• POSETTE: An Event for Postgres 2026 is a wrap — all 44 talks on YouTube — the free virtual event (Jun 16–18, organised by the Postgres team at Microsoft) ran four livestreams; every talk is now un-gated on YouTube. The internals-relevant picks for a hacker:
  • Postgres 19 Hackers Panel: What's In, What's Out, & What's Next — Álvaro Herrera, Heikki Linnakangas, Melanie Plageman, and Thomas Munro on what made the PG19 feature freeze and what stalled. (keynote)
  • random_page_cost — why the default is 4.0 and should you lower it? — Tomas Vondra dissects the planner's random-vs-sequential I/O cost ratio and when the SSD-era default misleads.
  • Fuzzing PostgreSQL — Adam Wolk applies fuzzing to libpq, the client library that parses every server message; a template for finding edge-case bugs in stateful C code.
  • The Wonderful World of WAL — Bruce Momjian walks the write-ahead log as the substrate for crash recovery, PITR, and physical/logical replication.
  • pg_stats: How Postgres Internal Stats Work — Richard Yen on the statistics model behind seq-scan-vs-index-scan planner decisions.
  • Journey of developing a performance optimization feature in PostgreSQL — Rahila Syed on systematically diagnosing and fixing an internals bottleneck end-to-end.
  • Past, Present, and Future: Logical Decoding and Replication — Hari Kiran traces the arc from the low-level decoding API to today's logical replication.
  • Exploring property graphs with SQL/PGQ in PostgreSQL — Ashutosh Bapat on the in-core graph-query syntax landing in PG19.
• pg_kpart 1.0 — new extension that rejects any query against a protected partitioned table that can't prune a single partition, turning "always filter on the partition key" from convention into a database-enforced guarantee. (Gilles Darold · HexaCluster)
  • throws a clear error (with a dedicated SQLSTATE) instead of letting a key-less query trigger a server-wide full-hierarchy scan
  • audit mode for staged rollout; blacklist/whitelist scoping to target specific tables
• powa-archivist 5.1.2 — the core collector of the PostgreSQL Workload Analyzer project ships a maintenance release. (PoWA team · Marc Cousin)
  • new: per-module retention (set different history-retention windows per collected extension)
  • compatible with all supported PostgreSQL versions

PostgreSQL mailing lists

• [hackers] Conflict log history table for Logical Replication — a lost-conflict bug surfaces — the PG20-targeted feature that lets a subscription log apply-time conflicts to a system-managed table (in a pg_conflict schema) instead of just the server log saw heavy review this week. vignesh C reported a real defect: a concurrent ALTER SUBSCRIPTION that switches the conflict-log destination drops the log table while the apply worker still holds its cached OID, so table_open() fails, the worker exits (code 1), and the conflict record is logged nowhere; reviewers also flagged that arbitrary objects (extensions, collations, operators) can be created in pg_conflict and that DROP of the log table errors inconsistently with INSERT/UPDATE. (vignesh C, Peter Smith, Shlok Kyal, Amit Kapila · pgsql-hackers) [patch posted]

CommitFest (open: PG20-1, #59)

• Balance (Jun 15–21): PG20-1 is in its collection phase (in-progress Jul 1–31) and saw net inflow — several new patch records, a few committed/withdrawn. This is the first week the digest tracks the CommitFest, so the week-over-week delta baseline starts now; absolute totals will be diffed from next run. (source: per-CF activity log)
• New this week: Global temporary tables (deanr) — a fresh patch record for the long-requested GTT feature: tables whose definition is shared cluster-wide but whose rows and statistics stay private to each session (the SQL-standard / Oracle-style global temp table Postgres still lacks). Also new: PROXY protocol support (#6916 — lets the backend see the real client address behind a load balancer) and pg_dump: skip pg_init_privs entries for non-existent roles (#6915, Hüseyin Demir).
• Closed in-window: Fix PL/Python crash when a SRF is redefined mid-execution (#6917, filed and committed same day by Tom Lane) and Fix publisher-side sequence permission reporting (#6911, committed); pg_restore --no-globals output for pg_dumpall archives (#6899) withdrawn.

