What’s new in this version?
PostgreSQL 19 adds a built-in REPACK command for online table reorganization, two planner-advice extensions (pg_plan_advice and pg_stash_advice) to stabilize query plans, parallel index vacuuming, and read-your-writes consistency on replicas via a new WAIT FOR LSN command.
The REPACK command replaces VACUUM FULL and CLUSTER without the downtime those tools required, and the planner-advice extensions let teams fix slow queries without changing application code.
4 details4 sources1 disagreement
- REPACK (CONCURRENTLY) reclaims bloat and reorders table data to match an index while reads and writes continue; it replaces VACUUM FULL and CLUSTER, which remain for compatibility. It has limits: it needs a non-deferrable primary key or index-based replica identity, temporary disk space for a table copy, and fails if more than ~104 million tuples change during the operation [databricks.com](https://www.databricks.com/blog/postgresql-19-available-lakebase).
- pg_plan_advice lets you experiment with planner choices using a mini-language; pg_stash_advice stores advice and auto-applies it by query identifier, so you can fix a slow plan without touching the application [postgresql.org](https://www.postgresql.org/about/news/postgresql-19-released-3449/).
- Autovacuum can now vacuum indexes with parallel workers (autovacuum_max_parallel_workers) and uses a scoring system to prioritize tables; queries that read a table can mark pages all-visible during execution, cutting future vacuum work [postgresql.org](https://www.postgresql.org/about/news/postgresql-19-released-3449/).
- Logical replication no longer requires a restart — the effective WAL level rises to logical automatically — and adds sequence-value replication, an EXCEPT clause in CREATE PUBLICATION, and WAIT FOR LSN so standby sessions can wait for replay before reading [postgresql.org](https://www.postgresql.org/about/news/postgresql-19-released-3449/).
- Concurrent REPACK has real operational constraints — Databricks notes the ~104 million tuple change limit during the operation and the need for free replication-slot capacity, so the headline "no downtime" claim is narrower than it sounds for busy tables.
