postgres=# select * from sessions where video_id = 'F9-oFC0H9Fg';

PG Roast: challenge the idea before reviewing the code

· Hacking process

Designing an adversarial proposal reviewer, a CommitFest Shortlist, and evidence-first feedback that saves committer time.

session summary

Nik Samokhvalov, Kirk Wolak, and Andrey Borodin explored PG Roast: an adversarial reviewer that asks whether a Postgres proposal is worth accepting before examining its implementation. The call connected precedent-backed objections and practical testing with a curated CommitFest Shortlist, expert routing, and shorter human-owned mailing-list replies. The agreed pilot starts privately on Andrey’s own CommitFest patches. These are discussion notes and proposed work, not evidence that the complete system has shipped.

Based on November 11 proposal, call notes, and closing decisions (not a verbatim transcript). Caption-backed timecodes are not available yet.

timecodes

Timecodes pending.

discussion and decisions

Challenge the idea before the implementation

Start with the problem, user demand, and whether the idea belongs in Postgres; then examine design and implementation. Andrey pointed to Robert Haas’s “Committer Paranoia” as a review rubric. Prior rejection threads should supply evidence, not an excuse to manufacture objections.

Seven perspectives, then a critic

The proposed first version covers demand, standards and compatibility, prior attempts, replication and on-disk changes, security, maintenance and fit, and existing alternatives. A critic attempts to refute the objections. High-risk subsystem changes call for expert review, not automatic rejection.

A shortlist, not mass rejection

The group proposed selecting roughly 10–20 patches worth committer attention. Correctness and fit are separate from priority: useful but neglected work may need a champion, bundling, or a design rethink. Age alone is not grounds for closing an entry. “CommitFest Shortlist” was the preferred working name.

Find the actual blocker and the right reviewer

Each entry should explain the real unresolved pushback with message links. Personalized or subsystem-based lists and visible claiming can avoid duplicated effort. Git history, trailers, and discussion participation can suggest relevant experts; contact stays human-mediated.

Preflight, field tests, and deeper follow-ups

The discussed pipeline starts with mechanical checks, reusing cfbot evidence where available, then design critique and scenarios that exercise a patch as if committed. Supervisor rounds ask what was missed. The B-tree merge discussion motivated explicit subsystem invariants instead of assuming an agent will infer them.

Short human-owned mail, optional detail

Keep the message body short and take responsibility for its claims. Clearly labelled AI detail can live in attachments or an opt-in report. The discussion favored calm, evidence-first language rather than publishing the “roaster” persona to the list.

Start privately and measure the results

The closing pilot starts with CommitFest entries and Andrey’s own patches, with the author involved. Historical proposals and reverted commits were suggested for evaluation. A post-commit testing watcher and a reduction in the active backlog are proposed goals, not measured results.

follow-up from this session

Owners and targets from the session notes. Targets are plans, not completed releases.

  • Nik — Finish the first PG Roast prototype and share results with the group.Prototype and human-reviewed findings; completion not verified here.
  • Kirk — Develop the post-commit testing agent and the short-body / AI-attachments email proposal.Follow-ups recorded in the call wrap-up.
  • Andrey — Share existing CommitFest summary/state tooling and pilot reviews on his own patches.Avoid duplicate tooling and get author feedback.
  • All — Choose historical proposals and reverted commits for blind evaluation.Measure useful objections and false alarms before expanding the pilot.