Observed behaviour
The Run CI workflow duration jumped from ~17–20 minutes to ~35–40 minutes, first seen on 2026-03-31. The increase is consistent across all three Python matrix jobs.
Root cause
PR #742 (#742) accidentally removed @pytest.mark.block_network from test_openai_compat in test/formatters/granite/test_intrinsics_formatters.py during the marker taxonomy audit.
That test deliberately connects to an unreachable URL (http://example.com:98765/...) and expects an openai.APIConnectionError. The comment in the test still reads "Note that network access is blocked for this test case."
With the marker in place, pytest-recording intercepts the socket call immediately. Without it, the OpenAI client makes a real TCP connection and waits for the OS-level connect timeout — approximately 31 seconds per call. The test makes 2 calls per parameterised case across 17 cases:
17 cases × 2 calls × ~31 s = ~18 minutes extra per matrix job
Evidence
|
Run Tests |
Passed |
Skipped |
| 2026-03-30 (pre-regression) |
711 s |
721 |
151 |
| 2026-04-07 (post-regression) |
1666 s |
873 |
125 |
Pre-regression each test_openai_compat case completed in ~2.8 s; post-regression each takes ~63 s.
Observed behaviour
The Run CI workflow duration jumped from ~17–20 minutes to ~35–40 minutes, first seen on 2026-03-31. The increase is consistent across all three Python matrix jobs.
Root cause
PR #742 (#742) accidentally removed
@pytest.mark.block_networkfromtest_openai_compatintest/formatters/granite/test_intrinsics_formatters.pyduring the marker taxonomy audit.That test deliberately connects to an unreachable URL (
http://example.com:98765/...) and expects anopenai.APIConnectionError. The comment in the test still reads "Note that network access is blocked for this test case."With the marker in place,
pytest-recordingintercepts the socket call immediately. Without it, the OpenAI client makes a real TCP connection and waits for the OS-level connect timeout — approximately 31 seconds per call. The test makes 2 calls per parameterised case across 17 cases:Evidence
Pre-regression each
test_openai_compatcase completed in ~2.8 s; post-regression each takes ~63 s.