@charris : Just a signpost for future discussion. Started in #8071 but should continue here.

@charris : Any updates / comments on this one? This fix does seem straightforward IMHO.

The actual change makes sense to me. I found the refactoring a bit confusing, though, because of how context is/isn't propagated between different shell scripts -- so e.g. there's this pushd/popd stuff copied into the new before-install script, but it isn't really needed, because working directory changes don't propagate beyond single script, and then there's the need to remember to re-activate the virtualenv each time.

I think it would be simplest to just merge the stuff that's currently in before_install: directly into travis-test.sh, so we just have one script that does everything.

@njsmith : Makes sense to me. I made a separate commit to test what happens on Travis.

Travis failed, not sure why.

@njsmith : I forgot to copy over the pip installs. 😄

Changes look good to me, but Travis is still failing. Maybe send a ping when you've got it passing?

@njsmith : Will do. In the meantime, do you think you could cancel #170595892? It's an old build. Thanks!

@njsmith : So the chroot build is failing because some of the commands from before-travis-install.sh cannot be run, as can be seen here. Unless we want to have at it with linux32, I am more inclined to keep these original commands separate. Thoughts?

That failure can be fixed by just telling the shell that we're okay with the information gathering steps at the beginning failing... something like

set -x
# It's okay if these fail, and when we run in a chroot they do
uname -a
free -m
ulimit -a
# From here on, we want to know if anything fails
set -e
mkdir builds
...

A bigger question is whether it's useful to be running the virtualenv stuff inside a chroot. Potentially the answer is yes, this is actually better... but the chroot handling is confusing enough that I'm not 100% sure :-/. I guess we could make the simple change above and see what happens.

@njsmith : Didn't work either. TBH, I would rather push the patch in first, as I don't see any easy way right now to integrate the two together since the chroot seems to mess everything up that before-travis-install.sh does.

Ugh. Well, I guess this is so confusing that we're not making it any worse. (What could possibly go wrong.)

Thanks for sticking with it...