Repository navigation
Commit 25e75e1
committed
Give the nightly E2E test job a realistic timeout
The Windows legs of the python-nightly matrix run right up against Azure
Pipelines' implicit 60-minute job timeout, so they get cancelled even though the
test suite itself passes. This is what has been showing up as a "flaky" nightly.
Measured on build 161689 (a clean run, setup and cleanup both green):
py39_windows_mqtt 60.4 min -> CANCELLED
py39_windows_mqttws 52.9 min -> passed, ~7 minutes of headroom
the seven linux legs 31.9 - 34.7 min
The cancelled leg had already printed its pytest summary before being killed:
========= 74 passed, 33 skipped, 67924 warnings in 3437.00s (0:57:16) =========
So nothing actually failed -- 0 failures, 74 passed -- the job simply ran out of
wall clock during teardown. Whichever Windows leg happens to be slower on the day
is the one that dies, which is why the failure appears to wander between
py39_windows_mqtt and py39_windows_mqttws across runs.
vsts/python-nightly.yaml sets no timeoutInMinutes anywhere, so every job inherits
the 60-minute default. Set 90 minutes on the Test job.
This is a CI-budget fix and is unrelated to the CodeQL change in this PR, but it
has to live on this branch for this PR's own python-nightly check to pass. Happy
to split it into its own PR if preferred.1 parent a9364c9 commit 25e75e1
1 file changed
Lines changed: 11 additions & 0 deletions
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
62 | 62 | | |
63 | 63 | | |
64 | 64 | | |
| 65 | + | |
| 66 | + | |
| 67 | + | |
| 68 | + | |
| 69 | + | |
| 70 | + | |
| 71 | + | |
| 72 | + | |
| 73 | + | |
| 74 | + | |
| 75 | + | |
65 | 76 | | |
66 | 77 | | |
67 | 78 | | |
| |||
0 commit comments