Bug in read_csv with duplicated column names

[rafaljozefowicz]

Tested on 0.13.0, 0.13.1 and 0.14.0rc1:

from StringIO import StringIO
import pandas as pd

# this is correct
print(pd.DataFrame([[0, 1, 2], [3, 4, 5]], columns=["a", "b", "a"]))
# and this is fine as well
# (although, changing the column names to a,b,a.1)
print(pd.read_csv(StringIO("a,b,a\n0,1,2\n3,4,5")))
# but this is not correct
print(pd.read_csv(StringIO("0,1,2\n3,4,5"), names=["a", "b", "a"]))

The last one returns:

Out[5]: 
   a  b  a
0  2  1  2
1  5  4  5

I would expect all 3 methods to return the same DataFrame. I noticed this when I wanted to read csv file that had a separate file with a header (and a duplicated column in it). BTW is there a better way to do it than to read the header file first and pass the output into 'names' parameter of read_csv?

[jreback]

read_csv needs to interpret many formats so that's why it changes to not have duplicate columns
(and some code that assumes they r unique)

so this needs some work

marking as a bug for 0.15 - feel free to submit a pr

[gfyoung]

@jreback : This bug still exists in master. What sort of output should be expected from this? Is the second one supposed to give the exact same output as the first and third (when patched) ones?

[jreback]

yeh I think the problem is that the names->columns are passed back as a dict and not as 2 lists, so it gets lots. Its in the post-processing code in python somewhere.

I would expect as the OP suggests. Note tthis is different that if usecols has dupes. names is merely acting as the header here.

[gfyoung]

FYI: for the second example, that output is correct because mangle_dupe_cols defaults to True, meaning that all columns become unique with .{x} labelling as expected. The bug also surfaces if you set mangle_dupe_cols=False.

[gfyoung]

@jreback : Question, what does the as_recarray option do exactly? And, should we be raising a ValueError (as we do now in master) in the third example if we set as_recarray=True?

[jreback]

as_recarray returns a numpy rec-array. But AFAIK its pretty much unused / not tested very well / somewhat buggy. Originally for numpy compat (and of course would be important if we eventually extracted the read_csv as a separate indepedent module (so that numpy could use it directly).