Semantics of `similar` about dataarrays.jl HOT 5 CLOSED

IMO, it should just do allocation.

For ordinary Julia arrays, similar(a) returns something whose contents are completely random/undefined. People can use fill and whatever to initialize the content. The same should be similar for data arrays. People can initialize everything to NA or not NA using additional statements.

from dataarrays.jl.

Let's see what @simonster thinks. If he agrees, we should make the change. After that, we need to review the codebase to make sure that we set NA properly in anything generated by similar.

from dataarrays.jl.

All other things being equal, I think it's more user-friendly to do NA setting. I don't think there's much point in consistency with Base for consistency's sake, since the behavior is not even entirely consistent there (bits types get random values, but other types get #undefs, which presumably requires zeroing the array of pointers). NA setting seems unlikely to cause much confusion; code that expects an uninitialized array should work just fine with a DataArray full of NAs.

If there's a performance impact of NA setting, that would be a good reason not to do it, but my intuition is that it would be minimal. For Ints and Float64s, there are 64x as many bits in the data array as in the na array. For code that does nothing but fill the data array, the cost to fill the na array should be <2% of the total time, and for anything more complicated it is likely immeasurably small.

from dataarrays.jl.

@simonster's argument is fair enough to me.

from dataarrays.jl.

Yeah, I think this is a reasonable hypothesis. Let's use Simon's auto-NA semantics until there's a legitimate performance problem. I'll add a few tests to make sure we're doing this correctly.

from dataarrays.jl.