· via Hacker News – Front Page (native)
PostgreSQL 19 reaches general availability with same-day rollout on Neon and Lakebase
PostgreSQL 19 is generally available, and managed platforms Neon and Databricks Lakebase shipped it the same day. The release targets table bloat, unstable query plans and long-standing logical replication gaps.

PostgreSQL 19 reaches general availability
PostgreSQL 19 is now generally available, and it landed on managed platforms the same day rather than weeks later. According to a Databricks blog post, the new major version is already selectable when creating a Lakebase project, an unusually fast turnaround given that major PostgreSQL releases typically take weeks to reach managed services. A Hacker News post announcing the release notes that Neon has it available as well.
Databricks credits the same-day delivery to engineers on the Lakebase team who are PostgreSQL committers and contributors and who worked on the release directly, across query planning, shared memory, locale handling and release management.
The company describes version 19 as pragmatic rather than headline-driven: no single feature changes what PostgreSQL is. Instead, the release goes after problems database teams know well — bloated tables, unstable query plans, replica lag, incomplete statistics and maintenance that is risky to run in production.
REPACK reclaims space without blocking traffic
Because PostgreSQL uses multiversion concurrency control, an UPDATE writes a new row version and a DELETE leaves a dead tuple behind. Vacuuming makes that space reusable, but it does not necessarily return space to the operating system or restore physical row order, so heavily updated tables tend to bloat. The built-in fixes, VACUUM FULL and CLUSTER, hold an ACCESS EXCLUSIVE lock for the whole operation; the lock-free alternative so far has been external tooling such as pg_repack.
Version 19 adds a native REPACK command. Per Databricks, with CONCURRENTLY it rebuilds a table while normal reads and writes continue, recording ongoing changes via logical decoding and taking only a brief exclusive lock at the end to swap the rebuilt files into place. Used with an index, it also physically reorders rows the way CLUSTER does.
The constraints are real, though. A concurrent repack needs a free replication slot, temporary disk space for the table plus all its indexes, and a primary key or suitable replica identity. The upstream implementation fails if more than roughly 104 million tuples are updated or deleted in the target table while the repack runs. Databricks also warns that concurrent repacking is not MVCC-safe: transactions using snapshots taken before the repack began can see the table as empty afterward, and physical clustering degrades as new writes arrive.
Plan advice for bad estimates
The cost-based optimizer relies on per-table statistics, which cannot capture cross-table correlations. The blog's example: enterprise customers who make up 0.1 percent of customers but generate 30 percent of orders can push the planner into repeated index probes where a hash join belongs.
PostgreSQL 19 introduces pg_plan_advice and pg_stash_advice. The first explains a plan as a set of advice items, applies advice you supply, and reports whether the planner followed it. The second stores advice keyed by query ID and re-applies it automatically when a matching query is planned. Unlike inline optimizer hints, the advice is kept separate from the queries applications send, so it can be inspected and tested before being stashed. Databricks frames it as a way to stabilize a plan after diagnosis, not a replacement for the optimizer: a compliant plan is not necessarily a better one.
Parallel autovacuum and more visible statistics
Autovacuum can now use parallel workers when vacuuming indexes, controlled by autovacuum_max_parallel_workers, and a new scoring system decides which tables need attention most urgently. New views expose lock statistics (pg_stat_lock), recovery status (pg_stat_recovery) and autovacuum priorities (pg_stat_autovacuum_scores). On the diagnostics side, EXPLAIN (ANALYZE, IO) reports asynchronous I/O activity, auto_explain.log_io adds the same to logged plans, and byte counts for full-page images now appear in pg_stat_wal, EXPLAIN (ANALYZE, WAL) and vacuum logs.
Logical replication closes long-standing gaps
Several old limitations in logical replication are gone. Logical decoding can be activated without a restart when wal_level is replica; creating the first valid logical slot raises an effective_wal_level of logical. Sequence values can be synchronized from publishers to subscribers, publications support FOR ALL SEQUENCES and FOR ALL TABLES EXCEPT with a named exclusion list, subscriptions can reuse connection details from a postgres_fdw server, and a retain_dead_tuples option preserves enough information to detect update_deleted conflicts.
Why it matters
Same-day availability on Neon and Lakebase shows how competitive managed PostgreSQL has become, and how entangled platform vendors now are with the core project — Databricks employs committers who shaped this release. For teams running Postgres in production, the substance matters more: REPACK brings online bloat cleanup into core instead of extensions, plan advice offers a supported escape hatch when the optimizer picks a bad plan, and the logical replication changes remove friction that has complicated change-data-capture pipelines and migrations for years. Individually these are incremental; together they cut the operational cost of running PostgreSQL at scale.
- #postgresql
- #databases
- #cloud
- #open-source