@jnothman I've reopened #12648, that PR is consistent with 0.20.X

There's a comment at the beginning of the test: # check that fit.predict gives same result as fit_predict.
Perhaps it's acceptable to merge #12648 since that PR is already approved and merged into 0.20.X?

It was approved as a short term fix, and no other fix was merged on the basis that the problem seemed to have gone away

I'll remove the reference to fit from the comment, but the point we to check the iterative updating when fitting matched assignment by pairwise_distances_argmin

I don't think the intention of this test is to check the determinism / idempotence of fit/fit_predict

Actually it was my intention :)
There's already a test which correspond to what your changes test: test_predict

Oh!
Hmm. If the point is to test the consistency of multiple fits, why does comparing fit(X).predict(X) to fit_predict(X) come into it??? I agree there is redundancy between those two tests, but I think test_k_means_fit_predict should not be doing what it is, nor named as it is, if its intention is to check the consistency of multiple calls to fit.

My intention was just to test that calling fit(X).predict(X) gives the same result as calling fit_predict(X). There are very similar tests for other estimators, like test_bayesian_mixture_fit_predict for instance.

My intention was just to test that calling fit(X).predict(X) gives the same result as calling fit_predict(X).

But under the assumption that fit_predict(X) and fit(X).labels_ return the same thing (which they do, even if it's not obviously tested), then test_predict tests exactly that... given that the fit is consistent across calls.

Can we rename the failing test to be test_fit_idempotence or something? And then skip it?? :|

given that the fit is consistent across calls.

I think it is. If I recalled correctly the failures appear when using elkan algorithm because the predict method always uses lloyd algo and even if both algo should give the same results, there might be some numerical instabilities.

I'm ok to rename and skip the test

@jnothman maybe close this one and merge #12648?