Case Studies

Postgres 19 Slips as AI Reviewers Outpace Human Fixers

Postgres 19's release has been pushed back because AI code review tools surface far more bugs per day than human maintainers can triage — a bottleneck problem, not a failure of AI.

Font size

What happened: how AI review tools delayed Postgres 19

On October 16, The Register reported that Postgres 19, originally scheduled for release this month, has been delayed. The cause isn't missing features — it's that AI-powered review tools are flagging more bugs than the core team can address[1].

How bad is it? The report cites a ratio of roughly 50 issues per day from automated scans versus about 10 that a human developer can seriously work through in a week. That number isn't an exaggeration; Postgres committers have mentioned similar figures repeatedly on the mailing list. The bottleneck clusters around the query optimizer and concurrency control — exactly the kind of code where AI excels at nitpicking but humans have to think hard about correctness.

No official delay window has been announced, but from public statements by maintainers, the slip is at least a few months, not weeks. Postgres has been around for decades and has seen release delays before, but this is the first time the reason is "the AI found bugs faster than we expected."

On Hacker News, the thread hit 82 upvotes with 17 replies. The most-liked comment: "Welcome to the future of software engineering, where the bottleneck is no longer finding problems but triaging them."

Why "finding bugs faster than fixing them" is the actual problem

The conventional wisdom says finding bugs is good. AI tools help you scan, you fix, everyone wins.

This time the equation flipped.

The root cause is how AI review tools work — to avoid missing real issues, they err on the side of reporting more. A potential null pointer dereference gets flagged. A edge case that has never triggered in history gets flagged. A bug that would require a race condition plus specific input plus a specific kernel version to reproduce also gets flagged.

The result: more than half of incoming issues turn out to be false positives. But you can't just dismiss them, because the cost of missing a real bug far outweighs the cost of reviewing noise. So each one has to be investigated, taking anywhere from thirty minutes to two hours.

I ran CodeRabbit on a 5,000-line project overnight and got 23 issues. After going through the code, 18 were confirmed false positives. Of the 5 real ones, 2 required constructing six layers of nested boundary inputs to trigger.

Scaled up to a 30-year-old project like Postgres — millions of lines of code, dozens of contributors — incoming work where noise outweighs signal becomes a disaster.

Software engineering has hit this problem for the first time: the bottleneck has moved from discovery to absorption.

Postgres isn't alone: the real cost of AI review tools

Today's popular AI code review tools all share a "report first, filter later" philosophy. The reason: the model doesn't know your codebase's context, doesn't know your project's tolerance for errors, doesn't know whether a branch is test code. It can only maximize recall and leave precision filtering to you.

I've looked at the major players:

  • CodeRabbit: PR comments default to expanding "nit"-level suggestions, leading to volume explosion.
  • Sourcery: useful for Python projects, but frequently reports PEP 8 issues as correctness bugs.
  • GitHub Copilot Review: best integration, but also the highest false positive rate.
  • CodeAnt: experimenting with tighter context to suppress false positives — results pending.

When maintainers face dozens or hundreds of PRs each tagged "this might have a problem," morale tanks with them. In a Postgres mailing list thread, a long-time contributor wrote: "I used to love code review. Now I dread opening my inbox."

This is an organizational problem, not an individual one. A 20-person team suddenly absorbing 5x the incoming work, where noise outweighs signal, will see any process grind to a halt.

The real ceiling of AI software engineering

Finding bugs, generating code, running tests — AI keeps getting better at all of these, visibly and quickly.

But "deciding whether to fix, how to fix, and who owns the risk" is still human work.

These are fundamentally different tasks. The former is mechanical work; the latter is judgment work. AI isn't doing judgment anytime soon — not because it's technically impossible, but because nobody is willing to hand judgment over to it. When something breaks in production, who takes the blame? On a project like Postgres, whoever's code lands in the release notes owns the consequences.

Postgres 19's delay shows that AI's marginal utility has hit a wall at "decision support." This isn't a compute problem; it's an organizational capacity problem.

More concretely: AI generates 5 PRs per day, but the team can only properly assess 2 PRs of real risk per week. Until that throughput mismatch is resolved, AI making things faster is a net negative.

A few possible fixes:

  • AI triage agent: use another AI to filter out obvious false positives first, leaving only the suspicious ones for humans.
  • Risk-based scoring: have AI report not just the bug, but "how likely is a real user to hit this," letting humans prioritize by risk.
  • Maintainer-specific learning: each maintainer has a different noise tolerance; tools could learn individual filtering preferences.

But all of these fixes are 18 months away. For now, projects at Postgres's scale have to grit through with "let AI scan less, and have humans pre-filter what does come through."

This is a new kind of toil for the AI era, similar to the early days of SREs manually handling alert floods, except the source has shifted from monitoring alerts to AI reviewers.

Editorial thoughts

This story is different from every "AI boosted productivity by X times" narrative we've heard.

Those older stories are about AI replacing human work, saving time. The Postgres story is about AI creating more work than it saves. The cognitive cost AI saves on bug detection is close to zero, because review overhead eats all the time it freed up. This is Jevons paradox playing out in software engineering — when some step gets cheaper, it gets called more often, and total workload goes up.

I bet a new role will emerge in the next year: AI bug triage engineer, acting as a filter between AI review tools and human maintainers. The job title might end up as "AI code review specialist" or "AI tooling reliability engineer," or it might get absorbed into SRE responsibilities. Essentially, it's a human false positive filter.

I also bet Postgres won't be the last open source project to slip a release. Big, long-running projects like CPython, Apache Kafka, or the Linux kernel will see at least one release schedule adjustment over the next 12 months because "AI finds bugs faster than humans can handle them." I have no evidence for this; it's a prediction, written down to revisit in a year.

How is this different from before? Past "AI productivity" stories were one-sided: AI does more, humans do less, total cost drops. This one is two-sided: AI does more, and humans do more too, because they have to absorb AI's output. The shape of the production function in software engineering has changed. AI lets f(x) produce more code per x hours of development, but it also creates a nonlinear explosion in maintenance burden.

What I'm uncertain about: I can't tell whether Postgres is uniquely unable to absorb this, or whether all large projects will hit the same wall. Postgres has a well-known review process — every commit needs at least one reviewer approval plus a 72-hour cooling-off period — and that process amplifies the review burden AI creates. Projects using "merge first, review later" workflows might get off lighter. Most mainstream open source projects lean toward the former, which means this problem will be widely observed.


[1] The Register, "Postgres 19 delayed after AI reviewers find more bugs than humans can fix", 2026-10-16. https://www.theregister.com/databases/2026/10/16/postgres-19-delayed-after-ai-reviewers-find-more-bugs-than-humans-can-fix/

Source: https://www.theregister.com/databases/2026/10/16/postgres-19-delayed-after-ai-reviewers-find-more-bugs-than-humans-can-fix/5303471