@views annotation affect how elements are printed

[ronisbr]

Hi guys!

The print behavior of string in DataFrame changes if the @view annotation is used. For example, consider the following code:

using DataFrames

df = DataFrame(:a => "Teste 1", :b => "Teste 2")

for i = 1:2
    println(df[1,i])
end

@views for i = 1:2
    println(df[1,i])
end

The output is:

Teste 1
Teste 2
"Teste 1"
"Teste 2"

I would expect that the behavior would not change if the @views annotation is used. This is caused because of the type data[1,i] when @views is used. In the first case it is String whereas in the second it is SubArray{String,0,Array{String,1},Tuple{Int64},true}.

I found this bug by analyzing the cause of this one: ronisbr/PrettyTables.jl#16

[bkamins]

The quick solution: please use getindex instead of [ syntax.

Regarding the core of the problem I will have to discuss it with @mbauman (I hope I will get a chance to do this).

Essentially Base now disallows the following:

julia> x = [1,2,3]
3-element Array{Int64,1}:
 1
 2
 3

julia> x[1] .= 10
ERROR: MethodError: no method matching copyto!(::Int64, ::Base.Broadcast.Broadcasted{Base.Broadcast.DefaultArrayStyle{0},Tuple{},typeof(identity),Tuple{Int64}})

In data frames we allow it:

julia> df = DataFrame(zeros(2,2))
2×2 DataFrame
│ Row │ x1      │ x2      │
│     │ Float64 │ Float64 │
├─────┼─────────┼─────────┤
│ 1   │ 0.0     │ 0.0     │
│ 2   │ 0.0     │ 0.0     │

julia> df[1,1] .= 100
0-dimensional view(::Array{Float64,1}, 1) with eltype Float64:
100.0

julia> df
2×2 DataFrame
│ Row │ x1      │ x2      │
│     │ Float64 │ Float64 │
├─────┼─────────┼─────────┤
│ 1   │ 100.0   │ 0.0     │
│ 2   │ 0.0     │ 0.0     │

I am not clear which is better (but this is related to other 0-dimensional broadcasting cases that were uncovered here #1890).

The fix is easy and quick to do (so hopefully all can be cleaned up during JuliaCon and a patch release tagged), but I want to make an informed decision before I move forward.

[ronisbr]

Thanks @bkamins ! In fact, in my specific case, it turns out that I do not need @views at all. But I though it might be interesting to report since we usually expect that such annotation does not have any effect on the code when printing.

[bkamins]

Thank you for reporting this. Actually this is a crucial part of consistency with Base regarding broadcasting for DataFrames.jl (i.e. how do we handle 0-dimensional containers and if/and how to materialize them). These are corner cases, but when you have hundreds of users they pop up quite fast 😄.

[bkamins]

Actually we should change the DataFrames.jl behavior to mimic what happens here:

julia> x = [[1,2,3]]
1-element Array{Array{Int64,1},1}:
 [1, 2, 3]

julia> x[1] .= 100
3-element Array{Int64,1}:
 100
 100
 100

julia> x
1-element Array{Array{Int64,1},1}:
 [100, 100, 100]

I will add it to #1890.

[bkamins]

fixed on 0.19.1