#hacking process

B-tree merge: correctness first

Test Salma El-Sayed’s B-tree leaf-page merge prototype: snapshots, parallel scans, VACUUM, and standby safety.

Nik Samokhvalov, Kirk Wolak, and Salma El-Sayed tested the B-tree leaf-page merge prototype and agreed to prioritize correctness before benchmarks. The discussion covered boundary pivots, merge-group metadata, concurrent readers, and deferred VACUUM cleanup. Nik reported standby and parallel-scan failures; the group examined index-only-scan snapshot safety and the lifetime of pinned pages. The next step is focused reproduction, small fixes, and retesting against exact revisions, not a claim that the prototype is ready to ship.

Summary from session notes; timecodes pending captions.

Agentic Hacking Postgres: Committer in Training

Standby snapshot corruption, seven upstream bug fixes, and a plan for adversarial CommitFest review.

Nik Samokhvalov, Andrey Borodin, and Kirk Wolak discussed standby snapshot corruption and how to review AI-assisted patches without losing human judgment. Nik reported about 30 bug reports with proposed fixes, seven committed fixes, and two false positives. The main proposal was Committer in Training: a small, low-risk CommitFest queue reviewed by adversarial agents, with red-green tests and public evidence. The group agreed to start a static bug-fixing record and work toward a patch-addressable CommitFest overlay.

Summary from session notes; timecodes pending captions.

Agentic Hacking Postgres with Andrey, Kirk, and Nik

Review the PG20-2 backlog: minor upgrades, global temporary tables, wait instrumentation, and pg_stat_log.

The group turned a tour of the PG20-2 CommitFest into concrete review homework. They reconciled Andrey Borodin's prolific August as 25 currently authored entries, explored bounded pre-shutdown checkpoints and safe cache handoff for minor upgrades, and examined why global temporary tables require deep catalog, relcache, planner, statistics, and lifecycle work. They then compared Ilmar Yunusov's EXPLAIN wait-event patch with Dmitry Fomin's broader timing/tracing series, insisting on honest attribution and workload benchmarks. The most actionable outcome was pg_stat_log: Nikolay will independently benchmark and review its SQL-queryable error counters, Kirk will support the thread, Andrey will revisit Dmitry's series, and Kirk will revise his pg_dump placeholder interface.

Agentic Postgres hacking with Andrey, Kirk, and Nik

Agentic Postgres hacking with Andrey Borodin, Kirk Wolak, and Nik Samokhvalov. Detailed notes pending captions.

Transcript summary and timecodes pending captions.

Agent hacking Postgres

Agent-assisted patch work around the wait-event timing submission.

The group reviewed Kirk Wolak's pg_dump patch for keeping loadable placeholders when table data is excluded, grounding the discussion in his large-table restore workflow. They debated a positive interface, filter-file actions, arbitrary dump queries, PII transformation, and referential-integrity risks, then posted supportive review feedback to pgsql-hackers. The final segment examined Salma El-Sayed's B-tree leaf-page merge design, including merged-page markers, scan correctness, vacuum cleanup, and review plans.

Hacking Postgres with Kirk, Andrey, Nik

Custom command tags and logical decoding from an archive-fed recovery server.

The workshop first revisited customizable command tags, confirming legacy behavior remains the default and its added branch should be negligible, then explored extensible client protocols and better psql completion. The main direction was extracting data from managed Postgres without burdening production: decode archived WAL into a logical stream. They compared a full recovery server with a standalone tool, concluding the latter would need schema state at a starting LSN, checkpoint/full-page-image handling, and careful treatment of file swaps such as pg_repack.

GSoC ideas for Postgres

Turn workshop ideas into scoped GSoC projects, including wait-event coverage, pg_stat_statements scalability, and B-tree page merge.

The group turned Postgres GSoC brainstorming into a draft set of core-development projects. After Nikolay recounted his 2006 SQL/XML GSoC experience, they evaluated planner, parallel GiST build, TOAST slicing, test coverage, logged/unlogged tables, global temporary tables, and autovacuum ideas. They then added wait-event coverage, pg_stat_statements scalability, and B-tree page merge, using an AI-generated wiki draft. The draft still needed expert validation, clearer mentor ownership, and manual editing before merging into the official ideas page.

How to create Postgres patches with AI?

AI-assisted patch discovery, implementation, review, and submission.

The session developed a responsible AI-assisted patch workflow: treat prompts as engineering specifications, assign independent security, documentation, and testing reviewers, distrust praise, and use agents to summarize long CommitFest threads without surrendering judgment. They then applied that workflow to a patch that logs the LSN of DROP TABLE, enabling precise PITR after accidental deletion. The agent retrieved and rebased the patch, planned transaction, rollback, savepoint, and documentation tests, and continued adapting it to changed GUC infrastructure; final review remained pending.

