Serializing `Vector{Array}` flattens elements

[kleinschmidt]

julia> tab2 = (x = [rand(2,2) for _ in 1:10], );

julia> io2 = IOBuffer()
IOBuffer(data=UInt8[...], readable=true, writable=true, seekable=true, append=false, size=0, maxsize=Inf, ptr=1, mark=-1)

julia> Arrow.write(io2, tab2)
IOBuffer(data=UInt8[...], readable=true, writable=true, seekable=true, append=false, size=736, maxsize=Inf, ptr=737, mark=-1)

julia> seekstart(io2)
IOBuffer(data=UInt8[...], readable=true, writable=true, seekable=true, append=false, size=736, maxsize=Inf, ptr=1, mark=-1)

julia> tab2_in = Arrow.Table(io2)
Arrow.Table: (x = [[0.0854794818573692, 0.6156692467610922, 0.4816953048461168, 0.10732732583276494], [0.2699150559983352, 0.7190134247010207, 0.4620847068434688, 0.8086969500721466], [0.32599895190448236, 0.5913226835708243, 0.914412039189433, 0.04825070891100114], [0.19800356830304522, 0.30969840980302754, 0.7586621051179192, 0.8383737966971678], [0.46353328096692525, 0.7946259468169095, 0.9205855777866754, 0.30143398878132244], [0.7980760486814706, 0.3204359773634282, 0.5253557751574975, 0.8140218988518082], [0.9662048857122691, 0.17369565835410516, 0.18328845062159216, 0.5595773503180002], [0.5423562251958653, 0.6285041382249899, 0.31820822802763304, 0.42185178416765723], [0.6985239393699472, 0.8842247274247905, 0.5466899110422125, 0.2058087506887174], [0.352616487112392, 0.6636744200136235, 0.9140196029380305, 0.2594771305442407]],)

julia> tab2_in.x
10-element Arrow.List{Vector{Float64}, Int32, Arrow.Primitive{Float64, Vector{Float64}}}:
 [0.0854794818573692, 0.6156692467610922, 0.4816953048461168, 0.10732732583276494]
 [0.2699150559983352, 0.7190134247010207, 0.4620847068434688, 0.8086969500721466]
 [0.32599895190448236, 0.5913226835708243, 0.914412039189433, 0.04825070891100114]
 [0.19800356830304522, 0.30969840980302754, 0.7586621051179192, 0.8383737966971678]
 [0.46353328096692525, 0.7946259468169095, 0.9205855777866754, 0.30143398878132244]
 [0.7980760486814706, 0.3204359773634282, 0.5253557751574975, 0.8140218988518082]
 [0.9662048857122691, 0.17369565835410516, 0.18328845062159216, 0.5595773503180002]
 [0.5423562251958653, 0.6285041382249899, 0.31820822802763304, 0.42185178416765723]
 [0.6985239393699472, 0.8842247274247905, 0.5466899110422125, 0.2058087506887174]
 [0.352616487112392, 0.6636744200136235, 0.9140196029380305, 0.2594771305442407]

At first I thought this was just some cleverness with how Arrow was serializing a "column matrix" (e.g. a n-by-1 matrix). I guess it's probably not possible to store the metadata about the size of the elements if each element can potentially have its own size? Maybe using StaticArrays is called for?

[quinnj]

Oof, yeah, this is tricky. Sorry for the delay in responding btw. We've had the same issue w/ JSON3 serializing; it's just tricky to know what kind of Matrix you want to deserialize. I have some ideas on introducing a Arrow.lower and Arrow.construct that would allow custom types to "hook" into the serialize/deserialize process; that would enable you do wrap a Matrix in a custom type and then on deserialize, you could reinterpret to the size you want, assuming you know the sizes.

[quinnj]

Though perhaps the actual right way to support Matrix is do implement the Tensor serialization: https://arrow.apache.org/docs/format/Other.html

[ericphanson]

I ran into this with LegolasFlux; there I worked around it with a manual "FlatArray" type: https://github.com/beacon-biosignals/LegolasFlux.jl/blob/80569ab63a8248a8a063c76e0bbf701f4ada9bd4/src/LegolasFlux.jl#L15-L36

This is also coming up in Lighthouse.jl: https://github.com/beacon-biosignals/Lighthouse.jl/pull/45/files#diff-790bf7705c3c1ba479ac3f9931bc82ec801c90f412d218ab4f9f88020fc128beR1-R16

[quinnj]

@ericphanson, what do you think, should we do a custom serialization for AbstractArray{T, N} where N > 1? We could serialize N and what size returns to be able to correctly deserialize?

The alternative, as I mentioned above, would be to try and use the official tensor message type.

Pros of tensor message:

• official arrow way of doing tensors Cons of tensor:
• Not all implementations support the tensor message

Pros of using extension type (metadata):

• Pretty trivial implementation (basically what @ericphanson linked to above w/ the FlatArray type) Cons:
• Would make non-vector AbstractArrays Julia-specific w/ their serialization

We could probably still support the tensor message type later if we go w/ the extension type method, but it would be an explicit thing where you wrap your matrix/array in a Arrow.Tensor and then write that out.

[ericphanson]

Hm, I think wrapping is inconvenient when you might have a nested structure you would need to traverse to wrap each leaf. Interop with pyarrow would be nice though, especially if one could do zero-copy IPC to numpy arrays or similar. I know @omus is looking into interop with Ray so maybe he has an opinion about what would work best for that.