Failed tests leak CSI resources

[timoreimann]

The vast majority (all?) of the tests in csi-test do not seem to clean up resources properly when an assertion failure occurs. I'm not too familiar with Gomega, but my understanding is that once an EXPECT fails, the test execution comes to a halt immediately, so none of the cleanup logic usually done by the end of a test (e.g., deleting volumes or snapshots) gets a chance to execute.

This is a problem because the created CSI resources leaking this way may affect other tests to the extent that they may now fail as well even though they otherwise wouldn't. Specifically, a test that does not create any fixtures would now have left-over objects from the failed tests and thus fail too, aggravating identification of why a test run fails.

It'd be good to come up with a means to reliably clean up resources regardless of what the outcome of a test (and any of its assertions) is.

[msau42]

/kind bug
/help

[k8s-ci-robot]

@msau42:
This request has been marked as needing help from a contributor.

Please ensure the request meets the requirements listed here.

If this request no longer meets these requirements, the label can be removed
by commenting with the /remove-help command.

Details

In response to this:

/kind bug
/help

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes/test-infra repository.

[timoreimann]

@pohly you mentioned on Slack that there is some kind of cleanup taking place already. Looking forward to hearing about it.

[pohly]

That code is in https://github.com/kubernetes-csi/csi-test/blob/master/pkg/sanity/cleanup.go

Whenever you see a call to MaybeRegisterVolume you can be sure that the volume is going to be deleted even if the test code itself doesn't do it.

[pohly]

Have you encountered a specific test where the cleanup didn't work?

[timoreimann]

@pohly I had experienced the problem when I was running ListSnapshots-related tests a few days ago. Oddly, I can't reproduce the issue anymore locally right now, although it seems as if it should be fairly easy to do so: I don't see a register functionality for snapshots, so presumably if snapshots aren't cleaned up, they should possibly affect other tests?
Example: test A creates a snapshot but fails its assertion so the cleanup step is skipped. If test B is executed afterwards, doesn't create any snapshots, and expects ListSnapshots to return an empty list, it would fail now because it still has the snapshot from test A stored.

What am I missing? Is this possibly as mock driver vs. hostpath driver issue?

[pohly]

I don't see a register functionality for snapshots, so presumably if snapshots aren't cleaned up, they should possibly affect other tests?

Yes. Looks like those tests were written without taking into account how cleanup is supposed to work. So this issue is valid, we just need to be more specific about which tests have to be updated.

[timoreimann]

@pohly at a quick glance, it looks like all *Snapshot* tests (e.g., ListSnapshots, CreateSnapshot) are affected to the extent that create volumes are not registered and snapshots cannot be registered.

Other tests may be affected as well. I spotted the CreateAndControllerPublishVolume helper function that creates a volume but does not register it.

AFAICS, there are two TODOs:

1. Ensure all existing code registers volumes properly.
2. Create a cleanup manager for snapshots and use it.

Going further, I wonder if we should extend the (test) API so that test authors are less susceptible to missing resource registrations. For instance, we could provide a convenience function that creates a resource and registers it for cleanup in one go. This may also allow us to get rid of MaybeRegisterVolume because the convenience function would encapsulate the check for error being non-nil and act accordingly. What do you think?

Feel free to assign me to this one. I should have some bandwidth by the end of this week to work on whatever we determine is the most desirable approach, and update all volume/snapshot usages in the test code systematically.

[pohly]

Going further, I wonder if we should extend the (test) API so that test authors are less susceptible to missing resource registrations. For instance, we could provide a convenience function that creates a resource and registers it for cleanup in one go.

That sounds useful.

Feel free to assign me to this one. I should have some bandwidth by the end of this week to work on whatever we determine is the most desirable approach, and update all volume/snapshot usages in the test code systematically.

Thanks! I'm not sure whether I can assign it to you, but we'll see:

/assign @timoreimann

[NicolasT]

Running into a similar issue: I have a CSI plugin that's far from complete/spec-compliant, so many tests like Ensure VolumeId is not provided are still failing.

Now, when such test fails, it appears /tmp/csi-mount gets created but not cleaned up. Any subsequent test which attempts to create the very same directory then fails (file exists).