Wider DBMS & distributed data

• Why we built Azure HorizonDB for PostgreSQL — Dingding Lu (POSETTE) on Microsoft's new cloud-native, Postgres-compatible engine that decouples storage to keep latency predictable under spiky traffic and scale reads without external bolt-ons; a useful look at how a major vendor is re-architecting Postgres for the cloud. (Microsoft · posetteconf.com)
• pg_lake: Postgres as a lakehouse — Marco Slot (POSETTE) on the extension approach to making Postgres a competent analytics/lakehouse front-end (Iceberg/object-storage interaction) rather than bolting on a separate warehouse. (posetteconf.com)

Migration experience

• Migrating VLDBs from Oracle to Azure Database for PostgreSQL — Adithya Kumaranchath (POSETTE) shares field-tested strategies for moving multi-terabyte Oracle databases onto a managed Postgres service, where the cloud target adds real constraints over a self-managed migration. (posetteconf.com)

International (non-English sources)

• PostgreSQL 19 ushers in a new era of temporal-table design (Part 1) — converts pgbench's standard schema to a temporal design using tsrange valid_at columns, PRIMARY KEY … WITHOUT OVERLAPS, btree_gist, and PG19's UPDATE … FOR PORTION OF; EXPLAIN ANALYZE shows point-in-time lookups stay cheap (Index Scan / Bitmap Index Scan on a single GiST index). Argues Postgres is now uniquely suited to temporal tables — the only major engine combining a range-overlap index with WITHOUT OVERLAPS primary keys — versus SQL Server's split current/history-table approach. (harukat · SRA OSS · qiita.com) [ja] (orig: 「PostgreSQL19でテンポラルテーブル設計の新時代へ(1)」)

New sources added this week

• pghackers.com — PostgreSQL Hackers Explorer — AI-assisted search/exploration over the pgsql-hackers archive; potentially a faster way to triage in-window threads than the raw monthly index. (surfaced via search; evaluate over coming weeks)

11 items · sources scanned: ~31 (Planet PostgreSQL, PostgreSQL news archive, pgsql-hackers via mail-archive.com, CommitFest PG20-1 activity log, pganalyze, thebuild.com, POSETTE, Qiita [ja], Habr [ru], modb.pro [zh], arXiv cs.DB) · international: 1 (ja) · filtered out as marketing/ads: ~6 (incl. recycled Crunchy↔Snowflake from 2025, vendor product talks). Community pulse, Research, and a ru/zh International item omitted: nothing could be confirmed inside the Jun 15–21 window without fabricating links or engagement counts.

• How the Other Half Counts — Christophe Pettus surveys how Oracle, Db2, MySQL, SQLite, DuckDB, and Snowflake gather planner statistics vs. PostgreSQL's sampling ANALYZE; the real axis turns out to be stats freshness and who owns it, not the join algorithm. (thebuild.com)
• All Your GUCs in a Row: deadlock_timeout — clarifies that deadlock_timeout sets how often PostgreSQL checks for a deadlock, not how long a deadlock is tolerated; part of a near-daily GUC-internals series. (Christophe Pettus · thebuild.com)
• All Your GUCs in a Row: DateStyle — on why DateStyle conflates two unrelated behaviours (input parsing and output formatting) in a single GUC. (Christophe Pettus · thebuild.com)
• All Your GUCs in a Row: data_sync_retry — the GUC born from the Postgres fsync-reliability durability saga, and why the default is what it is. (Christophe Pettus · thebuild.com)
• My Three Top PostgreSQL 19 Features — Stefanie Janine walks through INSERT ... ON CONFLICT DO SELECT, the new in-core REPACK (with CONCURRENTLY), and SQL/PGQ graph queries landing in PG19. (proopensource.it)
• PostgreSQL Anonymizer 3.1: local differential privacy — Dalibo's data-masking extension adds local-differential-privacy masking techniques (announcement light on detail; verify changelog before relying on it). (Dalibo · postgresql.org)

