[potiuk]

The scheduled Spot interruption notice run failed when GitHub dropped a connection mid-request: https://github.com/apache/airflow/actions/runs/40590296911/job/122492480657

Could not handle run https://github.com/apache/airflow/actions/runs/40178639592: stream error: stream ID 3; CANCEL; received from peer

Only (HTTP 5xx) counted as transient, so this HTTP/2 stream reset was neither retried nor skipped with a warning like a server error, and the run exited 1 although the next scheduled run does the same work again. Dropped connections (stream resets, connection resets, unexpected EOF, TLS handshake and I/O timeouts) now count as transient too. A POST is still not retried, since one that lost its connection may have created the comment.

---

Was generative AI tooling used to co-author this PR?

• Yes — Claude Code (Claude Opus 5)

Generated-by: Claude Code (Claude Opus 5) following the guidelines

🤖 Generated with Claude Code

[FrankYang0529]

It looks like only stream error is covered here. How about adding a test for is_transient like:

@pytest.mark.parametrize(
    ("stderr", "expected"),
    [
        pytest.param("gh: Server Error (HTTP 502)\n", True, id="server-error"),
        pytest.param("stream error: stream ID 3; CANCEL; received from peer\n", True, id="stream-reset"),
        pytest.param('Get "https://api.github.com/repos/x": read tcp …: read: connection reset by peer\n', True, id="connection-reset"),
        pytest.param('Get "https://api.github.com/repos/x": unexpected EOF\n', True, id="unexpected-eof"),
        pytest.param('Get "https://api.github.com/repos/x": net/http: TLS handshake timeout\n', True, id="tls-timeout"),
        pytest.param('Get "https://api.github.com/repos/x": dial tcp 140.82.112.6:443: i/o timeout\n', True, id="dial-timeout"),
        pytest.param("gh: Not Found (HTTP 404)\n", False, id="client-error"),
    ],
)
def test_is_transient(spot, stderr, expected):
    assert spot.is_transient(subprocess.CalledProcessError(1, ["gh"], stderr=stderr)) is expected