Cache build is rerun un-necessarily for experimental multi-output recipe

[carterbox]

I am testing the new multi-output cache with the libmagma recipe. This recipe has three outputs: one development and two runtime outputs (sparse linalg and non-sparse linalg are separate libraries). I have one cache build at top, and I want to split this cache build into the three outputs. However, the cache build is run twice instead of once.

At first I thought this might be because libmagma_sparse depends on libmagma, but the libmagma-devel also depends on other outputs and the cache isn't rebuilt before creating that output.

From the logs it appears that the cache build for libmagma and libmagma-devel have the same cache build hash, and libmagma_sparse gets a cache build with a different hash.

The converted recipe is available here: https://github.com/carterbox/libmagma-feedstock/tree/debug-rattler

The logs are attached as a text file.

build-logs.txt

rattler-build build --experimental -m .ci_support/linux_64_c_stdlib_version2.17cuda_compiler_version12.9.yaml -r recipe/recipe.yaml &> build-logs.txt

rattler-build 0.51.0

[wolfv]

Thanks for trying and opening this issue. Will have to check. The cache-key should indeed point to the folder and be the thing that is used to invalidate the cache hit/miss.