Skip to content

EPIC 20 merged/h: five-step rollout (supply, arm, flakes, cap, runners) #5510

Open
79 of 100 issues completed
Open

Description

@noahgift

owner: aprender-pace (set 2026-10-19 by aprender-triage-worker at the cop's word; picks up children only as batch constituents after its current timed task)

EPIC: 20 merged PRs an hour (FLOW-005 v2)

Today aprender merges 0.42 PRs/h (git, last 24 h). Spec: FLOW-005 (proposed) v2, "aprender at twenty merged PRs an hour", 2026-10-18 (cop relay). Little's Law: lambda_max = cap / W. At cap 10 and W ~ 22 h the ceiling is 0.45/h (measured 0.42/h); 20/h needs ~27 PRs open. Order matters: step 1 shrinks W; the cap rises only after a live merge-queue test confirms the model. The full test set stays on the merge-queue commit throughout, except a test demoted GATE→LAB under C324 (e.g. test_mega_long_context_memory_bound, #5862): it is #[ignore]d on PR, merge-queue and release jobs, runs nightly with --ignored, and a red opens one ticket with one owner. That demotion is a scope change, not a waiver; deleting a test is a waiver.

Rollout: one gate at a time (v2 order)

A step starts only after the previous gate's number holds for 24 hours. Rule changes (ruleset, cap, runner groups) go to a quorum with a non-author sign-off.

# Step Date Ticket Gate (24 h, verbatim)
1 Shrink W: pull, ready queue >= 2N, idle andon 10 min, wsweep (no PR > 24 h, 16 -> <= 10), auto-arm (#5490), red PR pushed within 30 min today #5498 W <= 5 h · merged >= 2/h · opened >= 3/h · arm p90 <= 10 min
2 Queue test: maxEntriesToBuild 5 -> 1 + one 8-PR burst under cap 10 10-19 #5505 (see #5496) >= 7 PRs per green pass at depth >= 8
3 Flakes: root fixes, infra#1753, #5489, x86-main own lane 10-21 #5506 flake <= 2% · cancels <= 10% · pass p50 <= 30 min · runs/PR <= 1.3
4 Cap 10 -> 30, sized from W measured after step 1 10-22 #5508 merged >= 8/h · W <= 3 h
5 Capacity: PR pool 38 or fast tier 12 10-26 #5509 merged >= 20/h for 4 h · W <= 90 min
parallel aprender-simulate SIM-1..5 (model, contract tests, drift andon) with step 1 #5511 SIM-3 green; SIM-4 wired; SIM-5 row in two consecutive batches

Each ticket carries its falsifier verbatim from v2.

Rules

  • No new Python. The reference model.py/sim.py stay outside the repo.
  • Questions in any step go to a planted quorum (C314); a gate-script change needs a non-author sign-off.
  • PRs are <=200 lines, one ticket each (Step 1/5 Shrink W: pull, andon, wsweep, auto-arm, red PR pushed in 30 min; <=200-line PRs #5498).
  • LAB until the operator gates it (C324): throughput numbers are measured, never block a PR.
  • Each gate result is logged: number, timestamp, source command, or not measured and why.

One-way doors and enforcement (operator order, 2026-10-18)

Each needs a planted quorum (lanes not the author model) and a non-author sign-off:

Children tracked here (REDUCE-1022)

Activity

  1. added this to the 0.73.0 milestone on Oct 18, 2026
  2. 74 remaining items

  3. added sub-issues on Oct 22, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    epicEpic — multi-story umbrella

    Type

    No type

    Projects

    No projects

      Milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions