BUG: python 3 compression and read_fwf

[TomAugspurger]

Getting a TypeError: Type str doesn't support the buffer API when using read_fwf on a compressed file with python 3 only; works in python 2.7

Example:

with a file fwf.bug.txt like

1111111111
2222222222
3333333333
4444444444

import pandas as pd

print(pd.__version__)

widths = [5, 5]

!gzip -d fwf_bug.txt

df = pd.read_fwf('fwf_bug.txt', widths=widths, names=['one', 'two'])
print(df)

!gzip fwf_bug.txt

# python 3 throws an error here.
df2 = pd.read_fwf('fwf_bug.txt.gz', widths=widths,
                  names=['one', 'two'], compression='gzip')

print(df2)

Versions:
python3: 0.11.1.dev-4d06037 should be most recent
python2: 0.11.1.dev-3ebfef9

I can paste the full traceback if you'd like.

[TomAugspurger]

Doing a bit of digging around, looks like it's a unicode thing.

The error is coming here, line 1919 in pandas/build/lib.macosx-10.8-x86_64-3.3/pandas/io/parsers.py:

    def next(self):
        line = next(self.f)
        # Note: 'colspecs' is a sequence of half-open intervals.
        return [line[fromm:to].strip(self.filler or ' ')
                for (fromm, to) in self.colspecs]

here, line is a bytecode string:

ipdb> line
b'1111111111\n'

I'm not sure what the preferred way of dealing with this is, but

ipdb> line.decode('utf-8')[fromm:to].strip(' ')
'11111'

works.

[cpcloud]

should probably be

import pandas.core.common as com
com.pprint_thing(line[fromm:to]).strip(self.filler or ' ')

[jtratner]

basically, the problem is that you can't mix bytes and str in Python 3:

In [12]: b'abcd'.strip('')
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-12-e487f65c9b72> in <module>()
----> 1 b'abcd'.strip('')

TypeError: Type str doesn't support the buffer API

In [13]: # Whereas it works with bytes

In [14]: b'adbc'.strip(b'')
Out[14]: b'adbc'

I'm reopening, since after staring at #4784 for a bit I think it's (and my +1 of it) wrong.

The use of next(f) when f is a BytesIO object seem dodgy,
and since next(f) doesn't strip the newline (it can't reliably), the decoding
may fail. I'm also not sure that the use of strip() is well-defined here either.

An example to illustrate some of this:

#!/usr/bin/env python3.3
from io import BytesIO
from encodings.aliases import aliases

for enc in set(aliases.values()):
    try:
        # print(enc, next(BytesIO("1234\nabcd".encode(enc)))[:-1].decode(enc)=='1234')
        bs=BytesIO("1234\nabcd".encode(enc))
        line=next(bs)
        res=line.strip().decode(enc)
        if res!='1234':
            print(enc, res)
    except LookupError:
        pass
    except UnicodeDecodeError:
        print("%s failed" % enc)

I think TextIOWrapper is the correct solution here, you can't get lines until you have text

[jtratner]

@y-p TextIOWrapper is what I tried first (in #4783) and it works perfectly in 3.3 (though we'd need to edit it to use the specified encoding if and only if one is provided). However, bz2 doesn't play nice with it because it doesn't support a read1() method. (gzip is fine).

Two options:

1. Make a subclass of io.TextIOWrapper that detects whether the passed buffer defines read1() and calls read() instead if it's not defined.
2. Special case bz2 and create a wrapper class that proxies everything to the internal bz2 reader, but calls read() when read1() is asked for.

Either way would work - (1) is probably more explicit though. If bz2 is the only place where we'll need this, probably makes more sense to do (2)...

So you did, didn't see it earlier.

I would be fine with correct code that only works on 3.3 and raising an error (or just blowing up) otherwise.
Building a compat layer for 3.2 seems like wasted effort.