Numpy 2.0 compatibility

[maxnoe]

No description provided.

[maxnoe]

Python 3.9 builds are failing because the new version of blosc2 that supports numpy 2.0 is requiring python >=3.10

[maxnoe]

https://scientific-python.org/specs/spec-0000/ says to drop python 3.9 support in Q4 2023, @avalentino should we simply also drop 3.9 here?

[avalentino]

https://scientific-python.org/specs/spec-0000/ says to drop python 3.9 support in Q4 2023, @avalentino should we simply also drop 3.9 here?

OK for me, thanks

[maxnoe]

Ok, we are down to only a handful of failures under 2.0 now

[maxnoe]

The change is due to the type promotion changes in numpy 2.0, the tests expect in16, but according to the new rules, the values are promoted to int64.

@avalentino Do we need to keep the old behavior or is it fine to adapt the tests to check for int64 if running under numpy 2.0?

[avalentino]

@avalentino Do we need to keep the old behavior or is it fine to adapt the tests to check for int64 if running under numpy 2.0?

I would need to check on a chase by case basis.
As a preliminary consideration I would say that for all the test involving the expression evaluator it should be fine to follow the numpy2 convention (when run with numpy2).
@FrancescAlted do you agree? What was the approach in numexpr?

[maxnoe]

Actually, it's a bit different: the expected type changed, but pytables didn't. So the question would be: do we change the pytest expression engine rules to follow the new type promotions of bumpy 2.0

[FrancescAlted]

@avalentino Do we need to keep the old behavior or is it fine to adapt the tests to check for int64 if running under numpy 2.0?

I would need to check on a chase by case basis. As a preliminary consideration I would say that for all the test involving the expression evaluator it should be fine to follow the numpy2 convention (when run with numpy2). @FrancescAlted do you agree? What was the approach in numexpr?

IMO, the NumPy team did a good job in changing promotion rules for 2.0 only when it was common sense, so our approach for numexpr was to follow the new conventions. I also think that PyTables should adopt the new conventions and note that in the release notes.

[avalentino]

Thanks @FrancescAlted