PostgreSQL mailing lists

• [hackers] FK fast-path: OOB write + missed FK checks during batched RI checks — three defects in PG19's new batched foreign-key fast path (an out-of-bounds write and two paths silently committing FK-violating orphan rows); reproducers confirmed, the fix is pushed and on the PG19 Open Items list. Data-integrity regression in a just-shipped beta feature. (Amit Langote · pgsql-hackers) [committed]
• [hackers] Commit Sequence Numbers and Visibility (CSN snapshots) — Jeff Davis proposes concrete rules for CSN-based snapshot acquisition; Matthias van de Meent pushes back that coupling snapshot acquisition to durability would impose synchronous-flush waits on readers, counter-proposing separate durability-horizon WAL records. Core architectural debate for a future MVCC redesign. (Jeff Davis · pgsql-hackers) [open]
• [hackers] RFC: a constraint-enforcement layer for cross-partition uniqueness — Greg Burd reframes the perennial "global unique index on partitioned tables" problem as a separate pg_constraint_method abstraction (a lightweight WAL-logged membership map) instead of a full global index, citing prior global-index attempts. (Greg Burd · pgsql-hackers) [open]
• [hackers] Any JIT-compiled query crashes (SIGILL) on libLLVM 19 + ASAN CI build — cfbot failures blamed on the Row Pattern Recognition patch turn out to be a feature-independent crash: on the ASAN+LLVM 19.1 config any JIT-compiled query SIGILLs on plain master; minimal repro and regression test posted. Was masking real CI signal on in-flight patches. (pgsql-hackers) [patch posted]

Community pulse

• PgDog is funded and coming to a database near you — a funding post for a Rust connection-pooler/sharder for Postgres became the week's biggest DB thread; the substance is the 261-comment debate over whether sharding belongs in a proxy and how it compares to PgBouncer / PgCat / Citus. (Hacker News · 548 pts, 261 comments)
• Show HN: HelixDB — a graph database built on object storage — an OLTP graph DB on S3 with native vector + FTS drew a long back-and-forth on the latency trade-offs of running transactions against cold object storage. (Hacker News · 158 pts, 46 comments)
• Spanish traders set the standard for GnuCash database design — an essay on how centuries-old double-entry bookkeeping prefigured GnuCash's schema turned into an 82-comment thread on modelling money in a relational database. (Hacker News · 133 pts, 82 comments)

Wider DBMS & distributed data

• The only scalable delete in Postgres is DROP TABLE — argues that under MVCC, bulk DELETE doesn't cheaply reclaim space (dead tuples, VACUUM, bloat), so partition-drop / DROP TABLE is the only delete that scales; lively HN discussion. (planetscale.com)
• Our MongoDB TLA+ workshop — Murat Demirbas reports an order-of-magnitude better onboarding when LLM assistance removes TLA+ syntax as the barrier, letting engineers formally model real replication/storage subsystems hands-on; includes a distilled "AI-assisted TLA+ best practices" guide. (muratbuffalo.blogspot.com)
• Postgres by Example — example-driven Postgres reference repo that surfaced on HN; useful as a teaching/cheat-sheet resource. (github.com)

Commercial engines (SQL Server, Oracle, MySQL, …)

• RUM: storing more in the index — Franck Pachot deep-dive on Postgres RUM indexes (carrying extra payload in the index to serve ESR-style equality/sort/range queries the way other engines' covering indexes do). (dev.to)
• GIN: understanding PostgreSQL's inverted index and its limitations — Pachot walks the internals and trade-offs of GIN indexing, with cross-engine framing. (dev.to)

International (non-English sources)

