The failures are weird. Here is a summary of what I have seen so far:

• all tests pass on Linux,
• some test_cdist parametrization (but not all) fail on Windows and Mac,
• each time there is a failure we have something like:

            D_sklearn = dm.pairwise(X_csr, Y)
            assert D_sklearn.flags.c_contiguous
>           assert_allclose(D_sklearn, D_scipy_cdist, **rtol_dict)

D_scipy_cdist = array([[1.21796151, 0.99804582, 0.64796144, 1.09320258, 1.49914561,
        1.12392043, 0.47718256, 1.020455  , 0.9278..., 2.10374725, 2.20617501, 1.51727776, 1.13402   ,
        1.92153794, 1.95399778, 0.78790286, 0.78723968, 1.36559244]])
D_sklearn  = array([[1.21796151, 0.99804582, 0.64796144, 1.09320258, 1.49914561,
        1.12392043, 0.47718256, 1.020455  , 0.9278..., 2.10374725, 2.20617501, 1.51727776, 1.13402   ,
        0.        , 0.        , 0.        , 0.        , 0.        ]])

E       AssertionError: 
E       Not equal to tolerance rtol=1e-07, atol=0
E       
E       Mismatched elements: 100 / 500 (20%)
E       Max absolute difference: 2.44808142
E       Max relative difference: 1.

the relative error is often, but not exactly 1.0. Meaning that there are 100 out of 500 exact zeros in the scikit-learn CSR/dense distance computation.

So it looks like a pointer arithmetic bug to me.

Good catch. But why doesn't it fail on Linux?

Good catch. But why doesn't it fail on Linux?

Honestly, I do not see why tests were passing on configuration using Linux. Let's put it on hazard? I think this problem motivates the need to add global_random to those tests. What do you think?

On a side note, you mentioned in a review that n_features could be used instead of size for DistanceMetrics: do you think this PR is a good opportunity for this change?

Skipped jobs on the CI are due to an unusual way of raising awareness of Ubuntu 18.04 upcoming removal. See: actions/runner-images#6002

Honestly, I do not see why tests were passing on configuration using Linux. Let's put it on hazard?

It cannot be chance: it was deterministically happening for different linux builds with different base images and different numpy / scipy versions.

I think this problem motivates the need to add global_random to those tests. What do you think?

I am pretty sure we would have had the same problem with a different random seed. The distance values where non-zeros and the same on all the platforms.

Were those tests skipped for some reason on the linux builds?

Answer: not it wasn't the case. I just checked that for instance test_cdist[X2-Y2-minkowski0] did run and pass on the Linux CI using the filtering capabilities of the Azure Web UI:

https://dev.azure.com/scikit-learn/scikit-learn/_build/results?buildId=45755&view=ms.vss-test-web.build-test-results-tab

It probably has to do with a malloc happening after a recently freed buffer for the result array: the distances values of the previous assertion were not reinitialized and therefore left to the correctly computed values of the dm.pairwise(X_csr, Y_csr) case while on Windows and Mac malloc would not reuse recently freed memory in the same way.

It probably has to do with a malloc of a recently freed buffer for the result array: the distances values of the previous assertion were not reinitialized and therefore left to the correctly computed values of the dm.pairwise(X_csr, Y_csr) case while on Windows and Mac malloc would not reuse recently freed memory in the same way.

Indeed, we are using np.empty which is returning an uninitialised buffer allocated with malloc.

Indeed, we are using np.empty which is returning an uninitialised buffer allocated with malloc.

I tried to make the test stronger as following:

diff --git a/sklearn/metrics/tests/test_dist_metrics.py b/sklearn/metrics/tests/test_dist_metrics.py
index bb95681eb..c895143fe 100644
--- a/sklearn/metrics/tests/test_dist_metrics.py
+++ b/sklearn/metrics/tests/test_dist_metrics.py
@@ -113,14 +113,23 @@ def test_cdist(metric_param_grid, X, Y):
         assert D_sklearn.flags.c_contiguous
         assert_allclose(D_sklearn, D_scipy_cdist, **rtol_dict)
 
+        # Erase values in this results buffer to avoid successive free/malloc
+        # reusing a buffer with the correct distances in the next calls,
+        # potentially silently hiding problems.
+        D_sklearn.fill(0.)
+
         D_sklearn = dm.pairwise(X_csr, Y_csr)
         assert D_sklearn.flags.c_contiguous
         assert_allclose(D_sklearn, D_scipy_cdist, **rtol_dict)
 
+        D_sklearn.fill(0.)
+
         D_sklearn = dm.pairwise(X_csr, Y)
         assert D_sklearn.flags.c_contiguous
         assert_allclose(D_sklearn, D_scipy_cdist, **rtol_dict)
 
+        D_sklearn.fill(0.)
+
         D_sklearn = dm.pairwise(X, Y_csr)
         assert D_sklearn.flags.c_contiguous
         assert_allclose(D_sklearn, D_scipy_cdist, **rtol_dict)

but I cannot make it fail on 1c7bd26 (last commit prior to your fix)... Not sure what's going on.