Use the appropriate config dir for the registry

[samueldr]

It was reported on IRC that Nix would create a .config dir even though XDG_CONFIG_HOME was set.

This is the only occurence of .config I could grep for that wasn't in getConfigDir().

The test is a bit... ugly... And might be missing some other actions that could create the config dir...

But otherwise, it does test this particular case. Reverting the fix (tip of this branch) makes it fail as expected.

I guess we want to test other commands that can write to the config directory, are there others?

We probably also want to test the equivalent for the different caches.

[samueldr]

Yes, this export is now redundant considering the previous export.

The following tests, though, are not redundant. None are.

[edolstra]

We shouldn't use fixed paths like /tmp/home in the tests since that's racy and breaks in multi-user environments. I see that recently got introduced in 895516c but it's a mistake.

[samueldr]

Totally agree that it may be racy. I even changed it to reduce raciness assuming (verified through grepping) that the path is unique between all tests.

Though I see it's not foolproof. What strategy should be used to isolate each test files?

[samueldr]

Are we blocking this on the fixed paths?

If so, I'll need a suggestion to make this go forward. I'm not sure what approach could be used in these tests.

If we're not blocking on the fixed paths, anything else needs to be changed?

[thufschmitt]

What strategy should be used to isolate each test files?

I guess using $TEST_ROOT is good-enough (it's doesn't really bring much more guaranty if the test is run twice concurrently, but it's what everything else uses, so at least there's only one place to change if must be)

[samueldr]

Applied the fix to use $TEST_ROOT.

• Tested the new test still fails without the fix (the last commit)
• Tested the tests work with the fix

And how? nix-build.

[edolstra]

Thanks @samueldr!