Brian Reed: Hey — so I've been sitting with this PostgreSQL story since last night and I'm not sure I have it straight yet.
Eliza Ward: October 16, The Register. PostgreSQL 19 is delayed — November 19 now, instead of late September or early October. And the reason Eisentraut gave publicly is that LLM-assisted review turned up roughly 40 edge-case bugs the volunteer team couldn't get through in time.
Brian Reed: Hold on — Eisentraut is EDB, right? He's not just a committer.
Eliza Ward: VP at EDB, yeah. Which is the main commercial PostgreSQL company. So he's speaking with some institutional weight here.
Brian Reed: And he specifically went out of his way to say the reverts were not 'vibe coding.' I mean — that's an interesting thing to volunteer.
Eliza Ward: It tells you the framing was already out there. He knew how it would look from outside — AI tools in, bugs up, release slips — and he wanted to get ahead of the obvious read.
Brian Reed: So RC1 is already out as of October 15. RC2 comes November 12. The PostgreSQL Global Development Group basically built a second candidate into the schedule just to absorb this.
Eliza Ward: That two-RC window is the structural signal to me. That's not a slip — that's a decision.
Brian Reed: But here's what the two-RC framing actually hides — it wasn't the delay that cost the project. It was the features. Eight to ten of them, pulled from 19, deferred to 20. That's not a calendar problem.
Eliza Ward: Wait — deferred meaning PostgreSQL 20 now has to carry new features instead of being a clean bug-fix cycle?
Brian Reed: Exactly that. And, I mean — think of it like a building inspector who can scan every wall in a day instead of a week. The inspection didn't create the cracks. It just found them faster than the repair crew can patch them.
Eliza Ward: Right — so the headline 'AI broke the release' is actually wrong.
Brian Reed: That's the clean version, yeah. But — wait, actually SQL/PGQ specifically, the property graph query syntax added to the SQL standard in 2023, that one got pulled. And Tom Lane had already flagged it in September 2026, before the LLM findings even landed. He said shipping it in 19 would produce post-release bugs that couldn't be fixed until version 20. So that removal wasn't the AI's call.
Eliza Ward: Which means at least one of those eight to ten features was already on borrowed time.
Brian Reed: So the LLM review confirmed what Lane already suspected — or it piled on top of it. And the PostgreSQL Global Development Group had to untangle which problems were actually new findings and which were... already known and just harder to ignore now.
Eliza Ward: That's the part I'd actually push on. 'AI revealed a problem' versus 'AI made an existing problem unavoidable' — those are different governance failures.
Brian Reed: And the consequence either way is the same: PostgreSQL 20 is now a different release than it was supposed to be. Not a cleanup release. A feature release, absorbing the deferrals.
Eliza Ward: But that clean version — AI caught the bugs, team deferred the features, story over — that's actually the take I want to push back on. Because here's what people are saying: PostgreSQL is skeptical of AI tooling, so this is an outlier. An enthusiast ran an experimental tool at the wrong moment. Doesn't generalize. And I think that logic runs exactly backwards.
Brian Reed: Wait — backwards how?
Eliza Ward: The PostgreSQL Global Development Group is conservative. Structurally skeptical. That's documented. And LLM-assisted review still reshaped their release cycle. If anything that makes it a stronger industry signal, not a weaker one. A permissive team adopting AI tools and hitting problems — fine, expected. This team?
Brian Reed: Right — and there's a piece of this that I think actually makes the outlier framing worse. We don't know who introduced these tools. It's not confirmed whether the use was formally sanctioned by the Global Development Group at all. So it's not even 'the team ran AI review.' It might be 'someone on or adjacent to the team ran it and the findings had enough weight that the schedule moved anyway.'
Eliza Ward: That's its own governance story, actually. An unsanctioned — or at minimum unconfirmed — tool input reshaped a major open-source release. That's not a cautionary tale about misapplied tooling. That's a question about who has standing to introduce new process.
Brian Reed: And there's a Hacker News thread on exactly this — practitioners pushing back on the framing that this was just LLM static review. Some of them described deterministic simulation testing, scaling advanced techniques well beyond regression testing. So the tooling landscape here is, I mean, it's more varied than one model scanning code for bugs.
Eliza Ward: Which matters because — wait, the benchmark numbers actually land here. SynapseWire published a comparison in July 2026, 47 bugs tested across tools. GitHub Copilot came in at 74% overall detection. CodeRabbit at 80% on security issues specifically. These aren't experimental numbers anymore.
Brian Reed: No, and that's the part the 'outlier' framing can't absorb. If detection rates are that high across the ecosystem, every conservative team is eventually going to hit this same asymmetry — a tool that finds faster than humans can fix.
Eliza Ward: And that's where volunteer bandwidth becomes the actual constraint — which is a whole other thread. PostgreSQL 20 is the next observable signal on that.
Brian Reed: And that signal — PostgreSQL 20 — is actually more concrete than it sounds. Picture a database administrator at a mid-size logistics company. It's a Friday afternoon, she's got a maintenance window booked, migration to PostgreSQL 19 in October. She planned around REPACK CONCURRENTLY specifically — that's the feature that reclaims disk space without locking the whole table the way VACUUM FULL does. That one made it in. But SQL/PGQ didn't. So now she's redoing the project plan, rescheduling the downtime to late November, and explaining to her team why the graph query feature they scoped work around is sitting in version 20 instead.
Eliza Ward: And she has no visibility into when version 20 lands.
Brian Reed: None. And I mean — the structural problem is right there. AI raised the detection bar. But the repair capacity is still volunteer hours. That asymmetry doesn't close just because the tooling improves.
Eliza Ward: Commercially backed teams have paid engineers who can surge on this. A company like EDB can absorb a triage spike. The PostgreSQL Global Development Group — that's volunteers. The workload doesn't flex the same way.
Brian Reed: Right — so the tool scaled the problem-finding. The human side didn't scale with it. Forty bugs in a compressed window, and the team just... couldn't get through them inside the original window.
Eliza Ward: Which is actually — wait, that's the part that matters for version 20. Does the PostgreSQL Global Development Group re-adopt LLM review for that cycle, or do they modify how it's applied, or do they step back entirely? Because that answer tells you whether they read this as a quality win or a process problem.
Brian Reed: Those are genuinely different conclusions from the same event.
Eliza Ward: Exactly — and we don't know yet. That's the live question. Version 20 is the test case.
Brian Reed: And every volunteer-driven open source project is watching it, whether they'd say so or not.
Eliza Ward: And the honest version of where I've landed — PostgreSQL 19 is probably better software than it would have been. Eisentraut said the findings were real. Tom Lane's September warning lined up with them. So the bugs were there. AI just moved when they got found. That part I'll say with some confidence.
Brian Reed: But the cost of moving that discovery earlier — that landed entirely on volunteers who didn't agree to an AI-augmented workload.
Eliza Ward: Which is actually — I mean, that's the thing I can't resolve from what we have. Whether the PostgreSQL Global Development Group treats this as a net win and builds LLM review into version 20's cycle, or quietly steps back — that decision, when it becomes visible, is the actual story. Not October 16.
Brian Reed: And if the answer turns out to be 'it works, but only if you have paid engineers to absorb the surge' — EDB can do that, most volunteer projects can't — then AI code review isn't closing the gap between commercial and volunteer infrastructure. It's widening it.
Eliza Ward: We don't know yet. That's the live part.
