Sporadic build failures on GitHub Actions

[rolyp]

The build now runs ~20 jobs for various configurations, but one or two jobs often stall and never resume. Sometimes this may be down to bandwidth or download issues, but often it looks like the job has just stalled in the GitHub Actions runner. The workaround is to cancel the build when this happens and run again until all jobs complete.

Basic problem. In recent days, it seems that, with Python 3.6+ on Linux, test_recombine can no longer use 240 dimensions – the process hangs (as opposed to just taking a long time). 240 used to be a small enough number; now 90 reliably fails. 60 works, so that is the current setting. Further investigation is required to see if there is an underlying problem, or whether this is related to (changes in) the resources available to GitHub Actions.

[rolyp]

Not going to worry about these, as the difficulty of getting the system building in the first place is significantly greater than any problems associated with the build occasionally failing for a random reason. GitHub claim some of these are “in progress” issues.

[rolyp]

The Windows failures are (at least) to do with:

• curl: (28) Failed to connect to ayera.dl.sourceforge.net port 443: Timed out
• curl: (18) transfer closed with xxxx bytes remaining to read

CircleCI experiment (Windows):

• [x] switch from curl to curl.exe to avoid Invoke-WebRequest incompatibility
• sporadic failures:
  • curl randomly fails at 29% with exit code 1 :x: not reproducible
  • curl: (56) Failure when receiving data from the peer
• [x] curl fails with Invoke-WebRequest missing required parameter – switch to curl.exe
• [x] Write-Progress : Win32 internal error “Access is denied” 0x5 occurred while reading the console output buffer. in Expand-Archive – set $global:ProgressPreference to disable progress output
• [x] then no credits to run remaining tasks

[rolyp]

Reopened this issue, as the build now functions pretty well, and with 20+ platform targets, these sporadic failures means it usually takes multiple attempts to do a full build.

[rolyp]

Re-closing this issue. The sporadic curl failures ([Errno 14] PYCURL ERROR 18 - “transfer closed with x bytes remaining to read”) remain, but these aren’t too frequent (or easy to fix).