AI/Nov 20 edition/Reporting & analysis
InfrastructureCodingSafety

InfrastructureCompute, chips & cloud

PostgreSQL 19 ships two months late after AI code review surfaces a backlog of old bugs

PostgreSQL 19 arrived on 19 November 2026, about two months late, after LLM-driven code review surfaced latent bugs faster than maintainers could fix them. Graph queries and other features were cut, while replica read-your-writes consistency and non-blocking REPACK shipped.

From the Simon Willison article of November 19, 2026 and cited sources: The project missed its usual September window and pulled SQL/PGQ graph queries, temporal FOR PORTION OF updates, partition merge/split and online checksum toggling in Beta 4. [2] [5] [6]

SOURCES[2] The Register: Postgres 19 delayed after AI reviewers find… theregister.com[5] PostgreSQL 19 Beta 4 Released postgresql.org[6] Neon: Postgres 19 is late, and it's AI's fault neon.com

AI-assisted, not human-edited. Methodology Corrections Who owns this site

KEY FACTS
What happened
The project missed its usual September window and pulled SQL/PGQ graph queries, temporal FOR PORTION OF updates, partition merge/split and online checksum toggling in Beta 4. [2] [5] [6]
Original source
Simon Willison article, independent perspective
Brief by
AT Pulse research desk AI-assisted; publisher and contact on the desk page
When
November 19, 2026 article published
Evidence
9 sources across 6 sites

Reach: not measured yet

THE CORE IDEAS3 TAKEAWAYS
01

The project missed its usual September window and pulled SQL/PGQ graph queries, temporal FOR PORTION OF updates, partition merge/split and online checksum toggling in Beta 4. Committers say AI reviewers raised the quality bar after these features were committed. Most fixes were small, but there were a great many of them. The cut features are now aimed at PostgreSQL 20. [2] [5] [6]

02

The security load jumped. One November minor release fixed 52 CVEs and more than 140 bugs across versions 14 to 18, and 22 of those CVEs credit Anthropic or OpenAI security programs. Disclosed CVEs reached 96 in 2026, up from 7 in 2025. Back-branch commits rose about 60% year on year. The most severe flaw traces back to 2004. [3] [6] [8] [9]

03

Version 19 adds WAIT FOR LSN, so an application can read its own writes from a replica after a timed wait instead of maintaining custom stick-to-primary logic. It also adds REPACK (CONCURRENTLY) as a non-blocking replacement for VACUUM FULL and CLUSTER, and extensions that pin query plans without code changes. JIT is now off by default and TOAST compression defaults to lz4. [1] [4] [7]

WHY IT MATTERS

committers report much higher fix volumes, and 22 of the 52 November CVEs credit AI security programs.

Read the full assessment

Implication: AI auditing can become a release-schedule risk for open-source projects, while PostgreSQL users face urgent patching and can drop custom replica workarounds.

Executive brief

PostgreSQL 19 shipped on 19 November 2026, about two months after the usual September window. The delay came largely from LLM-driven code review. AI tools found old bugs faster than maintainers could fix them, and the project dropped SQL/PGQ graph queries and several other features to ship (The Register). The week before, a single minor release fixed a record 52 CVEs (PostgreSQL). Version 19 itself adds read-your-writes consistency on replicas (WAIT FOR LSN), non-blocking REPACK and plan-pinning extensions (PostgreSQL 19 announcement).

What changed and event timeline

  1. Beta 4 drops features

    SQL/PGQ, online checksum toggling, FOR PORTION OF temporal DML, MERGE/SPLIT PARTITIONS and the DDL-retrieval functions were reverted. At that point general availability was still planned for October ().

  2. Delay blamed on AI reviewers

    Committer Peter Eisentraut said LLM review had raised the quality bar after these features were committed. Each fix averaged about three lines, but there were many of them ().

  3. Committer quantifies the surge

    Robert Haas reported that back-branch commits rose 60% year on year (724 to 1,169) and master commits went from 1,037 to 1,443 ().

  4. Record security release and RC 2

    The 18.7/17.12/16.16/15.20/14.25 releases fixed 52 CVEs and more than 140 bugs. Version 14 reached end of life ().

  5. PostgreSQL 19 GA

    The release covers workload management, replication, performance and security (). Simon Willison picked out WAIT FOR LSN as its standout feature ().

Capabilities and access

PostgreSQL 19.0 is open source. Upgrades go through pg_dumpall/restore, pg_upgrade or logical replication (release notes).

Read the full section

PostgreSQL 19.0 is open source. Upgrades go through pg_dumpall/restore, pg_upgrade or logical replication (release notes). Headline features, all from the announcement:

  • WAIT FOR LSN for read-your-writes consistency on standbys.
  • REPACK (CONCURRENTLY), which replaces VACUUM FULL and CLUSTER.
  • pg_plan_advice and pg_stash_advice for pinning query plans.
  • Sequence replication via FOR ALL SEQUENCES.
  • TLS SNI through pg_hosts.conf.
  • Parallel index autovacuum.
  • INSERT … ON CONFLICT DO SELECT and JSON output from COPY.

