ENH: Changing Series and DataFrame `repr` for NaN values

[wesm]

With future pandas internal improvements in contemplation, I have often wondered if it would be worth changing the NaN outputs to be NA or NULL instead to reflect the actual semantics of the data. This could be something that's configurable in pandas.options (i.e. showing the semantic value or the physical value).

[jorisvandenbossche]

I would be in favor of that (NA instead of NaN).
The only thing I am wondering is if we already know whether we want to make a distinction between NA and NaN for float data in pandas 2.0, where this would theoretically be possible when using bitmasks for NA. Because if we want that distinction, it might make sense to wait with the other repr until 2.0 instead 1.0.

[wesm]

It's somewhat out of scope for this issue, but I've been thinking about the NaN problem in pandas 2.0, and I think for backwards compatibility reasons we're going to get forced to make NaN and NA / NULL equivalent during the transition period. We could later add warnings when NaN is being treated as NA in operations like s[...] = np.nan (which I can attest litters people's pandas code), and then later add an option (where the default is that NaN and NA are differnet), and then later remove the option.

I am pretty confident that using bitmaps everywhere will make our code much simpler and faster (i.e. can use SIMD operations on the bitmaps to deal with null analytics and propagation).

[wesm]

As one example of why things will be faster, we can use bitmaps to eliminate branching in aggregations:

sum_x += values[i] * GetBit(bitmap, i);

compared with

if (values[i] == values[i]) sum_x += values[i];

[jorisvandenbossche]

Yes, we could have a 'nan_as_missing' option that is first True and could possibly later change (similar to how we now have pd.options.mode.use_inf_as_null option that is False by default).

[jorisvandenbossche]

The issue to discuss this is probably: wesm/pandas2#46

[TomAugspurger]

FWIW, I think this will be effectively closed by #29964. That changes IntegerArray to use pd.NA, which has NA as its repr.

In [1]: import pandas as pd

In [2]: pd.Series(pd.array([1, 2, None]))
Out[2]:
0     1
1     2
2    NA
dtype: Int64

I don't think we'll want to move away from np.nan for ndarrays, and I don't think we want to display nan as NA, since they have different behaviors.