• Postgres Pro evolves a "Virtual DBA" LLM agent's environment with record/replay — Postgres Professional ran an evolutionary search over a Postgres-diagnosis agent's setup (prompts, analysis steps, MCP tools) by recording and fast-replaying real workloads; a headline finding is that handing the agent more tools can actually worsen its final verdict. (Postgres Professional · habr.com) [ru] (orig: «Промпты, RAG, LLM-тюнинг, Harness… Идём дальше?»)
• Three discovery problems with Postgres master/replica — and a lightweight fix — finding the master after failover, picking a replica within an acceptable lag, and guaranteeing read-your-writes, solved with a small HTTP service (pg-status) instead of DNS/HAProxy hacks that don't understand replication LSN. (krylosov-aa · habr.com) [ru] (orig: «Три задачи discovery при работе с PostgreSQL master/replica — и как их решить»)
• OceanBase lands a batch of papers at SIGMOD/VLDB 2026 — surfaced via the Chinese DB community 墨天轮; the Ant Group distributed-RDBMS vendor's top-venue haul, with three worth mining for our own engine: TaCo (query-aware subspace collision for high-dimensional ANN search), ESTune (Bayesian uncertainty-guided early stopping for knob tuning), and Secure Multi-Party Sampling over Joins (VLDB). (墨天轮 / modb.pro) [zh] (orig: «OceanBase 多篇论文被国际顶会录用»)

Research & cutting edge

• Multiversion Concurrency Control for Multiversion B-Trees — cMVBT redesigns the multiversion B-tree with optimistic latches for writers and latch-free range scans plus continuous GC, beating version-chain MVCC on write throughput and arbitrary-version range scans. (arXiv) [paper]
• Bespoke-Card: synthesizing workload-specific cardinality estimators — generate a per-workload cardinality estimator instead of tuning one generic model. (AIDB @ VLDB '26) [paper]
• Determination Provenance: From Ambiguity to Algebra — Joseph M. Hellerstein gives an algebraic treatment of determination/provenance. (arXiv) [paper]

23 items · sources scanned: ~30 · community threads scanned: HN (40 DB stories) · international: 3 (ru, zh) · filtered out as marketing/ads: 8

• PostgreSQL 19 Beta 1 released — first beta of PG19 lands (June 4) with in-core REPACK [CONCURRENTLY], the pg_plan_advice/pg_stash_advice planner-stabilization extensions, parallel autovacuum (autovacuum_max_parallel_workers) plus a new vacuum-prioritization heuristic, and adaptive async-I/O worker scaling (io_min_workers/io_max_workers) building on PG18. (Jonathan Katz · postgresql.org)
• Looking forward to Postgres 19: query hints, sort of — a tour of PG19's new pg_plan_advice / pg_stash_advice contrib modules that capture and reapply planner decisions by query id, ending the project's long "no hints" posture without actually being hints. (Shaun Thomas · pgEdge)
• pg_durable: durable execution as a Postgres extension — Microsoft open-sources a pgrx/Rust extension that runs Temporal-style durable functions entirely inside Postgres: state checkpointed to tables, crash/failover recovery via deterministic replay, a SQL DSL (~>, |=>), and a background worker — no Redis, no external orchestrator. Preview, PostgreSQL 17. (microsoft/pg_durable)
• The integer at the end of the universe: integer overflow in Postgres — what actually happens when an int4 sequence runs out of headroom, how to spot it before it bites, and how to migrate to bigint without a full-table-rewrite outage. (Crunchy Data)
• Async I/O in PostgreSQL 19: the year after — Christophe Pettus revisits the io_uring/AIO subsystem one release on, covering the new auto-scaling worker model and what changed for operators between 18 and 19. (thebuild.com)
• Handling graphs with SQL/PGQ in PostgreSQL — worked examples of the ISO SQL/PGQ property-graph syntax (CREATE PROPERTY GRAPH, GRAPH_TABLE) shipping in PG19, run over ordinary relational tables. (Cybertec)
• PostgreSQL and the OOM killer: why we use strict memory overcommit — Ubicloud argues for vm.overcommit_memory=2 so allocations fail predictably instead of the kernel killing a random backend (often the postmaster) under memory pressure. (Lev Kokotov · ubicloud.com)
• Authenticate PostgreSQL with SSH keys — a PAM module that lets Postgres verify connections against users' existing SSH public keys, reusing key infrastructure you already manage instead of yet another credential store. (ChronicallyJD)

