Strange behavior with multiple registry cache-from lines

When using multiple cache-from= lines, what is the expected behavior?

In a CI environment, we might want to have a separate cache for each feature branch. However, we want to mitigate the cold start problem by having a backup cache, usually from the truck (main) branch. So it might look like:

  --cache-from type=registry,ref=${{ env.REGISTRY }}:branch.${{ github.ref_name }}
  --cache-from type=registry,ref=${{ env.REGISTRY }}:main
  --cache-to type=registry,ref=${{ env.REGISTRY }}:branch.${{ github.ref_name }}

However, I've always observed this to behave inconsistently when i test this. Sometimes the build is blazing fast because nothing has changed on the branch, and the branch cache is obviously being used. Other times, the build appears to fall back to using the backup cache, despite nothing having changed in the branch. It actually appears as if the cache is selected at random, and Buildkit simply chooses which ever one it happened to pick.

When I remove the backup cache-from, the first build is slow due to the cold start problem, but subsequent builds are always blazingly fast.

I should add that i am actually using buildx and bake. Doing something like:

docker buildx bake \
  --set "target.cache-from type=registry,ref=${{ env.REGISTRY }}:branch.${{ github.ref_name }}" \
  --set "target.cache-from type=registry,ref=${{ env.REGISTRY }}:main" \
  ...

Perhaps my cache-from arguments are being reordered randomly...

Am I crazy?

Start by reproducing the issue with docker buildx bake and two registry cache-from references, using the branch and main cache examples in the report. Trace how multiple cache sources are ordered and selected; done means establishing whether selection is deterministic and documenting or fixing the observed behavior.