PostgreSQL 19 dropped a series of planned features during its beta cycle, and the project gives its reasons: reliability first, a predictable release schedule, and design questions that were not settled in time. A second explanation for the pattern comes from Peter Eisentraut, a PostgreSQL committer, in a post published on 9 October 2026. In his account, LLM-assisted code review now surfaces defects faster than volunteers can fix them, so the effective quality bar rises after the code was written and the project runs out of time to work through the findings. He presents that as his own analysis, not as a project statement.

Which PostgreSQL 19 features were reverted, and when

The reverts arrived in three waves. Beta 2, released on 16 July 2026, reverted non-text output formats for pg_dumpall. Beta 3, on 13 August, reverted GROUP BY ALL. Beta 4, on 24 September, removed the largest group: SQL/PGQ property graph support, online enabling and disabling of data checksums, temporal updates and deletes using the FOR PORTION OF clause, ALTER TABLE … MERGE PARTITIONS and SPLIT PARTITIONS, the pg_get_role_ddl(), pg_get_tablespace_ddl() and pg_get_database_ddl() functions, and a change that forced LC_COLLATE to C in the postmaster process. The changes list in the release candidate 1 announcement, published on 15 October, contains fixes only, so the revert set appears to have closed with Beta 4.

Robert Haas, a PostgreSQL committer, added another item in a pgsql-hackers thread: REPACK’s prerequisite, database-specific logical replication snapshots, “was reverted as fundamentally flawed”, which left REPACK CONCURRENTLY restricted to a single process cluster-wide in version 19.

The release still ships its headline features. Eisentraut picks out REPACK CONCURRENTLY, logical replication of sequences, itself a previously reverted feature, and planner hinting with pg_plan_advice.

The project’s stated reasons: reliability, schedule and unsettled design

The Beta 4 announcement states the project’s reasoning directly. “The PostgreSQL community strongly believes that, first and foremost, PostgreSQL must be reliable,” it says, adding that “the community strives for a predictable release schedule”. Features were removed so that they could “potentially be included in a later major release after more time working through the community development and review process”.

For SQL/PGQ, the PostgreSQL release management team’s message of 2 September 2026 lists the open questions: orphaned label and property metadata, whether graph scoping should be global or per label, whether property graphs should get a pg_class entry, and a dependency loop in pg_dump support. The team concluded that “there is insufficient time to fix these in time for the release” and that reverting “would make it easier to fix these things properly in 20 without having to be burdened by backwards compatibility and backpatching”.

The SQL/PGQ revert commit itself, b1f106c8 by Eisentraut on 7 September 2026, lists the commits it undoes and links to the discussion, but its message gives no rationale of its own. The feature being removed had been committed on 16 March 2026, before the April feature freeze.

No project source located in this research attributes the reverts to LLM-assisted review. The stated reasons are design questions and insufficient time, and the release management team’s “insufficient time to fix these” is the closest any of them comes to the capacity argument Eisentraut develops.

Eisentraut’s account: review throughput outrunning fix capacity

Eisentraut counts “about 8 to 10 significant features, depending on how you count”. The project’s own list mixes substantial features with smaller items such as the three pg_get_*_ddl() functions and the LC_COLLATE change.

The mechanism he describes is about processing capacity. A maintainer working through LLM-assisted review, he writes, ends up “staring at a list of like forty defect reports”, and “there is a fixed overhead to dealing with each of these, and then you run out of time”. The defects are being found, but they still have to be read, triaged, assigned and fixed by people with finite hours. He adds that most if not all of the affected features were written long before LLM-assisted review was as useful as it is now, which means, in his words, that “the effective quality bar has been raised after the time the feature was written and committed”. That is his framing rather than a measured threshold.

He also points to a workload collision. Many senior developers were busy from roughly April 2026, when the feature freeze for PostgreSQL 19 began and the security work intensified, until August, when the most recent security release shipped, so the usual extra review and testing during the beta period was reduced and then resumed. The workload itself is a project fact: the 13 August 2026 release fixed 28 CVEs and more than 110 bugs across supported versions.

He rejects the suggestion that the features were carelessly generated.

No independent defect count exists for PostgreSQL 19

The “forty defect reports” figure is illustrative prose from Eisentraut’s post, not a measured project count for the release. No announcement, release note, commit message or mailing list post located in this research provides an audited defect count for PostgreSQL 19 or for any reverted feature. The project wiki’s PostgreSQL 19 Open Items page, the natural place to look for a running total, records individual open and resolved items with their commits and fixes, but it publishes no defect total for the release. That page is a community-edited project wiki, not an official announcement, and it was last edited on 27 October 2026.

The nearest quantitative artifact is a pgsql-hackers thread of 25 August 2026 in which Haas asked Claude to rank the scariest version 19 patches by the number and type of bugs fixed after the feature freeze: roughly 16 fixes for RI fast-path foreign key checks, 28 for REPACK, about 25 substantive fixes for online data checksums, 17 for UPDATE/DELETE FOR PORTION OF, 17 for SQL/PGQ, and 7 for postgres_fdw statistics import. Those are fix counts produced with an LLM at one committer’s request, posted to prompt discussion, and dated before the September reverts.

The alternatives Eisentraut weighs against his own explanation

Eisentraut sets out the competing readings himself. The first is that the features were not good enough: “Maybe we just didn’t do a good enough job. Maybe the code was just not good enough.” Tom Lane, a PostgreSQL committer, said in the same 25 August thread that he would “bet dinner” that shipping SQL/PGQ in version 19 would produce “post-release bug discoveries that are unfixable until v20”. The second is coincidence: the project may have attempted several features that turned out to be too hard in the same cycle. The third is the reviewer diversion described above. The fourth is system complexity, an interaction matrix in which defects are easy to find and expensive to fix.

Alongside those readings, the project’s own reasons for the reverts remain unresolved design questions and a preference for keeping to the schedule.

Where PostgreSQL 19 stands: RC 2 on 12 November, GA on 19 November

The release candidate 1 announcement, published on 15 October 2026, states that “the planned date for the general availability of PostgreSQL 19 is November 19, 2026”, with a second release candidate planned for 12 November 2026 unless an issue requires a delay or an additional candidate. It explains the November date as a way to bundle the fixes from the scheduled 12 November update release into PostgreSQL 19 before general availability. It does not attribute the date to LLM-assisted review.

Some project pages lagged that announcement: the roadmap page still described PostgreSQL 19 as planned for October 2026 as of 26 October 2026. The date can still move, since the RC 1 announcement allows for another candidate or a delay.

The reverted features are expected to return in PostgreSQL 20

The Beta 4 announcement says removed features “can potentially be included in a later major release” once more development and review time is available, and the release management team described the revert as a way to “fix these things properly in 20”. Eisentraut writes that he is “confident that many of the reverted features will come back before too long” and that the project has “a few nice 80%-ready features queued up for PostgreSQL 20”. The Register reported that the reverted features “will likely make it into PostgreSQL 20”. None of this amounts to a commitment or a schedule.

What the PostgreSQL 19 reverts show about review throughput and fix capacity

PostgreSQL 19’s beta cycle is a worked example of a review process changing the outcome of a release without a single dramatic failure. When assisted review increases the number of usable defect reports reaching maintainers, a project faces more decisions than it has people to make them, and those decisions resolve into delays or removals. The release management team chose removals, protecting both the schedule and the unfinished design work.

The reverted features were written before LLM-assisted review became as effective as it is now, so the episode does not show that assisted review created defects or that PostgreSQL has an AI-specific quality problem.

Eisentraut’s own expectation is that the situation “should sort itself out for future development cycles” unless review capabilities keep improving at the current rate.