Allow `Base.filter` to work with all iterators

[pdeffebach]

This issue is motivated by this discussion on discourse.

My example of when this is a problem is when you are working with a subset of data that, by chance, does not contain missing values. When you work with the full data, you now have missing values and have to use skipmissing at various places in your code. Now all your filter commands are broken.

A heavy and breaking change would be to make Base.filter work on any iterator, and not just arrays. Other alternatives include pointing more people towards Iterators.filter for general use, or special-casing skipmissing so a collect call is included somewhere.

[ararslan]

One issue here is that filter is eager and, generally speaking, typeof(filter(f, x)) == typeof(x). That isn't really possible to implement in a general case, but I think it would be okay to define filter for SkipMissing to materialize and filter.

[nalimilan]

So I guess in that proposal filter(::SkipMissing{<:AbstractArray}) would work, but not with other wrapped iterators?

[ararslan]

That would make sense to me. SkipMissing is an odd enough case itself that I think special treatment of it would be fine. We could define it as

filter(f, s::SkipMissing) = filter(x->!ismissing(x) && f(x), s.x)

then let filter fail if the underlying iterator doesn't have filter defined on it.

[pdeffebach]

An alternative is that people could just change their filter commands to have ismissing, which is a bit more explicit anyways.

This could just be a documentation / tutorial problem.

[ararslan]

I was looking at implementing what I said above verbatim, but I realized that that definition leaves you with a container without any missing values but with an element type that allows missing. That seems suboptimal and is inconsistent with the result of filter(f, collect(s)).

This is difficult to do generically, since ideally you'd want to be able to use filter on any SkipMissing{T} where T supports filter and get a T back, but to provide useful and consistent behavior, I'm not sure how we'd swing it for anything but arrays...

[nalimilan]

Supporting filter only for AbstractArray sounds OK to me.

[ararslan]

Implemented in #31235.

[nalimilan]

One possible rationale for supporting filter(::SkipMissing{<:AbstractArray}) and not other methods is that SkipMissing{<:AbstractArray} is indexable. If an Indexable trait was added as discussed at #31020 (comment), it would make sense for filter to support all types with that trait. That would include in particular Broadcasted objects.

[ararslan]

We can't support filter for arbitary iterables and we now have filter for SkipMissing-wrapped arrays, so is there anything actionable left or can this be closed now that #31235 is merged?

[nalimilan]

As I noted it could make sense to also support Broadcasted. Not sure whether that warrants keeping this open.

[ararslan]

Ah right, sorry. It might make sense to have that as a separate issue, since it's very specific (and a difficult situation to find yourself in, since presumably a broadcast will materialize before filter is called).

[pdeffebach]

Broadcasted would materialize an array while lifting missing elements? Would be great!