PostgreSQL 19
postgresql.orgPart of why delaying is helpful is that more people will actually install the release than they do during the beta/RC cycle, and (in my read) most of the delay of the release was not necessarily because of significant bugfixing of features that remained in 19 (though e.g. REPACK certainly had its fair share of fixes), but rather the Postgres committers being overly busy with some of the security fixes that just went into 18.7 and the other minors, combined with some late removals of features that warranted an extra beta and RC release.
Unless everybody starts waiting for .1 :)
> was not necessarily because of significant bugfixing in 19
They discovered some serious bugs in new features which weren't really solvable in a way the devs were comfortable with (with regards to long-term maintainability and stability). So some of the new features were rolled back and that process also took some time.
To clarify what I meant with the second part: If you look at e.g. SQL/PGQ it was reverted pretty late (beta4), but that revert doesn't mean that other features don't have bugs that are yet to be ironed out. e.g. just this morning a bugfix went in that can crash EXPLAIN (BUFFERS) (without ANALYZE) when you have auto_explain turned on [0]. Edge case, but could happen, and the fix will only be in 19.1 in February 2027.
[0]: https://git.postgresql.org/gitweb/?p=postgresql.git;a=commit...
As I understand it, the new feature is not a continuous replication (for the above reason), but rather a one-off synchronization of the current sequence values of a leader to its followers, you'd need to run it again later for additional syncs.
This cannot be true because physical replication already supported sequences. Only the support for logical replication is new.
1) If you do BEGIN; SELECT nextval('someseq'); ROLLBACK; the value from the rolled back transaction won't be reused, despite the rollback. That's quite crucial for performance, as otherwise concurrent transactions would not be able to get a sequence value until the first transaction commits/rolls back.
2) If you crash (PG or HW), some sequence values may never end up being used. For performance only every 32nd sequence advancement is WAL logged.
- REPACK CONCURRENTLY to reclaim disk space, which also affects performance.
- Logical replication of sequences - Fantastic addition. This is a missing piece in the major-version upgrade story which required extra scripting, and enable additional usecases.
- WAIT FOR LSN is a game-changer for read-replicas. For read-after-write, where the write goes to the master and the read hits a random replica. And also read-after-read, where each read may hit different replicas.
Add to that additional performance improvement and the plan advice extension, this makes 19 one of the most exciting releases of PostgreSQL.
Big big big kudos to the contributors, thank you!
That said, I don’t see the point. If you need to read DB-generated data (like a PK, or timestamp) from a write, use RETURNING so you get it with one call, and pass that to another function if it needs it. If it’s data you wrote, you already have it. So many web apps are slow because they have a million services, and all of them are calling the DB for the same information instead of passing it to each other.
That was long overdue. The performance impact on perfectly fine tuned queries was horrendous.
It's also disabled by most OSM apps that make heavy use of PostGIS for exactly that reason.
Maybe it wouldn't be that bad if a copy-and-patch kind of JIT would be used?
Out of curiosity, are there any copy-and-patch JITs that are clear-cut successes? (Python's JIT is the most high-profile case, and it's definitely not one.)
I believe Wirth and Franz's JITs for Oberon were of the template/copy-and-patch variety, and I believe they were able to intertwine caching of expanded templates inside their bytecode decompressor, allowing it to JIT compressed bytecode off of disk into memory faster than reading native binaries from disk. Though, that was largely due to the ratio of CPU speed to disk read speeds on the workstations at ETH Zurich.
But an apologetic opinion: Postgres 18 was released in Sep 2025, and it's just Nov 2026. Half the companies I know haven't even updated to pg17 yet, let alone pg18's async I/O or anything. There are already so many features that are unadopted and unproven at scale.
IMHO the last software that I want to see feature creep or a large amount of major version bumps, is in my Database.
Postgres has a huge backlog of much requested features and I take it positively that they are finally shipping these.
Also, the definition of a major release in Postgres is no new features, and, by extension, no system catalog changes. When the Postgres system catalog changes, you need to make a major version upgrade (this is e.g. the case if you want to upgrade from Postgres 19 beta3 to GA).
A lot of Postgres' reputation is on their solid minor releases, and they will not compromise that.
just because you don't want to use anything new, doesn't mean they should stop development. postgres 15 is still officially supported, 14 is still community supported. being a few versions behind is fine and has nothing to do with new versions.