Community pulse

• pg_durable: Microsoft open sources in-database durable execution — the week's biggest DB thread debated whether durable execution belongs inside the database at all vs. an external orchestrator, and what it means for crash semantics and operability. (Hacker News · 474 pts, 108 comments)
• databow: a Rust CLI to query any database with an ADBC driver — discussion of using ADBC as a universal columnar wire layer, and whether it finally beats the ODBC/JDBC status quo for cross-engine tooling. (Hacker News · 116 pts, 24 comments)
• Is Azure capacity this constrained, or am I doing it wrong? — an operator stuck behind "backlogged" quota approvals for Postgres Flexible Server and vCPU in multiple regions; the replies are a field guide to cloud-region capacity war stories. (Hacker News · 11 pts, 15 comments)

Wider DBMS & distributed data

• databow: a Rust CLI to query any database with an ADBC driver — a single command-line tool that speaks to any database exposing an ADBC driver, returning Arrow columnar results without per-engine glue. (columnar.tech)
• Ladybug: the DuckDB of graph databases — an embeddable, single-binary property-graph engine pitched as the local-first analytics counterpart to DuckDB for graph workloads. (ladybugdb.com)

Migration experience

• Functionally migrating from Firestore to PostgreSQL — a from-the-trenches account of moving a document workload off Firestore onto Postgres: modelling collections as relational tables, preserving query semantics, and the cutover mechanics. (Valentin Mouret)
• Postgres serials should be bigint — and how to migrate — why serial/int4 keys are a latent time bomb and a concrete, low-downtime procedure to widen them to bigint on a live table. (Crunchy Data)

International (non-English sources)

• 2 + 2 = 6: fixing lost updates in Postgres — a systematic tour of race conditions under each isolation level with measured costs of each fix (SELECT FOR UPDATE, advisory locks, SERIALIZABLE, atomic UPDATE) on a real highload workload. (nikolayschaback · Habr) [ru] (orig: «2 + 2 = 6 и как мы это фиксим: lost updates в Postgres»)
• Killing the REFRESH MATERIALIZED VIEW queue in PostgreSQL — SKB Kontur's hour-long MV refreshes were piling into a connection-hogging queue; they used pg_try_advisory_xact_lock per-MV plus trigger-tracked dependencies so a redundant REFRESH skips instead of waiting. (makushevski · Habr / SKB Kontur) [ru] (orig: «Как мы убрали очередь из REFRESH MATERIALIZED VIEW в PostgreSQL»)
• VARCHAR(N) in PostgreSQL is a constraint, not a memory saving — myth-busting the varchar(255) habit: Postgres stores the actual string regardless of the limit, so N buys you a check constraint, not packed storage; when it helps and when it's cargo-culted. (alwaysdeterminated · Habr) [ru] (orig: «VARCHAR(N) в PostgreSQL: ограничение, а не экономия памяти»)

Call for papers

• PostgreSQL Conference Poland — PGConf.PL 2026 — Gdańsk, Nov 24 — the inaugural PGConf.PL; CfP opened this week (Jun 4), closes Jul 13, 2026. (community)

18 items · sources scanned: ~30 · community threads scanned: Hacker News (Algolia, postgres+database queries) · international: Habr [ru] in-window; Dalibo [fr] and Qiita [ja] had no in-window items · filtered out as marketing/ads: ~14

