The PostgreSQL Global Development Group released PostgreSQL 19 on November 19, 2026. The new major version of the open source database is aimed above all at operators of large, mission-critical installations: a new command called REPACK cleans up bloated tables without blocking ongoing work, and a new extension is designed to keep query plans stable. Developers get, among other things, “read-your-writes” on standby servers and a “get or create” pattern without an extra network round trip. Anyone planning to upgrade should read the list of incompatibilities, though, because several defaults change noticeably.
PostgreSQL 19: REPACK replaces VACUUM FULL and CLUSTER
Tables with many updates and deletes accumulate unused space, known in Postgres terms as “bloat”. Until now, VACUUM FULL and CLUSTER were the tools against it, but both locked the table for the duration of the cleanup. According to the official announcement, PostgreSQL 19 introduces the REPACK command for this job. It reclaims the space and can also reorder a table’s data to match an index. With REPACK (CONCURRENTLY), it does so without blocking reads or writes.
REPACK replaces the two old commands, which remain available for compatibility. Progress is shown in the new system view pg_stat_progress_repack, which according to the release notes also covers CLUSTER and VACUUM FULL. Many teams have so far relied on external extensions such as pg_repack for online cleanup. Having the feature in core should spare operators of write-heavy tables a fair number of maintenance windows.
Plan advice: pinning query plans without touching the application
The PostgreSQL query planner picks an execution plan for every query. When the data changes, the plan can change too, and sometimes the new one is slower. For those cases, version 19 ships the pg_plan_advice extension. It allows plan choices to be stabilized and lets users experiment with different plans through a “plan advice” mini-language. The companion extension pg_stash_advice stores advice and applies it automatically based on the query identifier. A slow plan can therefore be fixed without changing application code.
That is a notable step, because the project long resisted classic optimizer hints. Implementing the feature as an extension leaves default behavior untouched: if you do not load it, nothing changes.
Autovacuum and monitoring
On busy systems, autovacuum often struggles to keep up. PostgreSQL 19 can now vacuum indexes with parallel workers, controlled by the new autovacuum_max_parallel_workers setting. A new scoring system decides which tables are processed first, and the pg_stat_autovacuum_scores view makes that prioritization visible. Queries that read a table can now mark pages as all-visible in the visibility map during execution, reducing the work future vacuums have to do.
Monitoring gains more views as well: pg_stat_lock provides per-lock-type statistics and pg_stat_recovery shows recovery status. A new IO option for EXPLAIN ANALYZE and the matching auto_explain setting report asynchronous I/O activity. PostgreSQL 19 also reports the amount of WAL written for full-page images in pg_stat_wal, in EXPLAIN (ANALYZE, WAL) and in vacuum and analyze log events.
Replication: logical replication without a restart
Previously, enabling logical replication often required a server restart, because it needed wal_level = logical. PostgreSQL 19 automatically raises the effective WAL level to logical as soon as the first logical replication slot is created. The new effective_wal_level setting shows which level is in use.
Several long-requested features join it:
- Sequence replication: sequence values can be replicated logically via
CREATE PUBLICATION ... FOR ALL SEQUENCES. - EXCEPT clause: individual tables can be excluded when publishing
FOR ALL TABLES. - Connections via foreign servers:
CREATE SUBSCRIPTIONcan take its connection settings from a foreign server, such as one defined with postgres_fdw, using the SERVER clause. - Conflict detection: the
retain_dead_tuplesoption detects updates that conflict with a concurrent delete; these are counted inpg_stat_subscription_stats. - Read-your-writes: with the new
WAIT FOR LSNcommand, a session on a standby waits until the standby has replayed the WAL up to a specific log sequence number. Applications that send reads to replicas can thus read data they just wrote on the primary.
Queries on foreign tables benefit too: with the new import_stats option, postgres_fdw can import table statistics from the remote server during ANALYZE instead of sampling rows.
