PostgreSQL 19 came out on November 19, 2026. For most teams, upgrading is a planned migration, not an emergency. The release adds built-in non-blocking table repacking (REPACK), plan pinning, logical replication without a restart, and INSERT ... ON CONFLICT DO SELECT. It also turns JIT off by default, removes RADIUS authentication, and adds new conditions under which pg_upgrade refuses to run. If you're still on PostgreSQL 14, the date that matters more is November 12, 2026. That's when the community stopped supporting it.
This guide covers what changed, which checks to run before an upgrade, and how to time the move on self-hosted servers and managed platforms.
What shipped in PostgreSQL 19
The official announcement focuses on running large workloads with less disruption. These are the changes builders are most likely to use:
REPACKandREPACK (CONCURRENTLY). One command now reclaims space from bloated tables and can reorder a table to match an index. It replacesVACUUM FULLandCLUSTER, which still work for compatibility. The concurrent form doesn't block reads or writes. Until now, that required third-party extensions such as pg_repack.- Plan control. The new
pg_plan_adviceextension lets you pin or experiment with planner choices using a small "plan advice" language. Its companion,pg_stash_advice, applies saved advice automatically by query ID. That means you can fix a regressed query plan without touching application code. - Easier logical replication. The effective WAL level rises to
logicalautomatically when the first logical slot is created, so you no longer need a restart. Sequence values can be published withCREATE PUBLICATION ... FOR ALL SEQUENCES. A newWAIT FOR LSNcommand gives read-your-writes consistency on standbys. - Developer features.
INSERT ... ON CONFLICT DO SELECT ... RETURNINGreturns the existing row on conflict, which turns "get or create" into a single round trip.COPYcan output JSON.lead(),lag(),first_value(),last_value()andnth_value()acceptIGNORE NULLS.random()can generate dates and timestamps within a range. - Performance. The announcement says improved foreign-key checks made bulk inserts "nearly 2x faster" in testing, the only performance figure it gives. Other changes: the planner now turns
NOT INinto an anti-join when no NULLs are present, it can aggregate before joins, and asynchronous I/O workers now scale automatically betweenio_min_workersandio_max_workers. - Security. Server-side SNI lets one server present different TLS certificates per hostname, configured in the new
pg_hosts.conf. Successful md5 logins now produce a warning, ahead of md5's planned removal.
Some features expected earlier in the cycle didn't make it. Databricks' release write-up notes that SQL/PGQ property graphs, GROUP BY ALL, and partition MERGE/SPLIT were reverted between June and September 2026. Don't plan around them for this version.
Breaking changes that can stop or skew an upgrade
The release notes list a long set of incompatibilities. They fall into three groups.
Changes that block the upgrade
| Change | What happens | What to do |
|---|---|---|
inet/cidr default opclass moves from btree_gist to core GiST |
pg_upgrade refuses clusters with btree_gist inet/cidr indexes, because those indexes can exclude rows that should be returned |
Drop them and recreate them with the core GiST opclass before upgrading |
| CR/LF characters in database, role or tablespace names | pg_upgrade rejects the cluster |
Rename those objects |
MULE_INTERNAL encoding removed |
Databases using it can't be carried over | Dump and restore into a different encoding |
standard_conforming_strings always on |
Dumps made with a pre-19 pg_dump while this setting was off won't load |
Re-dump with a version 19 client |
| RADIUS authentication removed | pg_hba.conf entries using radius stop working |
Move to another authentication method first |
Changes that silently alter behavior
These won't stop the upgrade, but they can change performance or query results afterwards:
- JIT is off by default. The project says the cost estimates that triggered JIT "proved unreliable." If you run large analytical queries, benchmark with JIT on and off. Re-enable it with
ALTER SYSTEM SET jit = on;followed bySELECT pg_reload_conf();. - The
max_locks_per_transactiondefault rises from 64 to 128, and lock allocation works differently. According to the release notes, a custom value has to be roughly doubled to keep the same capacity. json_array()returns[]instead ofNULLwhen there are no rows. Check any application code that tests forNULL.postgres_fdwpasses onREAD ONLYandDEFERRABLE, so a read-only transaction can no longer write through a foreign table.- Parallel vacuum is throttled more. Leader and workers now share one cost-based delay limit, so a throttled parallel vacuum may run slower.
