initial proof-of-concept in a container

[mroth]

This is a potential alternate way to look at distribution of a dev tool with dependencies using Docker, versus the current vendoring/packaging methodology.

Instead of trying to vendor/bundle all the python/node dependency files in a way that will work on most systems, just build a self-contained docker image. This frees us from having to worry at all about if/what versions of Python/NodeJS are on the machine.

When the linter is updated, new images would be uploaded to Docker Hub (using that vs a private GCR since this is an open source project), end-users can docker pull khanacademy/khan-linter:latest (or a specific version/tag)

This also would provide an alternate way to run on a server without worrying conflicting dependencies between this and the app it is linting.

Note I haven't actually pushed the built image to Docker Hub yet, will do so if/after this is merged.

Future enhancements

For convenience of tagging docker images, it would be good to be more consistent with updating semantic version of releases, and then have the make image task here add the version number, to the image tag, so people could more easily reason about what version they have on their system, especially when doing a docker images list or some such.

What would it mean to fully embrace this?

In a mystical future world if the docker image was to become the "one true way" to use this, then there are some cleanup tasks that would probably become worth looking at:

• the py2 vs py3 branching and vendoring logic could be removed entirely.
• the update_check can be removed entirely in favor of pulling the latest docker image (it currently won't fire within the container because its not in a git repo, but the code could be removed for cleanliness then.)
• the big vendored directories can be removed from this git repo, as with the associated Make tasks.

[cjfuller]

Can you elaborate a little more on how the py2 vs. py3 thing would be solved? I would think we would need a separate container for each. (The key problem here is not the version under which it has to run, but the version it has to lint. As far as I know you have to lint code of the same version of python you're running?)

[mroth]

Can you elaborate a little more on how the py2 vs. py3 thing would be solved? I would think we would need a separate container for each. (The key problem here is not the version under which it has to run, but the version it has to lint. As far as I know you have to lint code of the same version of python you're running?)

Oh interesting, I had assumed that was just there for handle running under different environments, didn't realize our lint rules required running under the same version. That would obviously require some changes if true.

[cjfuller]

Yeah, as far as I know, the off-the-shelf python linters do version-matched linting. So, e.g., to lint our python3.6 code, we'd need to run under python3.6.

[mroth]

Okay, that will require a different approach then. While we could have separate images for py2 vs py3, my assumption is it would probably be most convenient for users to have a single image that has both dependencies installed and could use the proper one via a flag or some such?

[carterjbastian]

Wow! Very cool! This is a pretty sweet proof of concept that seems like it would make things significantly easier. I haven't worked on the linter, so I'm curious to hear what other changes we'd need to consider before committing to this way of doing things.

[csilvers]

(I commented in phabricator, but repeating here, ugh so confusing):

Cool! I am learning a lot already.

How long does it take to run runlint over a single file in the docker world as compared to directly? I ask because the git hooks/etc often run lint over jusy one or two files, if that's what's in your commit.

[csilvers]

Clearing out my github queue...