numpy does not actually support decimal! Consequently the caller should either cast to Float64 first (with the potential loss of exactness that implies) and then export to numpy, or accept that the resulting array will contain decimal values in an object-dtype array, which will be inefficient.

What does pyarrow do here then?

The Series to_numpy method exposed a "use_pyarrow" parameter that the DataFrame method did not (but should have done). When converting from DataFrame the use_pyarrow parameter was implicitly True in the underlying Series passthrough; now it is exposed explicitly with the same default. Note that if the dtypes don't require it (eg: all int/float) we still do the conversion ourselves without pyarrow, so there is no actual change from the previous default behaviour.

Can we default to False?

What does pyarrow do here then?

Creates an object-dtype numpy array - which it occurs to me we can trivially do ourselves so we can handle this with/without pyarrow; updating the PR accordingly (will update the change description in a sec) ;)

Can we default to False?

Not really; that would result in a potential change of behaviour since DataFrame.to_numpy currently calls Series.to_numpy directly with the default param values - which include use_pyarrow=True. Note that we only actually use the pyarrow conversion if/when our own conversion fails, (or we're exporting to structured arrays).

If/when we want to change that then it should be done in conjunction with changing the default for Series, which would require some thinking (such as categorising exactly where/why we need the fallback and adding equivalent native paths as necessary, etc).

Creates an object-dtype numpy array - which it occurs to me we can trivially do ourselves so we can handle this with/without pyarrow; updating the PR accordingly (will update the change description in a sec) ;)

Yeap, not ideal, but no data-loss, which I guess is what you care about when you use decimals in the first place.

Yeap, not ideal, but no data-loss, which I guess is what you care about when you use decimals in the first place.

Exactly; can cast to float before export for performance if exactness isn't the #1 concern.