FIX: Do not use Circle's premium layer caching

[oesteban]

Seems master is broken for that reason.

[djarecka]

@oesteban - circle ci required updates, but I'm pretty sure it doesn't solve the problem that some notebooks fail with new nipype

[oesteban]

Well, the first step was fixing the CI 👍

[oesteban]

Also, compared to old builds in master - it seems tests now go substantially slower. Could that be the cause?

[djarecka]

sure!
If you are interested, I can share with you what I found while testing it locally.

here are also some older tests from circleci: https://github.com/djarecka/nipype_tutorial/compare/master...djarecka:effigies/test1.5?expand=1

[djarecka]

@oesteban - well, the test_3 get stuck

[djarecka]

@oesteban - I was chatting with @effigies about it, I can send you some smaller versions of notebooks - one that passes, the other that fails (get stuck)

[oesteban]

Okay, master is already broken - let's merge this one and at least we'll be able to compare. It also adds a nightly build, so we notice when things break earlier.

Let's try to find the problem through some other PR.

[djarecka]

@oesteban - I've send an email with examples, but just to update conversation here, I believe this is the cell. And interestingly, when I remove the part “Masks and Probability maps” part it can work fine

[djarecka]

@oesteban - while you are at this, I was thinking that the way we handle Dockerfile is not the best, basically we have generate.sh bash script and Dockerfile in the repo. Should we remove Dockerfile and just generate it before running the test?

Having Dockerfile might be nice for the users, but we can add instruction how to create one

[oesteban]

@oesteban - I've send an email with examples, but just to update conversation here, I believe this is the cell. And interestingly, when I remove the part “Masks and Probability maps” part it can work fine

That could possibly be a memory issue. If we are hitting swap at points then it makes sense tests are, in general, suffering from a slow down.

We should try to pin nipype to the latest release after the previous successful build 8 months ago and see when we introduced the problem.

[oesteban]

@oesteban - while you are at this, I was thinking that the way we handle Dockerfile is not the best, basically we have generate.sh bash script and Dockerfile in the repo. Should we remove Dockerfile and just generate it before running the test?

Having Dockerfile might be nice for the users, but we can add instruction how to create one

Sounds good, but for a different PR

[djarecka]

I'm not sure how this could be a memory issue. The workflow shouldn't rerun the nodes at first place. The test_3 is not just slower, it doesn't finish

you can try to run the notebooks I sent you