Top-level skip not working anymore?

[h-vetinari]

Did you change something recently w.r.t. how skips are computed?

For torchaudio, we had

build:
  number: ${{ build_number }}
  skip:
    - win

in the recipe, which completely skipped all builds, until conda-forge/torchaudio-feedstock#43 came along, and suddenly windows jobs got added.

I had to add (conda-forge/torchaudio-feedstock@1dbc3ce)

@@ -157,6 +157,7 @@ outputs:
       name: torchaudio-tests
     build:
       skip:
+        - win
         - cuda_compiler_version != "None"

to another output to get rid of the windows jobs again. I'm hoping this is considered a bug, because re-applying top-level skips to every single outputs is a major annoyance in conda-build that I was hoping to put behind me with rattler-build 😅

[h-vetinari]

@wolfv @Hofer-Julian, by any chance, will this topic be addressed by the current refactor? I'm happy to provide test cases. It just happened again that skipping some variants for a single output caused more jobs to be created, because suddenly all other outputs were not covered by the global skip anymore. This is really annoying behaviour 😑

[Hofer-Julian]

Test cases would be much appreciated. We just merged the refactor, so would be good to check if it fixes the problems you encounter

[h-vetinari]

Test case would be the state of conda-forge/tinygrad-feedstock@1878f08; compared to its parent, it adds an output-level skip

   - package:
       name: tinygrad-tests
     build:
+      skip:
+        # opengpu server has sm_70 card that's not supported by CUDA 13 anymore
+        - match(cuda_compiler_version, ">=13.0")
+        # missing onnxruntime
+        - match(python, ">=3.14")

to the already existing global skip

  skip:
    # CUDA builds on GPU agents, non-CUDA builds on CPU agents; skip could be avoided
    # using https://github.com/conda-forge/conda-forge-pinning-feedstock/pull/6910
    - cuda_compiler_version == "None" and "gpu" in github_actions_labels
    - cuda_compiler_version != "None" and "cpu" in github_actions_labels
    - match(python, "<3.11")

However, by adding the output-level skip, the global condition match(python, "<3.11") does not apply anymore (which is what this bug is about), and thus smithy generates jobs for python 3.10 again. I then have to do stuff like conda-forge/tinygrad-feedstock@a7badeb to get things to work again. It's even more painful when it's only for quick debugging (e.g. conda-forge/tinygrad-feedstock@b0d12fd), which now needs to add some debug skip to all outputs.

I struggle to imagine how this could be intentional (aside from the fact that conda-build behaves like this, but that's broken due to the way how package: vs. outputs: grew historically, something that rattler-build is finally improving upon).

In short: the global skips (recipe-level build.skip) always need to be inherited by all outputs. As noted in the OP, this already worked previously in torchaudio (and elsewhere AFAIR), but something seems to have regressed.