DBMS Weekly — 2026-07-13 (week of 2026-07-13–2026-07-19)

Beta week: PostgreSQL 19 Beta 2 shipped Jul 16, and the price of late beta showed up two days later — Friday brought two reverts: GROUP BY ALL is out of PG19 entirely (wrong results with ORDER BY under non-default equality semantics), and the sequence-refresh concurrency guard was pulled pending a different fix. No browser was connected this run and most archive caches were stale; mailing-list coverage came via pgsql-committers (fresh) — see the source note at the bottom.

PostgreSQL

• PostgreSQL 19 Beta 2 released — second beta, GA still targeted around September/October. (PGDG · postgresql.org, Jul 16)
  • fixes since Beta 1: vacuumdb --analyze-in-stages regression on partitioned tables, several FOR PORTION OF and SQL/PGQ property-graph fixes, REPACK worker not cleaned up on FATAL exit, a race in logical-decoding activation, autovacuum's multixact-age score going infinite, postgres_fdw imported-statistics handling
  • reverted: non-text output formats for pg_dumpall
• GROUP BY ALL reverted from PG19 — post-commit review (reported by Chao Li) found the new syntax missed the special handling of entries that also appear in ORDER BY, returning wrong results when ORDER BY specifies non-default equality semantics; fixing it cleanly was judged too much churn for late beta, so the feature waits for v20. (Daniel Gustafsson, Tom Lane · postgres git, Jul 17) [reverted]
• Sequence-refresh concurrency guard reverted — the fix that made ALTER SUBSCRIPTION ... REFRESH SEQUENCES error out while a sequencesync worker was running (closing the stale-value race found in review two weeks ago) caused intermittent buildfarm failures and was judged too inconvenient to keep; the underlying race gets a different fix in a follow-up. (Amit Kapila · postgres git, Jul 17) [reverted]
• GiST index-only scans could return misaligned column data — fix for a column-alignment issue in GiST index-only scans, backpatched across the stable branches. (Peter Geoghegan · postgres git) [committed]
• Planner: edge case in remove_useless_result_rtes() with outer joins — wrong-plan edge case when removing useless RTE_RESULT rels under outer joins, fixed and backpatched; same week the grammar restored | and -> as prefix operators after a regression from earlier parser work. (Tom Lane · postgres git) [committed]
• Master grab-bag — pg_controldata now shows logical decoding status (Masahiko Sawada); REASSIGN OWNED fixed for subscriptions in other databases (Jeff Davis); uuidv7() rejects infinite/out-of-range interval shifts (Masahiko Sawada); case-insensitive LIKE prefix selectivity fixed (Jeff Davis); the ancient SQL function getpgusername() is gone (Michael Paquier); platform floors raised to Visual Studio 2022 and ICU 55 (Peter Eisentraut) — the matching Tcl 8.6 bump was reverted the same week. [committed]

Reading

• How to achieve pruning when querying by non-partitioned columns — if a non-key column correlates with the partition key (sequential IDs in time-partitioned data), per-partition CHECK constraints let the planner prune on that column too, with a BRIN-inspired trick for outliers. (Haki Benita · hakibenita.com) [publish date unverified]
• How SQL/PGQ rewrites to joins on PostgreSQL 19 — behind PG19's GRAPH_TABLE: how graph-query syntax compiles down to plain joins — timely, given the stream of SQL/PGQ fixes still landing in beta. (Hans-Jürgen Schönig · Cybertec) [vendor blog — substantive] [publish date unverified]

Ecosystem releases

• Odyssey 1.5.1 — Yandex's multi-threaded connection pooler; a large extended-protocol refactor headlines. (Yandex · postgresql.org news, Jul 14)
  • extended-protocol conformance fixes: correct Close handling, DEALLOCATE ALL, correct COPY; pipelining performance improved
  • new: simple shared pools, pool_pin_on_listen for transaction pooling, cpu_affinity, queue-time metrics (avg_wait_time/total_wait_time), NOTICE after configurable queue wait, server_drop_on_cached_plan_error
  • fixes: cancel-connection race, auth_query fixes, pool-size recheck after server creation
• plx: stored functions in PHP, JavaScript, Python, Ruby — even COBOL — transpiles other languages to PL/pgSQL rather than embedding interpreters; a curious alternative to the PL zoo. (Command Prompt · via Postgres Weekly #657) [publish date unverified]

CommitFest (open: PG20-1, #59 — In Progress Jul 1–31)

• Balance: not reliably measurable this week — every CommitFest page this run was served from stale caches (the /59/ status summary still showed the Jul 13 totals of 153 committed / 613 total; the per-CF activity log was still cut at Jun 24). The committers list shows review-month commit flow continued through Beta 2 week, including two feature reverts (GROUP BY ALL, sequence-refresh guard). Week-over-week deltas resume when a live read is next possible. (vs last week: n/a — stale cache)

Community pulse

• SQLite should have (Rust-style) editions — proposal to evolve SQLite's sharp edges (weak typing et al.) behind opt-in editions rather than breaking its legendary compatibility; the debate is a compact tour of why "never break anything" is both SQLite's moat and its curse. (Hacker News · 355 pts, 171 comments)

~14 items · sources scanned: pgsql-committers via mail-archive (thread index fresh through Jul 18; msg47258–47429 ≈ Stamp 19beta2 → week's end), postgresql.org news + events + release notes, Postgres Weekly #657, CommitFest app (all pages cache-stale this run), Hacker News via hntoplinks (pages 1–2 fresh, page 3 a months-old cache), publickey.jp [ja], blog.vonng.com [zh], Habr [ru] + Dalibo [fr] via native-language search (nothing DB-relevant in-window from any of the four) · filtered out as marketing/ads: ~6 · out-of-window: several (incl. malisper's "Four Horsemen" outage taxonomy and Pigsty 4.4, both ~Jul 11).

Source note: no browser was connected. pgsql-hackers was effectively unreachable — mail-archive's hackers index is frozen at May 18, its per-day search is blocked to cold fetches this run, postgresql.org /list/… and /since/… browse pages render empty to plain fetch, and the postgrespro.com mirror's month index stops at May — so list coverage came from pgsql-committers commit messages (fresh) rather than thread reading; a hackers-side design debate without a commit this week may be missing. pgsql-bugs/-performance remain without a fetchable archive. The CommitFest app served stale caches end-to-end (totals frozen at the Jul 13 snapshot). arXiv listings could not be date-verified, so the research section is omitted rather than guessed. No CFPs opened in-window; no event enters the 30-day horizon (Hyderabad Postgres Days, Aug 20–21, enters next week).