This is the new strings r/w behavior for 3.x series: https://docs.h5py.org/en/latest/strings.html

Just to reference the original ticket #44467

In the CI logs I see different unicode issues related to https://docs.h5py.org/en/latest/strings.html#what-about-numpy-s-u-type

TypeError: No conversion path for dtype: dtype

But CPU tests don't fail locally inside the official TF devel image tensorflow/tensorflow:devel.

It is hard to work locally where official devel images are not aligned with the CI.
I think an user can expect to collect the same test result of the CI Ubuntu CPU as locally with tensorflow/tensorflow:devel if not it will be hard to develop in a local reproducible system.

I think that probably this is caused unaligned tensorflow/tensorflow:devel with LANG_CTYPE=C.UTF-8 and other image used from the CI that are LC_CTYPE="POSIX"

I've tried to use POSIX locally in tensorflow/tensorflow:devel but tests are passing. I cannot reproduce CI fails..

So I've just reverted on one of the first commit to let you see that Linux is passing but not Win/Mac

@mihaimaruseac What was the Win and OSX issue?

Windows seems to be broken at HEAD due to an MLIR integrate from 12 hours ago.

MacOS seemed to pass. Probably the environment is different? Hopefully there is no rollback.

@mihaimaruseac Can you reopen this? Do you have Shell access on a win/MacOS machine?

I cannot reopen, GitHub does not let me :(

I can get access to a windows machine, but it will take a while as I have a few P0 items to finish this week/early next week.

Yes I know it is merged but I meant something like #45487