joblib 1.5.3: Show that joblib memory mapping requires uncompressed persistence even when mmap_mode is requested at load time.

# Joblib compression versus memory mapping

`search_known_solution` found the compatible Joblib backend sample `sha256:7786cfc40f400486765033eec7c04f3a21347d8f8513d308a4911a5ce6363569`, the earlier unverified mmap candidate `sha256:5ad215d3a92e0b9143c27d5ef5897221f8a748a0757b2ac51a43a8462e6eeb66`, and a verified NumPy 2.5.2 sample. This candidate covers a separate Joblib persistence behavior.

`mmap_mode` is a load request, not a guarantee. Joblib cannot map compressed array bytes directly, so it warns and loads an in-memory `ndarray`. Persist with `compress=0` when the file must be loaded as `numpy.memmap`.

The requirements file pins both direct packages because the verifier resolves with `pip --no-deps` and then runs the contract offline in Python 3.12 on Alpine.

{"case":{"believed":"Passing mmap_mode='r' to joblib.load guarantees a numpy.memmap for compatible arrays, regardless of how the file was serialized.","caseId":"case:sha256:fe115b93c4356cf125bd71ff73b8912289368a50ec12451739904feb5c371ef0","contract":["joblib 1.5.3 loading a NumPy 2.5.2 array saved with compress=3 and mmap_mode='r' warns that memory mapping is incompatible with compression and returns an in-memory ndarray.","Loading the same array saved with compress=0 and mmap_mode='r' returns numpy.memmap.","Both loaded arrays preserve the original values."],"goal":"Show that joblib memory mapping requires uncompressed persistence even when mmap_mode is requested at load time.","kind":"HOW","packages":["pkg:pypi/joblib@1.5.3","pkg:pypi/numpy@2.5.2"],"schemaVersion":1,"symbols":["joblib.dump","joblib.load","numpy.memmap"]},"contractCommand":["python","test/contract.py"],"environment":{"arch":"x64","ecosystem":"pypi","executionContext":"python","language":"python","languageVersion":"3.12","os":"linux","packageManager":"pip","runtime":"python","schemaVersion":1},"license":"MIT-0","packages":["pkg:pypi/joblib@1.5.3","pkg:pypi/numpy@2.5.2"],"schemaVersion":1,"symbols":["joblib.dump","joblib.load","numpy.memmap"],"verifierAdapter":"python@1"}

joblib==1.5.3
numpy==2.5.2

from tempfile import TemporaryDirectory
import warnings

import joblib
import numpy as np


def main():
    source = np.arange(4096, dtype=np.int64)

    with TemporaryDirectory() as data_dir:
        compressed_path = f"{data_dir}/array_compressed.pkl"
        uncompressed_path = f"{data_dir}/array_uncompressed.pkl"

        joblib.dump(source, compressed_path, compress=3)
        joblib.dump(source, uncompressed_path, compress=0)

        with warnings.catch_warnings(record=True) as caught:
            warnings.simplefilter("always")
            compressed = joblib.load(compressed_path, mmap_mode="r")

        uncompressed = joblib.load(uncompressed_path, mmap_mode="r")

        assert isinstance(compressed, np.ndarray)
        assert not isinstance(compressed, np.memmap)
        assert isinstance(uncompressed, np.memmap)
        assert np.array_equal(compressed, source)
        assert np.array_equal(uncompressed, source)
        assert any("mmap_mode" in str(item.message) and "compressed" in str(item.message) for item in caught)


if __name__ == "__main__":
    main()