Use Leafwing-Studios/cargo-cache in CI

[BD103]

Objective

• Our CI caches the ~/.cargo and target folders, to take advantage of faster, incremental, compilation.
• The use of caches is quite inconsistent, with different formats for keys. It also does not take advantage of cache fallback (using an old cache if the current one does not exist).
• Caches are being created in the merge queue, but can never be used.
  • To verify this, go to any validation job and check the actions/cache@v4 step. It will always say Cache not found for input keys: ....

Solution

• Switch to using Leafwing-Studios/cargo-cache, which automatically handles cache keys, fallback, dependency updates, and toolchain updates.
• Caches are no longer used in the merge queue.
  • Github Actions has something called cache isolation. Long story short, caches cannot be used if they were not created from the same branch or base branch.
  • For instance, a cache created on main can be used by any feature branch, but a cache created by feature-a branch cannot be used by the unrelated feature-b branch.

---

Changelog

• All CI caching is now done with Leafwing-Studios/cargo-cache.
• The CI no longer creates inaccessible caches for validation jobs.

[TimJentzsch]

You might be interested in this: https://github.com/Leafwing-Studios/cargo-cache

It uses a the hash of the updated lock file instead of the current date to calculate the hash key, also with several levels of fallbacks.

Honestly, debugging caching actions and figuring out why cargo keeps recompiling dependencies is super annoying, I hope that you'll get it to work!

[BD103]

You might be interested in this: https://github.com/Leafwing-Studios/cargo-cache

Thank you! I'll definitely take inspiration from that. I chose not to hash the lockfile because I figured that the amount of dependency change in 1 day cannot be so great that it would hurt performance, but that assumption will definitely require testing. It's good to see an alternate approach, though!

[BD103]

So far things are working well. I'm having some issues with the cache not being found, so I may have to move my testing over to my fork, where I can actually delete caches and cancel runs.

Building on MacOS and docs are the two longest jobs. I'm going to see if I can shorten their duration, if possible.

[BD103]

Following @TimJentzsch's interesting link, I've decided to switch this PR to use Leafwing-Studios/cargo-cache for everything. I think it's a lot more maintainable than my custom solution (and it's easy to use, too!).

Performance seems roughly the same, except for run-examples-macos-metal which didn't use caching before. It's hard to tell, since Github has two different MacOS runners with 3 and 4 CPU cores.

[TimJentzsch]

Previously attempted in #7627, but the action has also improved since then

[BD103]

Previously attempted in #7627, but the action has also improved since then

Is there a reason the original PR was closed?

[BD103]

If / when this gets merged, it would be good to wipe all of the existing caches here. This will make the migration a little quicker, and hopefully show the performance improvements sooner.

[alice-i-cecile]

Swapping to Controversial; I want to get Francois's sign-off here.

[janhohenheim]

Update to the newly released v2 before merging.

[TimJentzsch]

I think I would prefer an approach where we create cache only from the main branch, those should be available to every job

We can now do this via the save-if option, like this:

- uses: Leafwing-Studios/cargo-cache@v2
  with:
    save-if: ${{ github.ref == format('refs/heads/{0}', github.event.repository.default_branch) }}

[BD103]

Done!

@mockersf, could you please review this? You're the last step before this can be merged :)

[mockersf]

@mockersf, could you please review this? You're the last step before this can be merged :)

I was waiting for my previous review to be addressed! Now unless there's an urgent need I'll do it after the Jam 🙂