Deprecate get and haskey for AbstractDataFrame

[bkamins]

I propose to deprecate get method for AbstractDataFrame. I think it is not used and in general it is confusing to have it since we moved from AbstractDict interpretation.

[nalimilan]

Shouldn't we also deprecate haskey? It's useful, but it's quite inconsistent since we don't define keys? Or maybe we should define keys too. That could make sense if we decide to keep allowing df[col], but it could also be confusing since a data frame is two-dimensional in most situations.

[bkamins]

This was exactly my thinking that haskey is useful (as opposed to get).

We could rename haskey to hascol to make it more explicit and restrict its signature to allow only Symbol or Integer. Then we could decide if we go back to #1712 and decide if we want columnindex (or colindex if we prefer) as this would be a complementary function to hascol.

[nalimilan]

Could we just deprecate haskey(n, df) in favor of n in names(df) for now? Adding hascol is an interesting possibility, but that would only make sense if we also add a function to access columns like getcol or column. Better not rush that decision.

[bkamins]

I could do it, but the problem is that n in names(df) will be much slower for two reasons:

• names makes a copy
• it will be a linear search, while haskey uses a dict inbuilt into Index

In short - if we do it, then it should be only a temporary decision with the plan to provide a first class function that checks if some column exists in a data frame.

How do you see this?

[nalimilan]

Yes, it's much slower, but do we expect it to be used in performance-critical paths?

[bkamins]

Yes - I just wanted to make sure that we note down the consequences.

[bkamins]

I have deprecated haskey.