• 5 PostgreSQL locking behaviours that trip people up — counterintuitive production-outage locks from a PostgreSQL engineer: FIFO lock-queue chained blocking behind a waiting ACCESS EXCLUSIVE, invisible FK FOR KEY SHARE deadlocks, unique-check INSERT deadlocks, the wraparound-prevention autovacuum that refuses to be cancelled, and VACUUM's hidden truncate-phase ACCESS EXCLUSIVE (and how it force-cancels standby SELECTs via max_standby_streaming_delay). (Shinya Kato · dev.to)
• How soon is now()? A frozen-timestamp retry-loop bug — a real distributed-locking bug: now() returns the transaction start time, so a retry loop sharing one transaction re-evaluated last_updated < now() - timeout against a constant and never made progress; fix is clock_timestamp(), with the testing-seam lesson on why unit + e2e both missed it. (Oskar Dudycz · event-driven.io)
• TOAST: where PostgreSQL hides big values — reproducible walkthrough of the over-2KB out-of-line storage path: compression strategies, the per-column STORAGE knob, why a wide text/jsonb column silently changes your I/O profile, and how to see TOAST traffic. (Radim Marek · boringsql.com)
• Why Postgres lacks transparent data encryption — why TDE never landed in core (the buffer/WAL/temp-file surface area, key management, and the recurring pg_tde-style extension route requiring a patched server), and what at-rest encryption actually buys you versus filesystem/volume encryption. (pgEdge)

Community pulse

• SQLite is all you need for durable workflows — the week's biggest DB thread: a long argument over building durable-execution engines directly on an embedded DB instead of a dedicated workflow system, trading off single-node simplicity against scale. (Hacker News · 720 pts, 387 comments)
• Building durable workflows on Postgres — the Postgres-flavoured half of the same debate (DBOS): can the database be your only durable-execution substrate, and where do queues/outboxes still beat it. (Hacker News · 359 pts, 147 comments)
• CQL: Categorical Databases — category-theory-based data integration and schema mapping resurfaced and drew a genuinely technical thread on whether the formalism pays its way in practice. (Hacker News · 111 pts, 39 comments)
• SQLite does not accept agentic code — Simon Willison on SQLite's stance against AI-generated contributions, prompting a wider argument about provenance, testing rigour, and the project's famously high bar. (Simon Willison · Hacker News)

Wider DBMS & distributed data

• Multi-stage distributed query execution in ClickHouse Cloud — how ClickHouse moved from a single distributed-aggregation step to a multi-stage exchange model, with the shuffle/repartition mechanics that let joins and aggregations scale across nodes. (clickhouse.com) [vendor blog — substantive]
• SlateDB — an embedded database built on object storage — an LSM-style embedded store that puts its storage layer directly on S3-style object storage, trading local-disk latency for elastic, separated storage. (slatedb.io · via Hacker News)
• Show HN: Streambed — stream Postgres to Iceberg on S3, speaks the Postgres wire protocol — CDC from Postgres into Iceberg while exposing a Postgres-wire endpoint over the lakehouse; the thread digs into the wire-protocol emulation and where it diverges from real Postgres. (github.com/viggy28/streambed · via Hacker News · 129 pts)

International (non-English sources)

• Becoming postgres in someone else's cloud: a security crash-test of managed databases — a Yandex Cloud reliability lead documents two privilege-escalation vectors in managed PostgreSQL where the server runs code as superuser and bypasses the control-plane's restricted roles; both were found at two public cloud providers and fixed upstream and by the vendors before publication. (Evgeny Efimkin · Yandex Cloud / habr.com) [ru] (orig: «Как стать postgres в чужом облаке: краш-тест безопасности управляемых БД»)
• Why Postgres with an external OLAP engine isn't really HTAP — argues that bolting an external columnar engine onto Postgres (separate storage model, ETL/CDC, parquet hand-off) breaks the single-transaction-visibility property real HTAP needs, and sketches doing OLTP + analytics over shared storage inside Postgres instead. (Tantor Labs · habr.com) [ru] (orig: «Почему PostgreSQL с внешним OLAP — это не совсем HTAP?»)
• Polymorphic associations in PostgreSQL: helping the planner avoid performance cliffs — how the ORM-generated discriminated-FK pattern (type, id LEFT JOINed to every subtype) regresses, and three pgsql-hackers patches discussed in 2024–2026 that each target a different source of the regression in the optimizer. (Andrei Lepikhov · Tantor / habr.com) [ru] (orig: «Полиморфные ссылки в PostgreSQL: помогаем СУБД избежать провалов производительности»)

