per comments
- One commenter said a flood of bug reports delayed their release.
- Another found many issues running AI review on old codebases.
How to Apply
- Run LLM review on one old module
- Feed one old core module to an LLM today and ask for edge-case review.
- Classify results as real, already fixed or false, and count accuracy.
- Filter LLM findings before filing issues
- Set a team rule to file only findings a human reproduced.
- Track time per issue and add buffer to the release schedule.
- Move AI review earlier
- Add LLM review mid-development or before merge to avoid late floods.
- Run design-stage reviews too so structural defects surface before beta.
Try It
# Paste into Claude or ChatGPT (simplified version of the article's idea: LLM review then triage)
Review the code below. Focus on edge cases (empty input, huge input, boundaries, encoding, concurrency).
{{code}}
Output each finding in this format.
- Location:
- Problem:
- Reproducing input:
- Fix size (estimated lines and whether structural change is needed):
- Confidence (high/medium/low):
Next step: classify results yourself as real, already fixed or false, and count accuracy.Terminology
Related Papers
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.
90% of US working hours are spent on tasks not covered well by AI benchmarks
Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage
Show HN: Gemini 4 Argon plays StarCraft
LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.
A Plausible Explanation Is Enough to Mislead Multimodal Judges
Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.
Neural networks are clean
Looking at neurons directly shows most have one main job, challenging superposition.
Linus Torvalds's first LLM-assisted patch series
Linus posted an LLM-assisted 10-patch series giving an order-of-magnitude read() speedup, with a race already found.
Show HN: Parsimony – an open-source benchmark for how much code AI models add