CircleCI Optimization: skip costly nodes

[oesteban]

The two main bottlenecks on the CI workflow causing a slow development iteration are:

• Caching docker images (where there's little to do)
• Long-running nodes (mainly based on antsRegistration).

The idea of this request is to implement some mechanism that allows injecting pre-calculated transforms to antsRegistration-based nodes. We could maybe trigger the mechanism after the _run_interface has been called (to exercise all the inputs checking, etc.), then patch in the results and finally let the _run_interface run as if it had run the code.

That would shave off a great deal of the runtime.
EDIT: if this worked for sMRIPrep, it would be reasonable to upstream it to Nipype and downstream it to fmriprep.

[oesteban]

@satra can you drop some of your wisdom?

[satra]

i would use datalad/or some such to pre-populate the working directory and make sure to use content hash. if the working directory structure changes this wouldn't work.

almost seems like this is a test utility we could introduce in nipype.

[oesteban]

Re: content caching - it is on.

almost seems like this is a test utility we could introduce in nipype.

Yep. I thought sMRIPrep is a great testing ground before we upstream it to nipype.