Technical analysis for researchers and developers

  • Read-your-writes on replicas. After a write, the app records pg_current_wal_insert_lsn(). The release notes also list a no-throw option (release notes).
  • pg_stash_advice applies stored advice by query identifier, so a slow plan can be fixed without changing application code (announcement).
  • JIT is now off by default because its cost estimates proved unreliable. TOAST compression now defaults to lz4 (announcement).
Read the full section
  • Read-your-writes on replicas. After a write, the app records pg_current_wal_insert_lsn(). On the next request it runs WAIT FOR LSN '<lsn>' WITH (TIMEOUT …) on the replica before reading. If the replica hasn't caught up within the timeout, the command raises an error (Simon Willison). This replaces hand-rolled versions of the pattern, such as MediaWiki's code built on MySQL GTIDs. The release notes also list a no-throw option (release notes).
  • Plan stability. pg_stash_advice applies stored advice by query identifier, so a slow plan can be fixed without changing application code (announcement).
  • Changed defaults. JIT is now off by default because its cost estimates proved unreliable. TOAST compression now defaults to lz4 (announcement).

Claims and evidence

  • The roughly 2x speed-up on foreign-key bulk inserts is the project's own measurement (announcement). No independent benchmark of it has been published.
  • 22 of the 52 November CVEs credit AI programs: Anthropic's OSS program and Claude (alongside Ada Logics), OpenAI Codex Security and OpenAI's Security Research Team (HeroDevs).
  • PostgreSQL disclosed 96 CVEs in 2026, against 7 in all of 2025 (HeroDevs).
Read the full section
  • The roughly 2x speed-up on foreign-key bulk inserts is the project's own measurement (announcement). No independent benchmark of it has been published.
  • 22 of the 52 November CVEs credit AI programs: Anthropic's OSS program and Claude (alongside Ada Logics), OpenAI Codex Security and OpenAI's Security Research Team (HeroDevs).
  • PostgreSQL disclosed 96 CVEs in 2026, against 7 in all of 2025 (HeroDevs).
  • The causal link to AI comes from committers' own accounts (Neon, The Register). The Beta 4 notice itself gives reliability and a predictable schedule as the reasons for the reverts (Beta 4).

Context and prior work

  • CVE counts per PostgreSQL 18 minor release rose from 2 to 5, 11, 28 and then 52 (Vonng blog).
  • Haas describes the backlog as decades of latent bugs that LLMs are good at finding (Neon).
  • The most severe flaw, CVE-2026-6460, is a type-confusion RCE whose root cause dates to 2004 (HeroDevs).
Read the full section
  • CVE counts per PostgreSQL 18 minor release rose from 2 to 5, 11, 28 and then 52 (Vonng blog).
  • Haas describes the backlog as decades of latent bugs that LLMs are good at finding (Neon).
  • The most severe flaw, CVE-2026-6460, is a type-confusion RCE whose root cause dates to 2004 (HeroDevs).
  • Willison traces the read-your-writes pattern to MediaWiki, which already did this by hand on MySQL replication (Simon Willison).

Limitations, safety and contested findings

  • standard_conforming_strings is now forced on, RADIUS and MULE_INTERNAL encoding are removed, MD5 password logins trigger warnings, and some btree_gist inet indexes need rebuilding (release notes).
  • The November minors require reindexing GIN indexes and some bpchar B-tree indexes (PostgreSQL).
  • Contested reading of the CVE spike.
Read the full section
  • Breaking changes. standard_conforming_strings is now forced on, RADIUS and MULE_INTERNAL encoding are removed, MD5 password logins trigger warnings, and some btree_gist inet indexes need rebuilding (release notes).
  • Post-patch steps. The November minors require reindexing GIN indexes and some bpchar B-tree indexes (PostgreSQL).
  • Contested reading of the CVE spike. Ruohang Feng (Vonng) argues the high count is a healthy sign that the project is actively maintained, not evidence of decay (Vonng blog).

Business and practitioner implications

  • Teams on versions 14–18 should apply the November minors. Version 14 gets no further updates (PostgreSQL).
  • Read-scaled apps can drop custom "stick to primary after write" logic in favour of WAIT FOR LSN.
  • Re-enable JIT only where you've benchmarked a benefit.
Read the full section
  • Patch now. Teams on versions 14–18 should apply the November minors. Version 14 gets no further updates (PostgreSQL).
  • Replica routing. Read-scaled apps can drop custom "stick to primary after write" logic in favour of WAIT FOR LSN.
  • Analytics workloads. Re-enable JIT only where you've benchmarked a benefit.
  • Planning for AI-era security load. For open-source maintainers and vendors, PostgreSQL 19 shows AI auditing turning into release-schedule risk: expect more patches and more cut features. Reverted features such as SQL/PGQ are now aimed at PostgreSQL 20 (The Register).
FOLLOW THE EVIDENCE

The source trail.

Sources (9)
A LITTLE LESS NOISE. A LOT MORE CONTEXT.

Stay curious.
Follow the evidence.

Independent perspectives, the original sources, and room for the questions that don't have easy answers.

How we build the brief
Connect with us

Find us where you already read.