What is new for developers
For application code, INSERT ... ON CONFLICT DO SELECT is probably the most practical addition. The RETURNING clause can now return existing rows that conflict with an insert, so the classic “get or create” pattern no longer needs a second round trip. Other additions listed in the announcement:
COPYcan output data as JSON, including as a single JSON array.- The SQL/JSON path language gains string methods such as
lower(),upper(),replace(),split_part()andbtrim(). - The window functions
lead(),lag(),first_value(),last_value()andnth_value()can skip NULL values withIGNORE NULLS. random()can now generate random dates and timestamps within a range.
Many workloads should also run faster without any application changes. Foreign key checks were sped up, which in the project’s testing made bulk inserts nearly twice as fast. The planner converts NOT IN clauses into more efficient anti-joins when no NULL values are present and can perform aggregation before joins (“eager aggregation”). The asynchronous I/O subsystem introduced in version 18 now automatically adjusts the number of I/O workers between io_min_workers and io_max_workers; the old io_workers setting is gone. LISTEN and NOTIFY scale better in multi-channel workloads.
Security: SNI, MD5 warnings, RADIUS removed
PostgreSQL 19 adds server-side support for Server Name Indication. A single server can present a different TLS certificate depending on the hostname a client connects to, configured in the new pg_hosts.conf file. After a successful MD5 password authentication, the server now issues a warning, which can be turned off with md5_password_warnings. MD5 was deprecated in version 18 and is set to be removed in a future release; the project recommends SCRAM. The new password_expiration_warning_threshold setting warns users seven days before a password expires by default. RADIUS authentication has been removed; according to the release notes, PostgreSQL only supported RADIUS over UDP, which is “unfixably insecure”.
These changes can hurt during an upgrade
The release notes list a whole series of incompatibilities. The most important:
- JIT off by default: the cost estimates PostgreSQL used to decide on just-in-time compilation proved unreliable. Sites running many large analytical queries should re-enable JIT deliberately.
- max_locks_per_transaction: the default rises from 64 to 128. Because lock allocation has changed, custom settings effectively need to be doubled to match their previous capacity.
- standard_conforming_strings always on: dumps created with older pg_dump versions and
standard_conforming_strings = offwill not load properly. Create dumps with the version 19 tools. - pg_upgrade blocks: clusters with btree_gist indexes on inet/cidr or with line breaks in database, role or tablespace names cannot be upgraded via pg_upgrade.
- Full text search and pg_trgm: with ICU or builtin locale providers, results can change; a reindex of affected indexes after the upgrade is recommended.
- Other points:
json_array()returning no rows now yields an empty array instead of NULL, parallel vacuum workers share the cost limit and may run slower, and the MULE_INTERNAL encoding is removed.
LZ4 is also the new default TOAST compression method when PostgreSQL is built with LZ4 support. As with every major version, migrating requires a dump/restore, pg_upgrade or logical replication.
Support until 2031, PostgreSQL 14 reaches its end
According to the versioning policy, PostgreSQL 19 will receive updates until November 19, 2031. At the same time, PostgreSQL 14 has been out of support since November 12, 2026. Anyone still running version 14 no longer gets security fixes, and there have been plenty of those lately: in mid-November the project closed 52 flaws, as we reported in our article on CVE-2026-6460 and the PostgreSQL updates. Postgres-compatible cloud services are not immune either, as the critical Babelfish flaw in Aurora showed.
Assessment
PostgreSQL 19 is not a release of big headlines but one for day-to-day operations. REPACK (CONCURRENTLY) in core, plan advice, logical replication without a restart and read-your-writes on replicas solve problems that teams previously handled with extensions, workarounds or maintenance windows. The flip side is a set of changed defaults, above all JIT being switched off and the new math for max_locks_per_transaction. Teams that go through the incompatibilities and test on a copy before upgrading get a noticeably easier database to maintain. Anyone still on version 14 has no reason left to wait.




