Tables.dictcolumntable produces columns with #undef instead of missing in struct-valued columns

[kleinschmidt]

julia> using Tables

julia> struct SSS
           stuff
       end

julia> Tables.dictcolumntable([(; a=SSS(1)), (; b=2)]).a
2-element Vector{Union{Missing, SSS}}:
    SSS(1)
 #undef

julia> Tables.columns(Tables.dictrowtable([(; a=SSS(1)), (; b=2)]))
Tables.CopiedColumns{NamedTuple{(:a, :b), Tuple{Vector{Union{Missing, SSS}}, Vector{Union{Missing, Int64}}}}} with 2 rows, 2 columns, and schema:
 :a  Union{Missing, SSS}
 :b  Union{Missing, Int64}

julia> Tables.columns(Tables.dictrowtable([(; a=SSS(1)), (; b=2)])).a
2-element Vector{Union{Missing, SSS}}:
 SSS(1)
 missing

julia> Tables.dictcolumntable([(; a=1), (; b=2)]).a
2-element Vector{Union{Missing, Int64}}:
 1
  missing

[nalimilan]

Looks like dictcolumntable assumes that missing is used instead of #undef, but that's only the case for bitstypes. And even worse, singleton types may be used as the default instead of missing, giving misleading results. So better initialize all the time the entry with missing.

[quinnj]

Sorry for the slow response: #288