python3.6 test_ssl fails reliably on systems under load
Affects | Status | Importance | Assigned to | Milestone | |
---|---|---|---|---|---|
Auto Package Testing |
Incomplete
|
Undecided
|
Unassigned | ||
python3.6 (Ubuntu) |
New
|
Undecided
|
Unassigned |
Bug Description
$ autopkgtest-
put system under load; e.g. pull-lp-source boost1.67; cd boost1.67; DEB_BUILD_
$ autopkgtest -s --shell --apt-pocket=
test_ssl fails
To rerun test_ssl alone, one can do:
$ python3.6 -W default -bb -E -R -m test -j 1 -w -uall,-
in a racy manner, for many test cases, due to ConnectionRefused from the TestEchoServer
One way to solve this is to reduce the load, cause without load TestEchoServer keeps up just fine.
Symptoms are s.connect failing with ConnectionRefus
I'm not sure it's worth any time fixing this test-server implementation, as clearly it is testing the ssl server/client bits, that work correctly on a normal, not-under-stress systems. And thus the ssl bits are correctly validated.
I will try to create a reproducer which does not involve VMs and stressing host.
meanwhile it would be nice to run python tests on slightly bigger machines, e.g. mark it as big_packages.
It's true that ScalingStack hosts are slower now, but I'm not convinced that running python's tests on a bigger VM is a solution. This to me feels like the tests are buggy and/or racy, and they should probably better be fixed to deal with running in resource- constrained environments - you'd want to know that Python works well there anyway.