Vibe Hacking Postgres

Review B-tree merge blockers and prototype commit-LSN logging for DROP TABLE.

After reviewing UUIDv7 and amcheck support for GIN in Postgres 18, the group discussed responsible AI-assisted hacking and briefly parked the B-tree merge work over unresolved scan concurrency. Most of the session revisited recovery from accidental DROP TABLE: they asked an agent to extend transaction callbacks with the commit LSN and build a logging extension. The prototype found table names through object-access hooks, but cleanup timing and an invalid commit LSN prevented a working result, leaving the design unfinished.

Vibe Hacking Postgres

Test B-tree page merge against concurrent scans and expose correctness failures.

The group assessed the B-tree page-merge prototype after community feedback. They agreed benchmarks should emphasize extra buffer I/O, cache pollution, WAL, and sparse-page workloads rather than lookup latency alone. A new injection-point test then demonstrated the critical flaw: a concurrent forward scan returned duplicates while a backward scan missed tuples after pages were merged. Several recovery schemes were examined, but repeated vacuums and concurrent inserts defeated the simple designs. The session ended with correctness and concurrency still unresolved.

Vibe Hacking Postgres

Rebase GiST amcheck and design useful corruption tests.

The group revived the long-running amcheck work for GiST indexes after shared infrastructure and the GIN checker had been committed. They manually rebased the remaining GiST patch, repaired build files and SQL upgrade scripts, and got its basic checks passing. They then asked Cursor to adapt newly committed GIN corruption tests. The generated tests initially assumed ordering that GiST does not have and later passed without actually corrupting indexes, leaving only a useful page-header corruption idea and more manual work.

Vibe Hacking Postgres

Review commit-LSN logging for destructive DDL and its test strategy.

The workshop resumed the patch that logs a recovery LSN when destructive DDL commits. Review showed AI-generated schema support used the wrong hook, added an uncalled function, and supplied a nondeterministic regression test unsuitable for raw LSN output; a TAP test was deemed necessary. Rather than finish the substantial cleanup live, they switched to optimizing the regression-test schedule with Cursor. Its reordered groups caused failures and thermal throttling, demonstrating that faster scheduling must respect test dependencies and stable benchmarking conditions.

Vibe Hacking Postgres

Add a pg_resetwal system-identifier option and audit the agentic patch workflow.

Nik presented an AI-assisted pg_resetwal option for changing the system identifier, needed to replay archived WAL after DBLab created a replacement cluster. Review accepted the use case but called for removing interactive safeguards, shortening documentation, reducing redundant tests, and aligning option identifiers with other frontend tools. The session then demonstrated Nik's agentic patch workflow: small commits, specification-first documentation, compilation and manual tests, strict review by stronger models, and a final human review that caught issues the models missed.

Vibe Hacking Postgres

Merge improvements into commit-LSN logging for destructive DDL.

The session opened with AI-assisted Postgres development lessons from the committed pg_dump policy patch and a separate wal_compression-level prototype: write tests and documentation early, review every line, and use several strong models as independent critics. The practical task was merging Jim Jones's improvements into the patch that logs commit LSNs for destructive DDL. Cursor produced overbroad changes, but after resetting and supplying both patches, the group compiled and manually tested the result, found a callback-unregistration bug, and attempted test-driven repair.

Hacking Postgres with Andrey and Nik

Redesign DROP TABLE recovery logging around the commit LSN.

The session opened with recent Postgres work on GIN amcheck, corruption detection, read streams, and pg_stat_statements sampling, then returned to the DROP TABLE recovery patch. The patch logged an attempted drop before commit, producing misleading landmarks after rollback. They redesigned transaction callbacks to receive the commit LSN, manually changed the commit path after an AI attempt stalled, and demonstrated commit-only logging. Remaining work included two-phase commit, schema/OID formatting, tests, and incorporating Jim Jones's review before a new version.

Hacking Postgres — Andrey, Kirk, Nikolay

Review UUIDv7 and build deterministic transaction_timeout tests.

The first section reviewed UUIDv7's evolving RFC, monotonic counter, distributed generation, timestamp extraction, and blocked core support. They also surveyed transaction-timeout and SLRU-locking work. The hands-on section tackled flaky transaction-timeout tests by combining injection points with a background psql watch loop. Testing uncovered false success from malformed watch parameters and a broken-pipe cleanup problem when timeout killed psql. The harness remained unfinished, needing a reliable way to detect whether the background session is active.