wrapdims(::DataFrame) produces incorrect results when not all key combinations are present

[fredcallaway]

df = DataFrame([
    (x="a", y="a", z=1),
    (x="a", y="b", z=2),
    (x="b", y="b", z=3),
])
wrapdims(df, :z, :x, :y)

The result varies, but here's one example

                 ("a")  ("b")
  ("a")           1      2
  ("b")  4615108224      3

I think the matrix should be initialized with an undefined value rather than an arbitrary value (maybe missing)?

[fredcallaway]

Sorry, I should have read the documentation better. I think the default argument should be missing by default. The current default behavior can lead to errors from careless users like myself (and let's be honest, there are many of us).

[rofinn]

The issue with making the default missing is that this introduces unnecessary type complexity in many cases. If there are no missing values the returned element type is still Union{T, Missing} which may break downstream code. I'm not particularly bothered either way, that's the main reason we didn't make it the default.

[fredcallaway]

Yeah, I totally see why you would want to not use missing if you know there are no missing values. But I guess I think it's better to fail obviously (with an error) than to produce incorrect values. One could even make an argument for requiring the default value be passed explicitly.

The minimal action would be to add a more salient warning to the documentation (both the README and the docstring). I can draft that if you like.

[rofinn]

default value be passed explicitly

Okay, yeah, that's probably the best path forward.