Don't run travis for documentation-only changes

[zpao]

Or better, only run for changes in specific paths. This is the first step in the effort to reduce the number of jobs we use on travis ci. cc @jeffmo

[RReverser]

Travis itself doesn't currently allow this, you can only start job and immediately quite (i.e., before_script) if list of files changed in latest commit doesn't contain real code. Not sure if this helps much though.

Other two options that do not require changes to tooling could be:

1. Add [skip ci] to commit messages of all the doc-only commits.
2. Split code and docs into different branches and use Travis "branch only" feature.

[zpao]

Yea, I know Travis doesn't allow this natively (yet?). I've been doing the [skip ci] bit when I remember but it's tough to tell people submitting PRs that when I can't even remember. I was thinking of having a script that looked at the diff range and got the modified paths. If it wasn't src/ or vendor/, then exit early, otherwise continue. It wouldn't really only run for certain changes but it would be a quick noop. It sounds like you were thinking the same with before_script. If we can check the range and not just the latest commit, I think it would be really helpful.

I'm trying to avoid splitting code away from docs. I'll argue until I die that there's a lot of value keeping them as close as possible :) It is an option though, so thanks for bringing it up.

[RReverser]

It sounds like you were thinking the same with before_script.

Yeah, it was pretty much the same idea (however, task will be still created and queued, so it will still use some pool quota, but will release it faster).

If we can check the range and not just the latest commit, I think it would be really helpful.

Yes, we actually can. I did that somewhere for one of my pet projects, let me check and send PR with fix if you're fine to go with proposed solution.

[zpao]

Yea, I think that would be awesome!

[RReverser]

@zpao Fixed in #2000, please check it out (I've tested on own fork in different combinations, seems to work).

[RReverser]

Also, #927 seems to be particular case of this.

[zpao]

Hah, I knew I had filed this already. I'll point that one here.