Upcoming events

• Swiss PGDay 2026 — Rapperswil, CH, Jun 25–26 — the Swiss PostgreSQL conference, two tracks (EN/DE). Internals picks:
  • Deep dive into the PostgreSQL Front/Backend protocol (Xavier Fischer, EDB) — the wire protocol every pooler, proxy, and distributed router must speak.
  • Optimizer Hints in PostgreSQL (Michael Banck, credativ) — timely against PG19's new pg_plan_advice.
  • Zooming in on OOM: Postgres & Linux memory (Oleksii Kliukin & Dimitris Sideris, Tiger Data) — backend/OOM-killer interaction and memory accounting.

14 items · sources scanned: ~30 · community threads scanned: HN (postgres/database/sqlite/clickhouse queries, date-ranged) · international: ru (Habr) · filtered out as marketing/ads: ~10

• The four PG19 features you'll actually feel in production — ahead of the (then-imminent) PG19 beta, Christophe Pettus skips the SQL/PGQ graph headline and flags the operational changes: 64-bit MultiXact members (kills the "vacuum-or-die" member-exhaustion failure mode), parallel autovacuum index workers (watch the autovacuum_max_workers × autovacuum_max_parallel_workers × maintenance_work_mem worst case), temporal UPDATE/DELETE ... FOR PORTION OF, and jit = off by default (OLAP shops should re-enable and benchmark before upgrading). (thebuild.com)
• pgBackRest gets a multi-vendor lifeline after its maintainer steps back — the most widely-used Postgres backup tool was archived in late April when its sole maintainer of 13 years (David Steele) lost funding after the Crunchy Data acquisition; a coalition (AWS, Percona, Supabase, pgEdge, Tiger Data, Eon) stepped in to fund continued development. (The Register)
• Backrest's back, alright! — the anatomy of a sustainability gap — Jan Wieremjewicz reframes the panic: an archived repo isn't EOL, it's the "Nebraska guy problem" — critical infra carried by one person — and argues forks fragment trust where coordinated funding restores continuity. (Jan Wieremjewicz · percona.community)
• pg_infer 1.0.0 — transformer model internals as SQL relations — not text-to-SQL: a PG18+ index AM that brings small-transformer inference into the query plan as a planner-costable operator, exposing gate activations / learned associations / embeddings as scannable relations. (Greg Burd · postgresql.org)
  • <~> model-similarity operator is index-backed and composes with WHERE/JOIN/aggregation; describe(), walk(), implies() probe what a model "knows"
  • "model" indexes WAL-log the entire vindex inside PG pages, so backup/replication/PITR cover the model like any table; targets CPU inference (OpenBLAS, BitNet b1.58 ternary weights) and idle replica CPUs via a remote backend
  • PostgreSQL 18+; built with pgrx; credits the upstream LARQL project
• pg_tre 1.1.1 — approximate-regex index access method — a real IndexAmRoutine (USING tre) doing Levenshtein-distance regex matching ("within k edits of this pattern"), with WAL coverage, VACUUM awareness, and REINDEX CONCURRENTLY; three-tier filter funnel (BRIN-style range bloom → trigram postings → per-tuple bloom) over Laurikari's TRE library. Fills a gap pg_trgm/FTS/pgvector don't cover. PostgreSQL 18+. (Greg Burd · postgresql.org)
• pg_mentat 1.3.0 — Datomic-compatible Datalog inside Postgres — immutable datoms, a Datalog compiler that lowers to ordinary SQL, the pull API, time-travel (as-of/since/history), and ACID via PG transactions; built with pgrx in Rust for PG13–18, with soft integrations into pg_trgm/fuzzystrmatch/pgvector/PostGIS/etc. (Greg Burd · postgresql.org)
• pg_sorted_heap 0.14.0 — physically-sorted heap storage with zone-map pruning and a planner-integrated vector-search path. (Sergey Kuznetsov · postgresql.org)
  • sorted_hnsw made planner-visible via the Index AM path for covered KNN, including parent Merge Append over leaf HNSW indexes
  • composite zone-map metadata (first-key block ranges + bounded second-key summary); sorted_heap_bulk_load_ordered(...) trusted ordered ingestion
  • verified on PostgreSQL 16, 17, 18; now on PGXN
• plpgsql_wrap v1.0 — Oracle-style PL/pgSQL obfuscation — stored procedures written LANGUAGE plpgsql_wrap are AES-256-GCM-encrypted into pg_proc.prosrc; source is unreadable (including in pg_dump, whose output stays restorable) without the key, otherwise behaves identically to plpgsql. PostgreSQL 12+. (HexaCluster · postgresql.org)
• pg_statviz 1.0 — with opt-in AI analysis — minimalist in-database stats time-series extension, now with an opt-in --ai analysis flag. (pg_statviz project · postgresql.org)
  • --ai sends each chart's data/PNG to a vision LLM (Claude/Gemini/local Ollama) for a HEALTHY/WARNING/CRITICAL verdict plus remediation
  • notable design discipline: a deterministic rules engine runs first and a "severity floor" prevents the LLM downgrading a real finding; prompts wrap user data in <user_data> envelopes for prompt-injection containment
  • AI is fully optional; with no --ai flag there are zero extra dependencies and no external calls
• Barman 3.18.0 — new release of EDB's backup-and-recovery manager for PostgreSQL. (EDB · postgresql.org)
• check_pgactivity 2.10 — new release of the Nagios/monitoring plugin for PostgreSQL. (check_pgactivity project · postgresql.org)

