Postgres 19 delayed to November 19 — AI code review found too many bugs
Three releases, three different reasons to stop and look: a database delayed because AI review found more bugs than anyone could fix in time, an export-control draft that decides which models China may ship, and a browser assistant that answers without being able to read the question.
PostgreSQL 19 will not ship this month. The PostgreSQL Global Development Group has pushed general availability to November 19 and reverted between eight and ten significant features — and the more interesting news is a leading committer's own account of why: LLM-assisted code review surfaced a long tail of edge-case defects that used to take five years of point releases to find, and the project ran out of time to fix them all.
The mechanics first, because they are checkable. Release candidate 1 landed on October 15, release candidate 2 is scheduled for November 12, and the wiki now lists general availability as November 19 — against a published schedule that had GA on October 29, and an original Beta 2 target of September or October. Peter Eisentraut, a vice president at the consultancy EDB and one of PostgreSQL's leading committers, wrote on October 9 that the revert count is "about 8 to 10 significant features, depending on how you count," and that some of them carried architectural defects that could not be repaired "quickly and during beta." What changed between writing and shipping, in his telling, is the tooling: "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. Even if each of them requires only a three-line fix on average, there is a fixed overhead to dealing with each of these, and then you run out of time."
That sentence is the whole story, and it is a structural problem rather than a one-off. AI review raises the effective quality bar after a feature has been written, reviewed and committed, which means the backlog it exposes is retroactive and unbudgeted. Eisentraut expects the situation to sort itself out in future cycles only "unless the capabilities of LLM-assisted code review keep improving at the current rate" — which is a polite way of saying the tooling is winning the race.
The casualties are real features, not housekeeping. SQL/PGQ property graph queries, standardized in 2023 and the headline item of this cycle, are out over unresolved issues; contributor Tom Lane had warned in September that he would "bet dinner" the feature would produce post-release bugs unfixable until version 20. Also reverted: online enable and disable of data checksums, UPDATE/DELETE ... FOR PORTION OF, ALTER TABLE partition merge and split, the new role, tablespace and database DDL getters, and a forced C collation mode. Earlier betas had already pulled GROUP BY ALL and non-text output formats for pg_dumpall. What survives includes REPACK CONCURRENTLY, which reclaims disk space without the table-locking VACUUM FULL.
Two honest counterpoints. Eisentraut is explicit that this is not vibe coding — most of the reverted code was written before that was possible — and he names a second cause entirely unrelated to models: senior reviewers spent April through August on a run of security fixes, which cut the beta-period scrutiny the project normally relies on. And the project has been using models on the review side for months: in August, committer Robert Haas asked Claude to rank which version 19 patches looked scariest by the number and type of post-freeze bug fixes, a exercise that got its own write-up in LWN. The models were already in the room. What changed is how loudly they could say no.
China's commerce and science ministries opened a two-week consultation on Friday over a revised export-control catalogue that would put large AI models, agents and self-rewriting malware tools on the restricted list — and it decides whether a model counts as a controlled export by three numbers: 2 trillion parameters, 10^26 FLOPs of training compute, or a top-10 placing on public global rankings. The draft, published as Announcement No. 48 of 2026 with comment closing November 15, covers 40 technologies, 23 of them AI, alongside 12 solar and five biomedical entries, split into 14 outright prohibitions and 26 restricted items. The model controls reach weights, checkpoints, source code, architecture and training or inference methods, and they also catch models trained on data China classifies as "important" under its Data Security Law, with a carve-out of up to 0.1 percent of such data.
The carve-outs are where the policy gets interesting. Publishing restricted AI technology or weights as open source qualifies for automatic licensing rather than case-by-case review, as do Chinese firms running their own data centres abroad, taping out chips overseas, doing joint research with foreign partners and helping set international standards. Technology already in the public domain through an open-source release or a paper is exempt entirely, and using a restricted model through an API, an app, a website, a SaaS product or a trimmed on-device version sits outside the rules — while installing it on a server abroad needs a standard licence. Beijing's stated reason for the fast lane is not wanting to choke its own open-source boom, and that is the tension worth watching: the same catalogue that lets the state switch off foreign access to Chinese weights leaves the open-weight channel running by design.
For everyone outside China the arithmetic is uncomfortable. American companies that have leaned on Chinese open weights now have a rule that can be turned on; the European Union's sovereignty plan — run open models in domestic clouds instead of renting closed ones — assumes those weights keep arriving. We covered the soft-power half of this in September, when Beijing used an APEC ministerial to build regional institutions around its model stack — China turns APEC into an AI export channel for small business. Interim registration applies from Friday; the catalogue itself is still a draft.
Brave switched on end-to-end encryption for Leo, its browser assistant, so the machine answering your question cannot read it — and it runs that traffic on a model whose weights anyone can download. The feature is live in Brave's Release, Beta and Nightly builds as of October 15, scoped for now to GLM 5.3 Flash, the open-weight model Z.ai released under an MIT licence in August and which Brave still labels beta. Prompts, any page content you share, the conversation history and the streamed replies are encrypted on the device; the keys live only inside a confidential-computing enclave run by NEAR AI on Intel TDX processors and Nvidia Confidential Compute GPUs, with attestation through Intel's signed quote and Nvidia's remote attestation service. The transport uses Oblivious HTTP, so Brave's gateway relays the request without passing on your IP address and the enclave never learns it.
Read the fine print before you call it verified. The attestation check still happens on Brave's side rather than in your browser, and the log of enclave keys Brave says it will publish does not exist yet — so "verifiable" is a roadmap item, not something you can run today. Brave chief executive Brendan Eich's framing is the reason to care anyway: "AI has raised the stakes, yet too much of it still runs on 'trust me bro' promises, often at the user's expense." With closed assistants, privacy is a policy you accept. Here it is at least an architecture you can inspect, once the browser-side check ships.
What to watch: whether PostgreSQL 20 absorbs the reverted features and whether other projects adopting LLM review hit the same retroactive wall; whether China's catalogue survives consultation unchanged and whether any lab tests the open-source fast lane; and whether Brave moves attestation into the browser before someone finds a reason to doubt the enclave.
If AI review finds the bugs faster than humans can fix them, who pays for the fix — maintainers, the labs, or nobody? Tell us in the comments.
