Should `zonal` include a kwarg `if_all_missing`

[alex-s-gardner]

The following returns an error as all data is missing and the maximum is being requested:

dX = X(1:10)
dY = Y(1:10)
 
ras = Raster(fill(NaN, (dX,dY)); missingval=NaN)

ext = Rasters.Extent((X=(2,9), Y=(2,9)))
Rasters.zonal(maximum, ras; of=ext, skipmissing=true)

ERROR: ArgumentError: reducing over an empty collection is not allowed; consider supplying `init` to the reducer
Stacktrace:

should a kwarg be added to zonal to tell it what to do when no valid data is found? This would eliminate the need for try/catch statements to catch this error.

[rafaqz]

Right yes that's annoying.

Do you want a zero? Maybe an init keyword would help. Assumed to be some kind of identity like 0 for sum or 1 for prod, and only used if there are no other values. A missing is the other option that should work

[alex-s-gardner]

I think returning missing would make the most sense.

Not sure if that works for all possible reducing functions.

[rafaqz]

Yes so in this case using init in maximum fixes it. Should we allow passing that? It pretty simple to do. If set to nothing we won't pass it in.

[alex-s-gardner]

~~Is init too specific? What about a custom reducing function?~~

ya, init looks like the standard way to do this

[asinghvi17]

I guess you have some registry function that tells you whether to pass init?

[felixcremer]

We could add an innerkwargs keyword argument to zonal which would take a NamedTuple and forward this as keyword arguments to the reduction function. And then in your use case you can do Rasters.zonal(maximum, ras; of=ext, skipmissing=true, innerkwargs=(init=0,))

This would be more flexible and adaptable to other reduction functions.

[rafaqz]

But init=x is cleaner and close to universal

[asinghvi17]

I think in general an innerkwargs that gets passed to all functions is good, but we can special case init if we have a function registry that says which functions can use it. Can just be can_use_init(::Type{<: typeof(yourfunction)}) = true.