Kwarg to choose missing values for unstack

[kescobo]

Current behavior:

julia> df = DataFrame(key=[:a,:a,:b,:b], variable=["x","y","x","z"], value=rand(4));

julia> wide = unstack(df, :variable, :key, :value)
3×3 DataFrame
│ Row │ variable │ a        │ b        │
│     │ String   │ Float64⍰ │ Float64⍰ │
├─────┼──────────┼──────────┼──────────┤
│ 1   │ x        │ 0.464025 │ 0.672549 │
│ 2   │ y        │ 0.570984 │ missing  │
│ 3   │ z        │ missing  │ 0.400742 │

What I'd like to get out instead:

julia> for col in names(wide)[Not(1)]
           wide[!,col] = disallowmissing(replace(wide[!,col], missing=>0.))
       end; wide
3×3 DataFrame
│ Row │ variable │ a        │ b        │
│     │ String   │ Float64  │ Float64  │
├─────┼──────────┼──────────┼──────────┤
│ 1   │ x        │ 0.464025 │ 0.672549 │
│ 2   │ y        │ 0.570984 │ 0.0      │
│ 3   │ z        │ 0.0      │ 0.400742 │

Keyword suggestions: missing_values, fillmissing (I don't love either of these, but thought it would be good to get the ball rolling)

[pdeffebach]

This is a good idea. I ran into an issue recently where freqtables was slow (I can't reproduce right now), but a workflow using by and unstack was fast. But then you had to fill in 0 for missing values.

[kescobo]

Consideration that came up on slack: missings that were already in the data would be replaced with my version (and the more concise version from @alejandromerchan: coalesce.(unstack(df, :variable, :key, :value), 0)). If we only want to replace things that arise during the unstack, here's a different MWE to start:

julia> df = DataFrame(key=[:a,:a,:a,:b,:b], variable=["x","y","m","x","z"], value=[.1,.2,missing,.3,.4]);

And we'd want

julia> wide = unstack(df, :variable, :key, :value, missing_values=0.) # hypothetically
4×3 DataFrame
│ Row │ variable │ a        │ b        │
│     │ String   │ Float64⍰ │ Float64  │
├─────┼──────────┼──────────┼──────────┤
│ 1   │ m        │ missing  │ 0.0      │
│ 2   │ x        │ 0.1      │ 0.3      │
│ 3   │ y        │ 0.2      │ 0.0      │
│ 4   │ z        │ 0.       │ 0.4      │

Though actually, I'm not sure what we'd want at wide[1, :b] in this case...

[bkamins]

The same pattern was already discussed in #1864, so we would be consistent between the functions (that PR is currently extremely hard to look at but I hope you will see the idea).

[kescobo]

Oh, awesome - if the pattern has been considered and consensus reached, as long as it's consistent I think that's great. I don't have a strong intuition about it.

[metanoid]

I'd prefer something like filldefault or default instead of fillmissing because I think that makes it clearer that we're only inserting the default value where the row was omitted from the source df, not overwriting existing missings in the source df.

[bkamins]

This is now supported with fill kwarg