behaviour of `similar` with axes argument is inconsistent

[tiemvanderdeure]

MWE:

using DimensionalData, Rasters
ra = Raster(rand(X(1:10), Y(1:10), Z(1:1)); metadata = Dict("A" => 1), name = :myraster)[Z = 1]

ra2 = similar(ra, Int)
ra3 = similar(ra, Int, axes(ra))
ra4 = identity.(ra2)

# All of these are false
name(ra2) == name(ra3) # Symbol("") and NoName()
refdims(ra2) == refdims(ra3) # Z and ()
metadata(ra2) == metadata(ra3) # Dict and nothing
typeof(ra3) # DimArray not raster!

Making this consistent is not hard but will technically be breaking.

Can we fit in some changes here before the next breaking release #946?

[rafaqz]

Sounds good! Ra3 is because its hitting the fallback for non AbstractDimArray and getting the dims from our sneaky dimensional axes. But it should be calling rebuild on ra.

That code is some pretty wild nested evals already, let me know if you have trouble

[rafaqz]

Was this all fixed in #1082 ?

[tiemvanderdeure]

Basically, this now gives:

name(ra2) == name(ra3) == name(ra4) # true - all Symbol("")
refdims(ra2) == refdims(ra3) # false - Z and (), but hard to fix?
metadata(ra2) == metadata(ra3) == metadata(ra4) # true - all keep metadata
typeof(ra3) # is now a Raster!

So the only outstanding thing is the refdims - but I don't think the similar method with axes really can retain refdims, since the axes might have completely different dims (including some identical to the refdims).

Then there is the question of whether metadata should be retained after broadcasting, but maybe that is for a different issue.