Regression in serializing + deserializing union vectors introduced in 0.5.6

[KristofferC]

The following code gives different behavior on JLD2 0.5.5 vs 0.5.6:

using JLD2, DataFrames, Dates

df = DataFrame(date=Date(2014),boolcol=rand([true false],100),randcol=rand(100))
allowmissing!(df)

df[rand(1:100,10),:boolcol] .= missing
df[rand(1:100,10),:randcol] .= missing

jldopen("test.jld2","w") do f
    for n in names(df)
        f[n] = df[!,n]
    end
end

@show eltype(df.boolcol)


df2 = DataFrame()

jldopen("test.jld2","r") do f
    for k in keys(f)
        df2[!,k] = f[k]
    end
end

@show eltype(df2.boolcol)

On 0.5.5 this gives:

eltype(df.boolcol) = Union{Missing, Bool}
eltype(df2.boolcol) = Union{Missing, Bool}

while on 0.5.6 it gives:

eltype(df.boolcol) = Union{Missing, Bool}
eltype(df2.boolcol) = JLD2.InlineUnionEl{Missing, Bool}

So it seems some internal JLD2 type is now getting leaked into the constructed values.

[JonasIsensee]

Whoops, thanks for the report.
I'll have a look.

[JonasIsensee]

Version v0.5.9 with a fix should be out later today.