fix: keep stderr and stdout from interleaving

[henryiii]

I think this fixes #587.

[henryiii]

Looks like circleCI is still getting the pre-fix test/test_docker_images.py . Odd.

[YannickJadoul]

Looks like circleCI is still getting the pre-fix test/test_docker_images.py . Odd.

We've seen CircleCI act weirdly (or at least different from others) on these kinds of things, before, no? It also didn't restart automatically after closing-reopening.

[YannickJadoul]

Is there a way to check this? If so, this can be merged without further input, I'd say :-)

[joerick]

I have seen something similar to this before when I was doing the fold groups work, at the time I thought it was a Github actions bug... I might do a little testing to investigate...

[joerick]

I don't believe this PR fixes the problem. See this GHA run:
https://github.com/joerick/gha-testing/runs/1888060360?check_suite_focus=true

These examples try to print interleved stdout/stderr messages. So perfect output would look like:

1
stderr 1
2
stderr 2
3
stderr 3
4
stderr 4
...

As you can see, regardless of the ways I try to disable output buffering, there are still occasionally problems.

It seems to me that this is an inherent problem within GHA. Or perhaps there's a buffer somewhere that we haven't found the controls for yet?

[henryiii]

https://github.saobby.my.eu.orgmunity/t/stdout-stderr-output-not-in-correct-order-in-logs/16335

Maybe we could redirect stderr to stdout on GHA? Is there really a difference in GHA? (In fact, is there really a difference in any logging system?)

[joerick]

Yeah, we could do that. I actually don't think there's any difference between stdout and stderr on CI. I had implemented something like this when I was working on the fold group stuff. The only snag is that some of our tests understand the difference between out/err. But they could be worked around fairly easily.

btw, in our case, we also have to redirect the subprocesses stderr to stdout.

[joerick]

Just doing some tidying of PRs, so closing this one.

A simple fix for this would be to add 2>&1 to our sample configs. But that might just cause confusion. Otherwise, it's a refactoring to direct our logs to only stdout, perhaps just when CI=1 - and figure out how to manage this in the print-build-identifier tests