Use catalina for some osx jobs #3591
Conversation
Right now we're only testing in CI that things work on macOS high sierra but catalina has been out for a while and different osx versions differ enough that it warrants testing that the new version works too. This commit adds a new catalina job that runs only python 3.7 to verify that things work on newer osx.
|
I'm actually thinking it might make more sense to just bump the existing osx jobs to mojave since this worked and is not a brand new release. I think the issues I was looking at before when we first migrated to azure must resolved now. |
With the addition of python 3.8 jobs we're running at our job concurrency limit. To avoid making the run time of ci too long by having to wait on every commit this commit switches the second stage osx jobs to use mojave instead of adding duplicated jobs on multiple osx versions. We'll still use high sierra for the first stage's 3.7 testing to maintain coverage there.
|
It looks like this is failing because of: #2793 (which is why we ended up using high sierra in CI originally) There's not much we can do since the only consistent factor involved is osx 10.14 which we don't have any ability to influence and/or fix. |
|
Conda install the needed numpy and scipy will work. They use mkl |
|
The priority for this just went up, Azure Pipelines is going to stop offering the high sierra images starting March 23, 2020[1]. Which means we'll have to come up with a solution for running tests reliably on 10.14. I guess I'll have to pivot things to use conda instead of straight cpython and pip. [1] https://devblogs.microsoft.com/devops/removing-older-images-in-azure-pipelines-hosted-pools/ |
Catalina images have been available on azure pipelines for a few days now. This commit pivots to using those instead of 10.14 mojave which have issues with segfaults because of system libraries.
|
It looks like 10.15 is segfaulting too: https://dev.azure.com/qiskit-ci/qiskit-terra/_build/results?buildId=8099&view=logs&j=377f2c8d-8b6f-583f-e152-2ff025e2543a&t=28b3ff0d-9e76-5e94-89e9-eed7321a383e&l=4178 I guess I'll still have to switch it to conda with catalina |
mtreinish
force-pushed
the
add-catalina-job
branch
3 times, most recently
from
0c98245
to
1f3ba87
Compare
The pip installable numpy is causing segfaults on catalina and mojave environments. To workaround this issue this commit switches to using a conda install for the macos jobs only.
azure-pipelines.yml
Outdated
| @@ -234,8 +234,17 @@ stages: | |||
| inputs: | |||
| versionSpec: '$(python.version)' | |||
| displayName: 'Use Python $(python.version)' | |||
| - bash: echo "##vso[task.prependpath]$CONDA/bin" | |||
| displayName: Add conda to PATH | |||
There was a problem hiding this comment.
This looks like it just echos the conda path. Maybe it doesn't need to be added to PATH?
There was a problem hiding this comment.
|
It looks like this segfaulted again even with conda: https://dev.azure.com/qiskit-ci/qiskit-terra/_build/results?buildId=8866&view=logs&j=ce9a6331-8618-5e5b-ad8c-d3ab6621690a&t=b587a450-7392-5f12-37e1-082303ec0efb&l=4056 marking this as on hold until we figure this out |
10.15 fails in the same way with or without conda, we've tried 10.14 without conda but not with it. This commit switches back to 10.14 now that we've switched to using conda to see if the tests pass.
There is an OpenBlas threading issue affecting scipy.linalg.eigh that is holding up #3591. This removes the call to eigh and replaces it with eig that uses a different underlying blas call and just takes the real part of the resulting eigenvalues. This also made explicit two of the tests that were causing issues as they were off by a minus sign in the test suite but were fine when run explicitly. * fix threading crash * fix tests to be explicit
Now that we have a fix for the random segfaulting in the quantum info tests there shouldn't be a reason we need to keep running in conda. At the same time this switches back to 10.15 to test the newest version of macOS. This only reverted back to 10.14 briefly to see if the segfault was isolated to 10.15, which it was not.
|
Thanks to @nonhermitian this is now working on all versions of python with or without conda. This is ready to go whenever now |
There is an OpenBlas threading issue affecting scipy.linalg.eigh that is holding up Qiskit#3591. This removes the call to eigh and replaces it with eig that uses a different underlying blas call and just takes the real part of the resulting eigenvalues. This also made explicit two of the tests that were causing issues as they were off by a minus sign in the test suite but were fine when run explicitly. * fix threading crash * fix tests to be explicit
Right now we're only testing in CI that things work on macOS high sierra but mojave and catalina have been out for a while and different osx versions differ enough that it warrants testing the newest version works. This commit switches the osx jobs in the to use catalina. Additionally, azure pipelines is removing support for the high sierra environment which means we'll have to migrate to a newer version anyway. * Add catalina job Right now we're only testing in CI that things work on macOS high sierra but catalina has been out for a while and different osx versions differ enough that it warrants testing that the new version works too. This commit adds a new catalina job that runs only python 3.7 to verify that things work on newer osx. * Update azure-pipelines.yml Co-Authored-By: Kevin Krsulich <[email protected]> * Use Mojave for all second stage osx jobs With the addition of python 3.8 jobs we're running at our job concurrency limit. To avoid making the run time of ci too long by having to wait on every commit this commit switches the second stage osx jobs to use mojave instead of adding duplicated jobs on multiple osx versions. We'll still use high sierra for the first stage's 3.7 testing to maintain coverage there. * Pivot to macos 10.15 catalina Catalina images have been available on azure pipelines for a few days now. This commit pivots to using those instead of 10.14 mojave which have issues with segfaults because of system libraries. * Use conda for macOS jobs The pip installable numpy is causing segfaults on catalina and mojave environments. To workaround this issue this commit switches to using a conda install for the macos jobs only. * Install scipy via conda too * Switch back to 10.14 again 10.15 fails in the same way with or without conda, we've tried 10.14 without conda but not with it. This commit switches back to 10.14 now that we've switched to using conda to see if the tests pass. * Switch back to 10.15 without conda Now that we have a fix for the random segfaulting in the quantum info tests there shouldn't be a reason we need to keep running in conda. At the same time this switches back to 10.15 to test the newest version of macOS. This only reverted back to 10.14 briefly to see if the segfault was isolated to 10.15, which it was not. Co-authored-by: Kevin Krsulich <[email protected]>
Summary
Right now we're only testing in CI that things work on macOS high sierra
but mojave and catalina have been out for a while and different osx versions
differ enough that it warrants testing the newest version works. This
commit switches the osx jobs in the to use catalina. Additionally, azure pipelines
is removing support for the high sierra environment which means we'll have
to migrate to a newer version anyway.
Details and comments