context variable cannot be an empty string

[sebp]

When using an empty string in a context variable, I get the error:

Error:   × Failed to parse recipe

Error:
  × expected a scalar value.
   ╭─[recipe/recipe.yaml:7:8]
 6 │   # sometimes used to match PyPI releases
 7 │   tag: ""
   ·        ─┬
   ·         ╰── here
 8 │
   ╰────
  help: `context` values must always be scalars (booleans, integers or
        strings) or uniform lists of scalars

The recipe has been adapted from https://github.com/conda-forge/scikit-learn-feedstock/pull/290

schema_version: 1

context:
  version: "1.7.2"
  # sometimes used to match PyPI releases
  tag: ""

package:
  name: scikit-learn
  version: ${{ version }}

source:
  url: https://github.com/scikit-learn/scikit-learn/archive/${{ version }}${{ tag }}.tar.gz
  sha256: 363ca6f51164fce2ece13b59370ca4ec670a1b73188e0eb387f6ea99c2e2bf67

build:
  number: 1
  skip: python_impl == "pypy"

requirements:
  build:
    - if: build_platform != target_platform
      then:
        - python
        - cross-python_${{ target_platform }}
        - meson-python >=0.16.0
        - ninja
        - cython >=3.0.10
        - numpy
        - scipy
    - ${{ stdlib('c') }}
    - ${{ compiler('c') }}
    - ${{ compiler('cxx') }}
    - if: osx
      then: llvm-openmp
    - if: linux
      then: libgomp
  host:
    - python
    - pip
    - python-build
    - meson-python >=0.16.0
    - ninja
    - cython >=3.0.10
    - numpy
    - scipy >=1.8
    - joblib >=1.2.0
    - threadpoolctl >=3.1.0
    # See https://github.com/conda-forge/openmp-feedstock/issues/126
    - if: osx
      then: llvm-openmp
    - if: linux
      then: libgomp
  run:
    - python
    - numpy >=1.22.0
    - scipy >=1.8.0
    - joblib >=1.2.0
    - threadpoolctl >=3.1.0

tests:
  - python:
      imports:
        - sklearn
  - requirements:
      run:
        - pytest >=7.1.2
        - cython >=3.0.10
        - pytest-xdist
        - pytest-timeout
    script:
      interpreter: python
      content:
        - import os
        - import pytest
        - 'tests_to_skip = "_not_a_real_test"'
        # https://github.com/scikit-learn/scikit-learn/issues/20335
        - 'tests_to_skip += " or test_loadings_converges"'
        # Numerically unstable test numerical difference in test
        - if: ppc64le
          then: 'tests_to_skip += " or test_mlp_regressor_dtypes_casting"'
        # due to emulation through QEMU, the tests for aarch run very long;
        # azure has different linux agents, and the weaker ones can be a factor 2 slower;
        # skip longest-running ones; all run >60sec (on fast agents) resp. >90sec (on slow agents);
        # since pytest -k skips on string matches, try to match accurately to avoid skipping more than intended
        - if: aarch64 and (env.get("CI", default="False") == "True")
          then: 'tests_to_skip += " or (test_quantile and (test_asymmetric_error
            or test_quantile_estimates_calibration or test_quantile_sample_weight))"
            + " or test_kernel_pca_solvers_equivalence or test_n_iter_no_change_inf"
            + " or test_iterative_imputer_no_missing or (test_voting and test_gridsearch)"
            + " or test_svc_ovr_tie_breaking or test_check_estimator_clones"
            + " or (test_estimators and QuantileRegressor)"
            + " or (test_pairwise and test_cosine_distances)"
            + " or test_randomized_eigsh_reconst_low_rank"
            + " or test_randomized_svd_power_iteration_normalizer"
            + " or test_SGDClassifier_fit_for_all_backends"
            + " or test_classifier_chain_fit_and_predict_with_linear_svc"
            + " or (test_gradient_boosting and test_poisson)"
            + " or (test_ridge and test_lbfgs_solver_consistency)"
            + " or (test_linear_models_cv_fit_for_all_backends and MultiTask)"
            + " or test_spectral_embedding_callable_affinity"
            + " or test_spectral_embedding_precomputed_affinity"'
        # reduce contention for parallelism (xdist-workers, openblas, openmp) that
        # actually increases the runtime and can lead to timeouts
        - if: aarch64 or ppc64le
          then: 'os.environ["OMP_NUM_THREADS"] = "1"'
        - 'args = ["--pyargs", "sklearn", "-k", f"not ({tests_to_skip})", "-nauto", "--timeout=1200", "--durations=50"]'
        - if: not (ppc64le or aarch64)
          then:
            - 'print(f"running pytest with args: {args}")'
            - 'pytest.main(args)'
        # NOTE: test suite is skipped on ppc due to bugs in QEMU code that cause
        # CI to fail, even though the tests should run through on native hardware
        # NOTE(bis): test suite is skipped for aarch64 due to issue with lapack
        # We are not able to reproduce the issue locally

about:
  license: BSD-3-Clause
  license_file: COPYING
  summary: A set of python modules for machine learning and data mining
  homepage: http://scikit-learn.org/
  repository: https://github.com/scikit-learn/scikit-learn

I'm using rattler-build 0.49.0 on macOS.

[lesteve]

For a bit more context this seems like a variant of https://github.com/prefix-dev/rattler-build/issues/1748 which happens with recipe/variants.yaml.

They may share the same root cause, no idea.

[wolfv]

Will have to look into this. Thanks for the bug report, and it might indeed be the same issue.