- TOAST compression defaults to
lz4(when the server is built with LZ4), andlog_lock_waitsis on by default. Expect more lock-wait lines in your logs.
Changes that require reindexing
Full-text search and pg_trgm now classify characters using the database's locale provider. If you use ICU or the builtin provider with LC_CTYPE = C, results can differ. The release notes say to reindex the affected indexes and recompute stored tsvector columns. Indexes built on fuzzystrmatch functions such as soundex() and dmetaphone() need reindexing too. Several text-search stemmers, including Dutch, also produce different stems now.
Renamed settings and monitoring
io_workers is gone. It's replaced by io_min_workers, io_max_workers, io_worker_idle_timeout and io_worker_launch_interval. The BUFFERPIN wait event type is now BUFFER. In pg_stat_subscription_stats, sync_error_count is now sync_table_error_count. Update dashboards and alert rules, or they'll fail without any error. log_autovacuum_min_duration now covers vacuum only; use the new log_autoanalyze_min_duration for analyze logging.
Pre-upgrade checklist with queries
Run these against a production-like copy of your current cluster before you book a maintenance window.
- Find btree_gist
inet/cidrindexes (btree_gist's opclasses are namedgist_inet_opsandgist_cidr_ops):
Run it in every database.SELECT c.relname AS index_name, opc.opcname FROM pg_index i JOIN pg_class c ON c.oid = i.indexrelid JOIN pg_opclass opc ON opc.oid = ANY (i.indclass::oid[]) WHERE opc.opcname IN ('gist_inet_ops', 'gist_cidr_ops'); - Find names containing carriage returns or line feeds:
SELECT 'database' AS kind, datname AS name FROM pg_database WHERE datname ~ E'[\\r\\n]' UNION ALL SELECT 'role', rolname FROM pg_roles WHERE rolname ~ E'[\\r\\n]' UNION ALL SELECT 'tablespace', spcname FROM pg_tablespace WHERE spcname ~ E'[\\r\\n]'; - Check encodings:
SELECT datname FROM pg_database WHERE pg_encoding_to_char(encoding) = 'MULE_INTERNAL'; - Check authentication: search
pg_hba.confforradius. Move remaining md5 users to SCRAM now, before the warnings start filling client logs. - Check extensions. List them with
SELECT extname, extversion FROM pg_extension;and confirm each one has a PostgreSQL 19 build. Databricks warns that not every PostgreSQL 18 extension is supported on 19 yet. - Do a dry run. Run
pg_upgrade --checkagainst a copy, then run your slowest and most important queries on the upgraded copy. Compare plans and timings, especially if you relied on JIT. - Plan post-upgrade work. Schedule reindexing for text-search, trigram and fuzzystrmatch indexes. Update monitoring for the renamed settings and wait events. Check that rollback works and that backups actually restore.
The release notes accept three migration paths: pg_dumpall and restore, pg_upgrade, or logical replication. Logical replication gives the shortest cutover. Databricks points out that sequences aren't streamed continuously, though, so refresh them right before you switch traffic.
When to upgrade: timing by situation
New PostgreSQL major versions are supported for five years. The versioning policy page lists version 19 as supported until November 19, 2031. Versions 15 to 18 are still supported, and version 14 has dropped off the list. As the Netics Labs write-up puts it, a cluster on 14 "stops receiving fixes" as of November 12, 2026.
| Your situation | Suggested move |
|---|---|
| Self-hosted on PostgreSQL 14 | Upgrade now. Consider going to 18 to avoid 19's new incompatibilities, or to 19 if you want REPACK and plan advice |
| Amazon RDS on PostgreSQL 14 | Standard support ends February 28, 2027. Upgrade before then to avoid automatic Extended Support charges |
| On 15–18, stable | No rush. Test 19 in staging and plan production for 19.1 or later |
| Fighting table bloat or plan regressions | 19 is worth testing early for REPACK (CONCURRENTLY) and pg_stash_advice |
| Starting a new project | Use 19 if your host and extensions support it |
Managed platforms
Managed providers each follow their own schedule:
- Amazon RDS. The RDS release calendar says new major versions arrive "within 30 days of the community release for first minor of the new major version", which means after 19.1. Databricks expects 19.1 in February. The same calendar lists the end of RDS standard support for PostgreSQL 14 as February 28, 2027, with year-one Extended Support pricing starting March 1, 2027. RDS automatically enrolls instances that haven't upgraded.
