[python-committers] AppVeyor slowed down?

[Victor Stinner]

Hi,

Is it just me or AppVeyor became slower than it was a few months ago?

First of all, for AppVeyor, *all* builds are in the same queue and we
are only allowed to run one job in parallel. Currently, AppVeyor is
the obvious bottleneck of our workflow. It directly limits the number
of commits that we can merge per day.

Hopefully, thanks to Mariatta, backports are now merged as soon as CIs
pass if a core developer approved the PR. It's no longer needed to
watch the CIs to know when the Merge button will be enabed!

I looked more closely at AppVeyor last week since it took up to 2
hours to run tests on my CI, preventing me to merge it.

(Q1) AppVeyor runs again tests once the PR is merge. Was it the point
of running tests in each branch since... nobody look at test results
of these builds? Would it be possible to disable these jobs?

(Q2) AppVeyor runs tests twice on 3.6: VS 2015 and VS2017. It *seems*
(I'm not sure) that there are tested... sequentially. Each job takes
12 min so the who build takes 24 min... It seems like the official
binary on python.org is built on VS 2015: so maybe we could only keep
VS 2015 for the blocking CI, and maybe add a new buildbot (if there is
not already one) for VS 2017?

(Q3) How can we be allowed to run more jobs in parallel? Would it be free?

Zachary Ware wrote that he disabled the "largefile" resources of
Python test suite, so tests should now run faster.

Last week I found the cool build history:
https://ci.appveyor.com/project/python/cpython/history

It helped me to find some tests which fail randomly. I noticed that
they fail much more frequently than what I expected. Most developers
just schedule a new build.

Victor