Abstract away base Docker images without C-PAC

[shnizzedy]

Building a Docker image takes awhile (~½ hour on CircleCI usually). Building AFNI from source instead of from the latest binaries (#1454) nearly doubles that.

Most of the time, all we're really changing is in this little bit at the end:

C-PAC/Dockerfile

Lines 216 to 226 in 55ae58c

	COPY dev/docker_data/default_pipeline.yml /cpac_resources/default_pipeline.yml
	COPY dev/circleci_data/pipe-test_ci.yml /cpac_resources/pipe-test_ci.yml
	COPY . /code
	RUN pip install -e /code
	COPY dev/docker_data /code/docker_data
	RUN mv /code/docker_data/* /code && rm -Rf /code/docker_data && chmod +x /code/run.py
	ENTRYPOINT ["/code/run.py"]

I think we could create a base.Dockerfile that includes everything in our current Dockerfile except the above, build it with a GitHub Action only when that file changes and store it here, and have the regular Dockerfile build from that image instead of

C-PAC/Dockerfile

Line 2 in 55ae58c

FROM neurodebian:bionic-non-free

I've done this in branches when trying particular versions of dependencies, like

C-PAC/Dockerfile

Lines 1 to 2 in c173336

	#using neurodebian (Bionic non-free)+ ANTs 2.3.4 as parent image
	FROM docker.pkg.github.com/shnizzedy/neurodebian-bionic-non-free-ants-2.3.4/neurodebian_bionic-non-free_ants-2.3.4:latest

but I think it would make sense to do this for all of the dependencies, to save that build time on Circle.

Related #788

[shnizzedy]

Also, I think we could create an image with regression test outputs, update that image with each release, be clever with docker-compose, and really streamline regression testing.

[sgiavasis]

😮 yes

[sgiavasis]

Is there something I can put together or clarify that can help this along?

[shnizzedy]

For the multistage building images, I have a pretty good idea of what to do. For the regression test container, I haven't thought it all the way out yet, so if you get to it before I do, you're welcome to take the reigns. It's one of those things that I expect will take some time to design and build and then save time once it exists

[sgiavasis]

Now that the node blocks are modular and we can more easily run them standalone, I am picturing running correlations for regression testing on each one individually - but still keeping some of the main preconfigs for integration testing instead.

Basically, this would remove the necessity for those big regtest-N configs I created that attempt to test almost everything. This would make things leaner.

[shnizzedy]

One-by-one regression tests will be fantastic! Way easier to spot and diagnose issues.