- Databricks Lakebase made PostgreSQL 19 available on release day, November 19, 2026, according to Databricks. You choose it when creating a new project.
- Other providers (Neon, Supabase, Google Cloud SQL, Azure). Check each provider's own changelog. The Hacker News discussion mentions Neon availability, but I didn't confirm it from a primary source.
REPACK in practice: what it replaces and its limits
Most teams will feel REPACK first. VACUUM FULL and CLUSTER take an exclusive lock for the whole run. That's why so many teams installed pg_repack or pg_squeeze. Now that the concurrent version is built in, there's one less extension to keep compatible across major upgrades.
It has limits. Databricks lists them:
- It needs spare replication-slot capacity (the new
max_repack_replication_slotssetting) and temporary disk space for the new copy of the table. - The upstream implementation fails if more than about 104 million tuples are updated or deleted while it runs.
- It isn't MVCC-safe, and long-running transactions can delay the final lock it takes.
Run it during quieter periods. Watch it with the new pg_stat_progress_repack view, and don't remove pg_repack until you've tested REPACK on your largest tables.
Plan advice has a similar caveat. It's for locking in a plan you've already diagnosed, not a general tuning tool. Databricks also notes that a "matched" result only means the plan followed your advice, not that it was faster.
What to do now
- On PostgreSQL 14: community support has ended. Pick a target version this week. RDS users have until February 28, 2027 before Extended Support charges begin.
- On 15–18: run the checklist above against a copy of production now. That way, the btree_gist indexes, renamed settings and reindexing work are known well before 19.1 arrives, likely in February.
- Building something new: start on 19 where your host offers it.
ON CONFLICT DO SELECT, JSONCOPYandWAIT FOR LSNremove workarounds that many app backends carry today.
If you're comparing hosted Postgres, database tooling or AI data assistants as part of the move, the AI data tools and AI coding assistants categories are a good place to start.
Frequently asked questions
When was PostgreSQL 19 released?
The PostgreSQL Global Development Group released PostgreSQL 19 on November 19, 2026. According to the official versioning policy, it will be supported until November 19, 2031. Many teams wait for the first minor release, 19.1, before upgrading production.
Can I use pg_upgrade to go straight to PostgreSQL 19?
Yes. The release notes list three supported paths: pg_dumpall with a restore, pg_upgrade, or logical replication. pg_upgrade will refuse a cluster that has btree_gist indexes on inet or cidr columns, or database, role or tablespace names that contain carriage returns or line feeds. Fix those first, then run pg_upgrade --check on a copy.
Why are my analytical queries slower after upgrading to PostgreSQL 19?
PostgreSQL 19 turns JIT compilation off by default because the cost estimates that triggered it proved unreliable. If your workload runs many large analytical queries, compare timings with jit on and off. If JIT helps, turn it back on with ALTER SYSTEM SET jit = on and reload the configuration.
Is PostgreSQL 14 still supported?
The community stopped supporting PostgreSQL 14 on November 12, 2026, and it will get no more community fixes. On Amazon RDS, standard support for version 14 runs until February 28, 2027. After that, RDS automatically enrolls instances that haven't upgraded in paid Extended Support.
What does REPACK do in PostgreSQL 19?
REPACK reclaims space from bloated tables and can reorder a table to match an index. It replaces VACUUM FULL and CLUSTER, which still work for compatibility. REPACK (CONCURRENTLY) does the job without blocking reads or writes, but it needs spare replication slots and temporary disk space.
Sources
- PostgreSQL 19.0 Release Notes— postgresql.org
- PostgreSQL 19 Released! (official announcement)— postgresql.org
- PostgreSQL Versioning Policy and Supported Versions— postgresql.org
- Databricks: PostgreSQL 19 is available on Lakebase— databricks.com
- AWS: Release calendars for Amazon RDS for PostgreSQL— docs.aws.amazon.com
- DEV Community: PostgreSQL 19 Ships on 19 November and PostgreSQL 14 Leaves Support— dev.to
