DEP: Complete deprecation of invalid array/memory order

[sethtroisi]

Original deprecation was submitted in #6823 in 2015.

[seberg]

I think for errors I prefer the formatting:

PyErr_SetString(PyExc_ValueError,
        "this is a very long error message .... ... ... .. "
        "...");

As for the above comment, I do not think we typically put that in, not that it hurts, though, hmmm. The idea of the DEPRECATED comment is to find these places that need updating easier. What is important here is only the release note.

[sethtroisi]

I updated the PyErr_SetString, I'm happy to

• remove the both dates
• consolidate them into a single line
• /* more order formats were handled before DEP: Complete deprecation of invalid array/memory order #14596 */

[seberg]

Others may disagree, but if you do not mind, lets just remove the comments, they are not interesting internally for numpy, only for downstream. And downstream would read the release notes.

Can you reformat the other error setting as well?

The main thing to do is add the release note snippet, thanks!

[sethtroisi]

Removed comments and added a release note snippet.

[seberg]

Thanks. LGTM. Do you want to squash it into a single commit with DEP: ...? I can also squash merge, but thought I would mention just in case, since we slightly prefer merges.

[sethtroisi]

I'm happy to do what you prefer.

I'm squashing and force committing a single commit with the correctish title

[seberg]

Interesting. The test failures seem real. There must be some documentation around which actually uses the deprecated version, hah...

[sethtroisi]

There's one test failure in windows for TestMatmulOperator.test_matrix_vector_values. I don't know how flaky numpy's test (or the window tests) are. I gave it a quick glance and it doesn't look like I caused it. I'm closing and opening this to re-try the test.

[charris]

It's a Heisenbug that shows up randomly, usually on windows 32 bits.

[seberg]

Hmmm, there once was a bug with sqrt in the kernel calculation that was triggered in numpy. It seems that at the time that seems maybe not fixed everywhere. Although unless the machines run OpenBLAS with a different number of threads (on a different) cpu once in a while, it is surprising that it is so rare (although possibly it matters how quickly threads finish, hmmm). Maybe should just fix those things upstream in any case, might help...

EDIT: Anyway, made a PR for OpenBLAS, if we are lucky, the next update fixes it, I guess...

[mattip]

Thanks @sethtroisi

[eric-wieser]

This test is unfortunately nonsense, this passes a dict as the order argument.