Error when writing Array{Float64, 0}

[tomchor]

Describe the bug

I get an error when trying to write a type Array{Float64, 0} to an index of a variable with an unbounded dimension. The error is also reported in CliMA/Oceananigans.jl#2857

To Reproduce

using NCDatasets

a = reshape([1.], 1, 1, 1)
b = dropdims(a, dims=(1, 2, 3))

ds = NCDataset("/tmp/test.nc","c")
time = defDim(ds,"time",Inf)
v = defVar(ds,"temp",Float32,("time",))
ds["temp"][1] = b

close(ds)

In believe this comes from the function call T(data) in this line:

NCDatasets.jl/src/variable.jl

Line 308 in 69eda01

nc_put_var1(v.ds.ncid,v.varid,[i-1 for i in indexes[ndims(v):-1:1]],T(data))

Expected behavior

I expected the Array to be successfully writen to the NetCDF. For reference, this MWE works, which shows that the issue comes when using an unbounded dimension:

using NCDatasets
    
a = reshape([1.], 1, 1, 1)
b = dropdims(a, dims=(1, 2, 3))

ds = NCDataset("/tmp/test.nc","c")
v = defVar(ds,"temp",Float32,())
ds["temp"][] = b

close(ds)

Environment

• operating system: Manjaro
• CPU architecture Intel x86_64
• Julia version: julia 1.8.2, official binaries from https://julialang.org/downloads/
• NCDatasets version: v0.12.11
• Output of using Pkg; Pkg.status(mode=PKGMODE_MANIFEST): https://pastebin.com/gfbdcRu1

Full output

julia> a = reshape([1.], 1, 1, 1)
1×1×1 Array{Float64, 3}:
[:, :, 1] =
 1.0

julia> b = dropdims(a, dims=(1, 2, 3))
0-dimensional Array{Float64, 0}:
1.0

julia> ds = NCDataset("/tmp/test.nc","c")
NCDataset: /tmp/test.nc
Group: /



julia> time = defDim(ds,"time",Inf)

julia> v = defVar(ds,"temp",Float32,("time",))
temp (0)
  Datatype:    Float32
  Dimensions:  time

julia> ds["temp"][1] = b
ERROR: MethodError: no method matching Float32(::Array{Float32, 0})
Closest candidates are:
  (::Type{T})(::AbstractChar) where T<:Union{AbstractChar, Number} at char.jl:50
  (::Type{T})(::Base.TwicePrecision) where T<:Number at twiceprecision.jl:266
  (::Type{T})(::Complex) where T<:Real at complex.jl:44
  ...
Stacktrace:
 [1] setindex!(v::NCDatasets.Variable{Float32, 1, NCDataset{Nothing}}, data::Array{Float32, 0}, indexes::Int64)
   @ NCDatasets ~/.julia/packages/NCDatasets/sLdiM/src/variable.jl:308
 [2] setindex!(v::NCDatasets.CFVariable{Float32, 1, NCDatasets.Variable{Float32, 1, NCDataset{Nothing}}, NCDatasets.Attributes{NCDataset{Nothing}}, NamedTuple{(:fillvalue, :missing_values, :scale_factor, :add_offset, :calendar, :time_origin, :time_factor), Tuple{Nothing, Tuple{}, Nothing, Nothing, Nothing, Nothing, Nothing}}}, data::Array{Float64, 0}, indexes::Int64)
   @ NCDatasets ~/.julia/packages/NCDatasets/sLdiM/src/cfvariable.jl:765
 [3] top-level scope
   @ REPL[34]:1

julia> close(ds)
closed NetCDF NCDataset

I was wondering if this is something that should be dealt with on the NCDatasets side, or if this is by design and we should deal with it on the Oceananigans side.

Thanks

[Alexander-Barth]

Thanks a lot for the very detailed bug report!

I noticed that also in Base julia complains a bit about this:

b = dropdims([1], dims=(1,))
temp = [1.,2.,3.]
temp[1] = b # ERROR: MethodError: Cannot `convert` an object of type Array{Float64, 0} to an object of type Float64

But the following works in Julia:

b = dropdims([1], dims=(1,))
temp = [0.,2.,3.]
temp[1:1] = b # ok

In the master version (but not in v0.12.11) the following now work too with NCDatasets,

using NCDatasets
b = dropdims([1.], dims=(1,))

isfile("/tmp/test.nc") || rm("/tmp/test.nc")
ds = NCDataset("/tmp/test.nc","c")
time = defDim(ds,"time",Inf) # works also for fixed size arrays
v = defVar(ds,"temp",Float32,("time",))
ds["temp"][1:1] = b
close(ds)

Are you fine to use ds["temp"][1:1] in your code when using 0-dimensional arrays?
(thanks for your great work on Oceananigans :-) )

[tomchor]

Thanks a lot for the very detailed bug report!

I noticed that also in Base julia complains a bit about this:

b = dropdims([1], dims=(1,))
temp = [1.,2.,3.]
temp[1] = b # ERROR: MethodError: Cannot `convert` an object of type Array{Float64, 0} to an object of type Float64

But the following works in Julia:

b = dropdims([1], dims=(1,))
temp = [0.,2.,3.]
temp[1:1] = b # ok

In the master version (but not in v0.12.11) the following now work too with NCDatasets,

using NCDatasets
b = dropdims([1.], dims=(1,))

isfile("/tmp/test.nc") || rm("/tmp/test.nc")
ds = NCDataset("/tmp/test.nc","c")
time = defDim(ds,"time",Inf) # works also for fixed size arrays
v = defVar(ds,"temp",Float32,("time",))
ds["temp"][1:1] = b
close(ds)

Are you fine to use ds["temp"][1:1] in your code when using 0-dimensional arrays?

Thanks for the fix! I think this solution will work out (let's see how the discussion on the relevant PR (CliMA/Oceananigans.jl#2865) goes to be sure...)

(thanks for your great work on Oceananigans :-) )

Thank you! But I can't take the merit there. I'm only a very minor contributor :)
But yeah, I think the main developers there do a really good job!

[Alexander-Barth]

I am closing the issue; feel free to re-open if there are still some difference between Base.Array and NCDatasets regarding the 0-D arrays.