solve CI slowdown under Node 22

[turadg]

What

the test-all-packages workflow is about 20% slower under Node 22 than Node 20.

Using the recent successful Test all Packages workflow runs from .github/workflows/test-all-packages.yml, Codex pulled 44 successful runs (from the latest 100 runs), extracted all paired node-old/node-new job timings, and fit the model with job as a random effect.

Dataset

• 44 runs
• 9 job slots per run (test-boot and test-swingset shards)
• 792 job rows total
• 396 paired comparisons

Raw means

• Node 20 mean job duration: 562.77s
• Node 22 mean job duration: 679.31s
• Difference: +116.55s (about +1m56s, +20.7%)

Paired test (within run+slot)

• Mean diff: +116.55s
• 95% CI: [112.11s, 120.98s]
• p-value: 1.14e-177

Mixed-effects model (job random effect)

• Model: dur_s ~ node_new + (1 | slot) with run control
• slot random intercept + run fixed effects:
  • Effect (node_new): +116.55s
  • 95% CI: [112.66s, 120.43s]
  • p-value: effectively 0 (underflow; extremely significant)
• Cross-check model with run variance component gave nearly identical estimate:
  • Effect: +116.55s
  • 95% CI: [112.13s, 120.97s]

So the estimate is very stable: Node 22 is about 1 minute 56 seconds slower per matched job than Node 20 in this workflow sample.

Why

We should understand how Node 22 vs Node 20 affects production. And what impact Node 24 will have,

• adopt Node 24 and drop Node 20 #11702

How

TBD