FIX Support readonly sparse datasets for manhattan_distances

sklearn/metrics/_pairwise_fast.pyx

@@ -35,9 +35,15 @@ def _chi2_kernel_fast(floating[:, :] X,
                 result[i, j] = -res
 
 
-def _sparse_manhattan(floating[::1] X_data, int[:] X_indices, int[:] X_indptr,
-                      floating[::1] Y_data, int[:] Y_indices, int[:] Y_indptr,
-                      double[:, ::1] D):
+def _sparse_manhattan(
+    const floating[::1] X_data,
+    const int[:] X_indices,
+    const int[:] X_indptr,
+    const floating[::1] Y_data,
+    const int[:] Y_indices,
+    const int[:] Y_indptr,
+    double[:, ::1] D,
+):
     """Pairwise L1 distances for CSR matrices.
 
     Usage:

sklearn/metrics/tests/test_pairwise.py

@@ -2,6 +2,7 @@
 from types import GeneratorType
 
 import numpy as np
+from joblib import Parallel
 from numpy import linalg
 
 from scipy.sparse import dok_matrix, csr_matrix, issparse
@@ -15,7 +16,7 @@
     # should be used instead.
     from scipy.spatial.distance import minkowski as wminkowski
 
-from sklearn.utils.fixes import sp_version, parse_version
+from sklearn.utils.fixes import sp_version, parse_version, delayed
 
 import pytest
 
@@ -1549,3 +1550,14 @@ def test_numeric_pairwise_distances_datatypes(metric, global_dtype, y_is_x):
     dist = pairwise_distances(X, Y, metric=metric, **params)
 
     assert_allclose(dist, expected_dist)
+
+
+def test_sparse_manhattan_readonly_dataset():
+    # Non-regression test for: https://github.com/scikit-learn/scikit-learn/issues/7981
+    matrices1 = [csr_matrix(np.ones((5, 5)))]
+    matrices2 = [csr_matrix(np.ones((5, 5)))]
+    # Joblib memory maps datasets which makes them read only.
+    # The following call was reporting as failing in 7981, but this must pass.
+    Parallel(n_jobs=-1, max_nbytes=0)(
+        delayed(manhattan_distances)(m1, m2) for m1, m2 in zip(matrices1, matrices2)
+    )