Integer column randomly typed as string when threading enabled

[klaff]

The following code:

using CSV
using DataFrames
using Random

NCOLS = 30
NROWS = 150
Random.seed!(1)
fname = tempname()
f = open(fname, "w")
write(f, join("col".*string.(1:NCOLS), ","))
write(f, "\r\n")
for i in 0:NROWS
	write(f, join(string.(rand(Int16, NCOLS)), ","))
	write(f, "\r\n")
end
close(f)
df_by_threads = CSV.read(fname, DataFrame)
df_single_threaded = CSV.read(fname, DataFrame; ntasks=1)
print(eltype.(eachcol(df_by_threads)) == eltype.(eachcol(df_single_threaded)))

will print false , or at least it does on my Windows box, with 8 threads, running CSV v0.10.7.

If I reduce NCOLS or NROWS it will be true. If I choose a different random seed it may become true.

If one inspects the columns of df_by_threads, at least one column will be of type String7, but which column may vary with repeated execution and sometimes there are two such columns, even though the data written to the file is fixed.