- 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.
Huge respect to Postgres folks, and the features are amazing.
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.
Major upgrades for Postgres are more complicated. So it makes me wonder if they could do a bit less major releases and release some of these features in point releases?
First off, one does not need to upgrade to each major version, it is perfectly fine to jump over one or more major releases in one step.
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.
That's what LTS releases are for. And you don't need to adopt them at point .0, you can adopt them after one or two minor releases are out for the LTS version.
Right there with you. I'm still on Postgres 16, both for work and personal projects. I like being a major version or two behind given the pace of change.
okay, if you don't want to use new features, just don't...
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.
Copy-and-patch is going to generate worse code than the LLVM-based JIT, all things being equal. You'll get shorter compile times, but in my experience, that's only really a part of the JIT's problems. TBH it's a bit sad that they turned on the JIT in the current state to begin with; it clearly wasn't ready, and perhaps needed a couple more rounds with the design.
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.)
Personally I would say, for critical production systems, its still good to wait for the .1 release.
Part 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.
> Part of why delaying is helpful is that more people will actually install the release than they do during the beta/RC cycle
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.
Yep, agreed its bad if everyone does it - but e.g. you could upgrade pre-prod / staging and only do prod once the 19.1 release is out.
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.
Yes, my strategy is always last on dev + what I have on production (ie: I have PG18, PG16). Just changing the port make simpler to test problems both ways.
It really depends on your use-case.
Personal use database and you have regular backups, can tolerate some bumps - probably fine to test drive it, but keep an eye open.
Serious production system - let it "soak", as they say these days. My personal rule of thumb is 6 months for major versions, it has proven itself many times over many different technologies.
I always wondered why sequence numbers weren't replicated, it added another step to the stress of upgrading via pg_backup but thanks for adding it to PG19!
For performance reasons, sequences have always been non-transactional in Postgres. This implies that they are not in the transaction log which is being replicated to standbys.
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.
> For performance reasons, sequences have always been non-transactional in Postgres. This implies that they are not in the transaction log which is being replicated to standbys.
This cannot be true because physical replication already supported sequences. Only the support for logical replication is new.
- 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!
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.
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.)
Part 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.