macos-15-arm64 20260527.0100.1 — large performance regression in test execution vs 20260520.0085.1

[MortezaPoussaneh]

Summary

The macos-15-arm64 image version 20260527.0100.1 shows a roughly 22× slowdown in test execution compared to the prior release 20260520.0085.1, on the same workflow, same architecture, same Python version. The slowdown is reproducible across multiple runs (May 28 fast, June 1 and June 3 slow).

Timings

Date	Image version	Install dependencies step	Run tests step
2026-05-28	20260520.0085.1	16s	65s
2026-06-01	20260527.0100.1	15s	1411s (~23m31s)
2026-06-03	20260527.0100.1	15s	~1410s

Install dependencies is essentially unchanged. The regression is entirely in test execution.

Workflow

strategy:
  matrix:
    include:
      - os: macos-latest   # resolves to macos-15-arm64
        name: macos
runs-on: ${{ matrix.os }}
steps:
  - uses: actions/checkout@v5
  - uses: actions/setup-python@v6
    with:
      python-version: "3.13"
  - run: pip install -e .
  - run: <project's test runner>

The test suite is Python asyncio-based and uses multiprocessing for per-test subprocess isolation. The suite spawns many short-lived child processes during the run — if subprocess spawn or asyncio overhead changed in 20260527.0100.1, that matches the pattern (install untouched, exec dramatically slower).

Job logs (slow vs fast)

• Fast (20260520.0085.1, 1m29s): https://github.com/Eternego-AI/eternego/actions/runs/26570682141
• Slow (20260527.0100.1, 24m02s): https://github.com/Eternego-AI/eternego/actions/runs/26770558765
• Slow (20260527.0100.1, ~24m): https://github.com/Eternego-AI/eternego/actions/runs/26891059666

Happy to provide additional reproducer detail or run targeted timing experiments on request.