#43970 is dropping Python 3.8 for pyarrow, and as a consequence the tested minimum versions of numpy and pandas had effectively to be bumped numpy from 1.16 to 1.19.5 and pandas from 1.0 to 1.1.3.

But we should bump those minimum versions more officially, and update our code/packaging to do so:

• In pandas-shim.pxi we explicitly check for the pandas version to raise an error if too old, so this version can be updated as well
• We declare "numpy >= 1.16.6" in pyproject.toml, which needs to be updated.
• We have some tests that can be cleaned up (that check for the numpy/pandas version)

While for supporting Python >= 3.9, we effectively need to bumpy to pandas 1.1.3 and numpy 1.19.5, I think we could do a bit bigger bump, and the proposal is to bump mimimal support to:

• numpy 1.23
• pandas 1.5

For pandas, 1.5 is the last release in the 1.x cycle (and was released Sept 2022, so two years ago). And for numpy if we look at a similar time range as pandas 1.5 in terms of when it was released, then numpy 1.23 or 1.24 also makes sense. Although for numpy's range of 1.16 - 1.24, we don't have any special case in our tests, based on a quick look (except for skipped tests for dlpack), so we can also easily be more conservative for numpy and use a longer support window here.

(given we don't really have issues regularly with older numpy, I think I personally lean to be more conservative)