I'm a little confused how this diverges from the discussions in #19706.
ETA: is it cause a) targeted for 3.6, b)intended as a change all the examples together, c) now documented in the tests w/ #23221 ?

#19706 changed user-facing examples. The fear there is that users are sitting on older versions of matplotlib and numpy and get an error when they copy the example code.

Here the scope is limited to tests. We require numpy 1.17+ for the current codebase and that also applies to tests. People working on matplotlib cannot fall into the numpy-too-old trap.

I am 50/50 if the tests pass with the new generator (seems like a bunch of work and code changes for a low pay-off if numpy is not going to deprecate the old API) and 👎🏻 if this change requires any tests to be changed (because changing tests is a bad idea on principle).

I'm fine with closing. But then we should add a note as a follow up to #23221 that existing tests should be kept with the old API.

I honestly neutral on this. Despite my concerns, Jody is right that this is an easy first issue and people are at least as likely to find another test and copy-paste-edit rather than read the docs and write one from scratch so it will be easier on everyone if our tests all follow the documentation.

We should at least find out if it breaks stuff before before closing.

I tried this with one test and it fails because generated data is not the same. A lot of tests in that module that use random data are also using check_figures_equal, so would presumably be fine if the generated data changed. Some tests in the module use random data without setting a seed.

Isn't that because you didn't change the standard_normal on the next line, causing it to use the default (un-seeded) generator?

🤦‍♀️ now updated, but the test still fails.

Yes, it appears that default_rng uses PCG64, while the older global functions use MT19937, though I'm not sure how to best replicate that with the objects.

My 2 cents: If this cannot be simply replaced, it's not worth the effort.

I think the consensus here is that, because the new generator does not give the same numbers, we should not go forward with the replacement.

@tacaswell 👎🏻 if this change requires any tests to be changed

@timhoffm If this cannot be simply replaced, it's not worth the effort.

So I am going to close. Obviously re-open if you disagree.