In docs, note subsets are copies (unless of columns)?

[nickeubank]

Loving updated docs, and the (clear!) discussion up front about views v. copies (my greatest pet peeve about pandas).

Are people open to noting in Indexing that subsets of columns with ! are non-copies, while subsets of rows always generate copies (I assume?). Can PR if supported.

[bkamins]

Sure - go ahead with a PR please. And yes - getindex always creates a copy if row index is anything except ! or a single row selected using an integer (in which case it is either a single value if a single column is selected or a DataFrameRow if multiple rows are selected - in the last case note that DataFrameRow is a view). Feel free how much of this complexity should go to "getting started" and what can be left as "advanced stuff" 😄.

[nickeubank]

Thanks @bkamins! I definitely don't want to over complicate -- for me though, knowing when you have a view and when you have a copy has profound implications for that most critical of things -- data integrity -- so I think it's worth pushing up front.

[nickeubank]

But df[!, [firstcolumn, secondcolumn]] isn't a view, right? That's a copy?

julia> df = DataFrame("a" => [42, 47], "b" => ["y", "z"])
2×2 DataFrame
│ Row │ a     │ b      │
│     │ Int64 │ String │
├─────┼───────┼────────┤
│ 1   │ 42    │ y      │
│ 2   │ 47    │ z      │


julia> df[!,["a", "b"]] === df
false

[bkamins]

df[!, [firstcolumn, secondcolumn]] is a copy of a wrapper but not a copy of the contents.

As a data frame is a container we have THREE types of selection:

• df[:, [:a, :b]] - copy of columns in a freshly allocated DataFrame
• df[!, [:a, :b]] - alias (no copy, no view) of columns in a freshly allocated DataFrame (old and new data frames share columns :aand:band the columns pass===` test if compared)
• df2 = @view df[:, [:a, :b]] or df2 = @view df[!, [:a, :b]] are the same and create a SubDataFrame (containing all rows and two columns of df), you have that for columns :a and :b that parent(df2.a) === df.a.

The details are described here. In general we have indexing and broadcasting design that is fully compatible with Base (+ an extension of !) while at the same time it allows you to have the ability to express any kind of indexing behaviour that can be needed (copy - safe, alias - fast and unsafe, view - fast and explicit that it is unsafe).

[nickeubank]

@bkamins PR with a "Note" on this topic. Thoughts?

[nickeubank]

Thanks. Tweaked a few places you are definitely right, left language ambiguities for others to review.

[matthieugomez]

Do we really need the “df[!, cols]” syntax in the end? It sounds like the behavior can be obtained either by df.”$x” (to get vector) or select (to get dataframes).

[bkamins]

we need it for setindex! and broadcasting.