UndefInitializer -> cannot save as BSON / JLD

[kafisatz]

A dataframe of mine has the type below.
It has four unique entries (three strings and missing)
When trying to save this as BSON/JLD I get an UndeRefError.
Is there any way to keep working with SentinelArray and 'get rid of UndefInitializer'?
Or is there an option for CSV.read to avoid this?

I gather they are looking to make 'undef' work for BSON JuliaIO/BSON.jl#43
Then again, my data has no undef. Only the type has 'undef'...

this works for me (for the time being) :
dfraw.Instanz = convert(Array{Union{Missing,String},1},dfraw.Instanz);

df.Instanz
125845-element SentinelArrays.SentinelArray{String,1,UndefInitializer,Missing,Array{String,1}}:

#when saving to BSON or JLD file I get this error
ERROR: UndefRefError: access to undefined reference

I also mentioned this here
JuliaIO/BSON.jl#3 (comment)

[kafisatz]

MWE

using DataFrames
using CSV
using BSON

df = DataFrame(a=[1,2],b=[missing,"mi"])
fi=mktemp()[1]
CSV.write(fi,df);
dfr = CSV.read(fi,DataFrame,pool=false);
@assert isa(dfr[2],CSV.SentinelArrays.SentinelArray{String,1,UndefInitializer,Missing,Array{String,1}})

isfile(fi)&&rm(fi)
BSON.@save fi dfr

[quinnj]

By default, any "reference" type will use #undef as the sentinel value underneath, so SentinelArray([missing, "hey"]) will show that the array type is Union{Missing, String}, but it's actually using #undef under the hood to signal missing. Seems like you already found a work-around for the BSON.jl issue of not being able to handle #undef?

[kafisatz]

No. I do not have a workaround.
More interestingly the error also happens when there are no missing entries. It seems the error is related to the column type rather than the occurrence of missing entries.

[quinnj]

Yes, that is interesting.

I was confused when you said:

this works for me (for the time being) :
dfraw.Instanz = convert(Array{Union{Missing,String},1},dfraw.Instanz);

does that not work as a workaround?

[kafisatz]

Yes it does work.