Wider DBMS & distributed data

• Postgres is all you need for durable workflows — Peter Kraft argues external workflow orchestrators (Temporal, Airflow, Step Functions) are overcomplicated: if durable execution is really about checkpointing to a database, the database can be the orchestrator — workers cooperatively dequeue from a Postgres table via locking clauses, checkpoint each step, and recover each other's state, inheriting Postgres's replication/observability/security instead of adding a new point of failure. (Peter Kraft · dbos.dev)

Migration experience

• 21 CFR Part 11 audit logging in Postgres without triggers — Nathan Leung walks through why Harbor rejected pgAudit (no app context) and PL/pgSQL triggers (SET LOCAL plumbing, per-study-DB duplication, PgBouncer transaction-pooling wariness) and instead wrapped their Drizzle ORM in a JS Proxy that forces .withAuditLogCtx(...) before any mutation, writing the data change and a signed before/after audit row in one transaction (SELECT ... FOR UPDATE for a consistent snapshot). Honest trade-offs section: extra round trips, weaker guarantee than a DB-level trigger, and a note on eventually moving signing into pgcrypto. (Nathan Leung · runharbor.com)

Call for papers

• PGDay Israel 2026 — Tel Aviv, Oct 25 — CfP opened this week (May 21); submissions close Jul 1, 2026. (community)

Upcoming events

• DEBS 2026 — Lisbon, PT, Jun 23–26 — the ACM conference on Distributed and Event-Based Systems; close to Postgres logical-decoding/CDC and real-time analytics. Full program still going up; one accepted paper already of note:
  • Lean Gossip: Big Gains from Small Talk — a leaner gossip / membership-dissemination protocol; applicable to cluster membership and failure detection in distributed Postgres deployments.

12 items · sources scanned: ~30 · community threads scanned: HN (Algolia, postgres + database queries — a quiet week, no in-window thread cleared the engagement bar) · international: Habr/Qiita/Dalibo feeds checked but none reach back to the window (plain feeds only carry recent items) · filtered out as marketing/ads: ~7