Add Series.drop_nil about explorer

[cigrainger]

Do we need the regular version of this? It feels like clutter. The idea of minimal verbs is that you can do everything you need with them. Why not just Series.filter(not is_nil(_))?

[cigrainger]

Just thinking out loud here: I think there's got to be a more elegant way of dealing with lists than cleaving to Polars's API too closely. I'm not sure I like everything about purrr but I wonder if something like map_depth might help us out here?

There's also modify_tree. Something that indicates it's recursive might be useful?

[josevalim]

@cigrainger you are right, we don't need drop_nil at the root. However, the issue presented here is also available for many of the aggregate functions. How to distinguish between max of a list and maxof the series?

map_depth/modify_tree is definitely interesting. Implementing it is a bit less trivial. We would need to introduce some sort of LazyDepthSeries, that collects operations in a series, but as it relates to a struct or list field, and then translate that to polars. Do we want to go down this route?

[cigrainger]

Maybe. I need to explore how Polars handles this itself and in py polars.

[cigrainger]

So putting this a bit to the test with Python Polars:

It seems like nested lists may not be supported? That would fix the recursion problem pretty cleanly.

In [16]: df = pl.DataFrame({"values": [[None, 1, None, 2], [None], [3, 4], [[None, 1], [2], [None]]]})

In [17]: df
Out[17]:
shape: (4, 1)
┌────────────────────┐
│ values             │
│ ---                │
│ list[i64]          │
╞════════════════════╡
│ [null, 1, … 2]     │
│ [null]             │
│ [3, 4]             │
│ [null, null, null] │
└────────────────────┘

[josevalim]

@cigrainger all lists need to be nested equally. When it fails to cast to a certain type, it returns null instead of raising.

[billylanchantin]

@cigrainger Also, in case you missed it, there was an interesting saga of us discovering what Polars was doing WRT nested lists here:

• #857

Some take-aways:

• The first element of a nested list is the tie-breaker when the dtype is ambiguous.
• We decided to be more strict with our inference code than py-polars in certain situations.

[cigrainger]

I did! Thanks @billylanchantin