Bug fix: Validate cached julia binary exists and check PATH resolution

[krynju]

I noticed this in CI. It's related to some problems with juliaup being installed on the workers and being messed with by install-juliaup (but that's a different issue)

/home/actions-runner-12/.juliaup/bin/julia --version
Error: The Julia launcher failed to figure out which juliaup channel to use.
Error: The process '/home/actions-runner-12/.juliaup/bin/julia' failed with exit code 1

setup-julia failed here at the julia --version check, because path resolved to a faulty local juliaup installation. This makes me think the cached binary might not have been fully there (just a .complete marker present).

Changes

1. Added bin/julia exists check which causes a cache miss on failure prompting a new download

.complete	bin/julia exists	Before	After
absent	—	miss	miss
present	absent	hit	miss
present	present	hit	hit

2. Added a PATH resolution check (warning)

We're basically checking if the selected Julia version/binary is the one that would resolve from PATH. I'm not sure if this should be fatal, so leaving it as a warning in case such an issue appears again.

[DilumAluthge]

Can you put the changes that modify lib/* and dist/* into a separate commit?

[DilumAluthge]

Is there a way to implement this without try-catch?

[krynju]

Done

[DilumAluthge]

To clarify, in this PR, you're not actually running the Julia binary to make sure it's runnable, you're just verifying that the Julia binary exists as a file on disk in the expected place?

So, if I understand correctly:

Before this PR:

1. .complete is absent - we correctly conclude "cache miss".
2. .complete is present but bin/julia is absent - we incorrectly conclude "cache hit".
3. .complete is present and bin/julia is present, but if you tried to run bin/julia it would error - we conclude "cache hit".

After this PR:

1. .complete is absent - we correctly conclude "cache miss".
2. .complete is present but bin/julia is absent - we now correctly conclude "cache miss".
3. .complete is present and bin/julia is present, but if you tried to run bin/julia it would error - we conclude "cache hit".

Have I described things correctly?

I'm totally fine if that's the case¹, I just would like to make the PR description (and the code comments) a little more precise, to better describe what we're actually checking.

Footnotes

1. Actually running the Julia binary would eat up time that we probably shouldn't be spending. ↩

[krynju]

To clarify, in this PR, you're not actually running the Julia binary to make sure it's runnable, you're just verifying that the Julia binary exists as a file on disk in the expected place?

So, if I understand correctly:

Before this PR:

1. .complete is absent - we correctly conclude "cache miss".
2. .complete is present but bin/julia is absent - we incorrectly conclude "cache hit".
3. .complete is present and bin/julia is present, but if you tried to run bin/julia it would error - we conclude "cache hit".

After this PR:

1. .complete is absent - we correctly conclude "cache miss".
2. .complete is present but bin/julia is absent - we now correctly conclude "cache miss".
3. .complete is present and bin/julia is present, but if you tried to run bin/julia it would error - we conclude "cache hit".

Have I described things correctly?

I'm totally fine if that's the case1, I just would like to make the PR description (and the code comments) a little more precise, to better describe what we're actually checking.

Footnotes

1. Actually running the Julia binary would eat up time that we probably shouldn't be spending. ↩

I updated the PR description.

Also removed the part which checks if the julia binary is runnable.
Now the bin/julia check causes a cache miss if the binary doesn't exist. There's an additional PATH resolution check that ends up in a warning.

Also split the commit

[DilumAluthge]

Thank you @krynju!