CI perf: Gradle compatibility — 4 Windows hosted cells (36–42 min each) on every matching PR are 70% of the run (~5,000 Windows job-min/day) #1300

Measurement

Window: 2026-10-09 01:23 → 16:16 UTC. Runs are from workflows/gradle-compatibility.yml/runs?event=pull_request, first page of 100 (177 in 24h).

• Gradle patch compatibility on pull_request: 62 full runs in 14.9h (53 success, 9 cancelled after more than 8 min), or about 100 per day.

  • Since Skip duplicate Gradle compat legs on PRs #1255 merged at 14:09, there have been 6 full runs in 2.1h, about 70 per day.
  • Wall clock: p50 45 min, p90 48 (n=53 successful runs).

• After Skip duplicate Gradle compat legs on PRs #1255, a PR run is almost all Windows. For example, run 37948810769 (15:03, after Skip duplicate Gradle compat legs on PRs #1255) used 209 Windows job-min and 0.1 Linux.

  cell (windows-latest)	job min	Run the hosted suites
  gradle 8.14.3 / jdk 21 / hosted	42.2	41.8 (43 tests, finished in 2504.87s)
  gradle 7.6.6 / jdk 17 / hosted	41.1	40.8
  gradle 6.9.4 / jdk 11 / hosted	36.5	36.1
  gradle 9.8.0 / jdk 21 / hosted	26.2	25.8
  4 vendor cells	9.0 – 11.3 each, 41.7 total
  4 agent cells	3.6 – 5.3 each, 17.3 total
  build windows-latest	4.0

• The 4 hosted cells are 146 of the 209 Windows job-min (70%) and set the 45-min wall clock.

  • CI perf: Gradle patch compatibility — PR runs repeat ci.yml's Gradle e2e tier and run nightly-only extras (~17,000 Linux job-min/day) #1177's table from 62 runs agrees: 8.14.3 hosted p50 40.6 / p90 41.6 min, and 7.6.6 hosted p50 39.1 / p90 42.4 min.

• Failures: none of the 53 successful runs had a failed cell, and CI perf: Gradle patch compatibility — PR runs repeat ci.yml's Gradle e2e tier and run nightly-only extras (~17,000 Linux job-min/day) #1177 and Skip duplicate Gradle compat legs on PRs #1255 reported 0 PR failures in their windows. No Windows hosted Gradle cell failure on a PR was seen in this window.

Where the time goes

• All the time is in the e2e_redirect_gradle_build --ignored gradle_hosted_ step. Each of the 43 tests drives a real Gradle build on Windows, in groovy and kotlin DSL. In the log (job 113883777777), the tests take 60–180 s each.
• Setup, JDK selection, Gradle install and artifact download together take under 0.5 min per cell. There is nothing to trim outside the test step.

Root cause

• CI perf: Gradle patch compatibility — PR runs repeat ci.yml's Gradle e2e tier and run nightly-only extras (~17,000 Linux job-min/day) #1177 and Skip duplicate Gradle compat legs on PRs #1255 kept every Windows cell on pull_request on purpose. They are the only PR coverage of Gradle on Windows: gradle-compatibility.yml exclude block, and CI perf: Gradle patch compatibility — PR runs repeat ci.yml's Gradle e2e tier and run nightly-only extras (~17,000 Linux job-min/day) #1177's "Coverage" point 3.
• So each matching PR runs the full 43-test hosted suite four times on Windows, once per Gradle line.
• The paths: filter includes broad globs that most engine PRs touch:
  • crates/socket-patch-core/src/hosted/** and vex/**
  • commands/apply.rs, remove.rs, rollback.rs, vex.rs, scan/**
  • tests/common/**, Cargo.lock, crates/*/Cargo.toml

Proposed fix

In .github/workflows/gradle-compatibility.yml, cells, on pull_request only:

• Keep all 4 agent and all 4 vendor Windows cells. They are cheap: 59 job-min.
• Keep two hosted Windows cells: the oldest line, 6.9.4 / jdk 11, and the newest, 9.8.0 / jdk 21.
• Move 7.6.6 / jdk 17 / hosted and 8.14.3 / jdk 21 / hosted on windows to the nightly (17 4 * * *) and workflow_dispatch. One way is two exclude entries keyed on the event, like the existing macOS and ubuntu ones:

  - {os: "${{ github.event_name == 'pull_request' && 'windows-latest' || '' }}", gradle: '7.6.6', mode: hosted}
  - {os: "${{ github.event_name == 'pull_request' && 'windows-latest' || '' }}", gradle: '8.14.3', mode: hosted}

  Check that an exclude entry with an empty os value matches nothing on other events, as the existing entries do.
• Update the header comment ("Windows cells run on every PR that matches paths:").

Expected saving

• Per full PR run: 41.1 + 42.2 = ~83 Windows job-min (40%).
• At 60–100 full PR runs per day, that is ~5,000–8,300 Windows job-min/day. I use 5k (60/day) for ROI, in case PR volume falls after Run compatibility matrices on PRs only for their own files #1206.
• PR wall clock for this non-required workflow drops from ~45 to ~37 min, now bound by the 6.9.4 hosted cell.
• No change to ci-ok, clippy or the merge-queue critical path: this workflow is not required and does not run in merge_group.
• Linux and macOS: 0.

Coverage and risk

• Every test still runs on every PR:
  • The hosted suite runs on Windows on the oldest and newest Gradle lines.
  • It runs on ubuntu for all 4 lines in ci.yml's e2e PR tier and merge queue.
  • The vendor and agent suites run on Windows for all 4 lines.
• Windows hosted on 7.6.6 and 8.14.3 moves to the nightly. A regression that only hits Windows and only the middle two Gradle lines in hosted mode would surface the next morning instead of on the PR. In the windows sampled here and by CI perf: Gradle patch compatibility — PR runs repeat ci.yml's Gradle e2e tier and run nightly-only extras (~17,000 Linux job-min/day) #1177 / Skip duplicate Gradle compat legs on PRs #1255, no hosted Windows cell failed on a PR.
• Required check names ci-ok and clippy are untouched.

Effort

S. Two exclude lines plus a comment, in one file.

ROI

Saving: ~5k Windows job-min/day × 2 (Windows weight) = 10. Confidence: 0.6, since the run rate after #1206 is uncertain. Effort: S = 1. ROI = 10 × 0.6 / 1 = 6.0

Generated by Claude Code