Huge CSV and XML Files in Python

I, like most people, never realized I'd be dealing with large files. Oh, I knew there would be some files with megabytes of data, but I never suspected I'd be begging Perl to process hundreds of megabytes of XML, nor that this week I'd be asking Python to process 6.4 gigabytes of CSV into 6.5 gigabytes of XML¹.

As a few out-of-memory experiences will teach you, the trick for dealing with large files is pretty easy: use code that treats everything as a stream. For inputs, read from disk in chunks. For outputs, frequently write to disk and let system memory forge onward unburdened.

When reading and writing files yourself, this is easier to do correctly...

from __future__ import with_statement # for python 2.5
with open('data.in','r') as fin:
with open('data.out','w') as fout:
for line in fin:
fout.write(','.join(line.split(' ')))

...than it is to do incorrectly...

with open('data.in','r') as fin:
    data = fin.read()
data2 = [ ','.join(x.split(' ')) for x in data ]
with open('data.out','w') as fout:
fout.write(data2)

...at least in simple cases.

Loading Large CSV Files in Python

Python has an excellent csv library, which can handle large files right out of the box. Sort of.

>> import csv
>> r = csv.reader(open('doc.csv', 'rb'))
>>> for row in r:
...     print row
... 
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
_csv.Error: field larger than field limit (131072)

Staring at the module documentation², I couldn't find anything of use. So I cracked open the csv.py file and confirmed what the _csv in the error message suggests: the bulk of the module's code (and the input parsing in particular) is implemented in C rather than Python.

After a while staring at that error, I began dreaming of how I would create a stream pre-processor using StringIO, but it didn't take too long to figure out I would need to recreate my own version of csv in order to accomplish that.

So back to the blogs, one of which held the magic grain of information I was looking for: csv.field_size_limit.

>>> import csv
>>> csv.field_size_limit()
131072
>>> csv.field_size_limit(1000000000)
131072
>>> csv.field_size_limit()
1000000000

Yep. That's all there is to it. The sucker just works after that.

Well, almost. I did run into an issue with a NULL byte 1.5 gigs into the data. Because the streaming code is written using C based IO, the NULL byte shorts out the reading of data in an abrupt and non-recoverable manner. To get around this we need to pre-process the stream somehow, which you could do in Python by wrapping the file with a custom class that cleans each line before returning it, but I went with some command line utilities for simplicity.

cat data.in | tr -d '\0' > data.out

After that, the 6.4 gig CSV file processed without any issues.

Creating Large XML Files in Python

This part of the process, taking each row of csv and converting it into an XML element, went fairly smoothly thanks to the xml.sax.saxutils.XMLGenerator class. The API for creating elements isn't an example of simplicity, but it is--unlike many of the more creative schemes--predictable, and has one killer feature: it correctly writes output to a stream.

As I mentioned, the mechanism for creating elements was a bit verbose, so I made a couple of wrapper functions to simplify (note that I am sending output to standard out, which lets me simply print strings to the file I am generating, for example creating the XML file's version declaration).

import sys
from xml.sax.saxutils import XMLGenerator
from xml.sax.xmlreader import AttributesNSImpl
g = XMLGenerator(sys.stdout, 'utf-8')
def start_tag(name, attr={}, body=None, namespace=None):
attr_vals = {}
attr_keys = {}
for key, val in attr.iteritems():
key_tuple = (namespace, key)
attr_vals[key_tuple] = val
attr_keys[key_tuple] = key
<span class="n">attr2</span> <span class="o">=</span> <span class="n">AttributesNSImpl</span><span class="p">(</span><span class="n">attr_vals</span><span class="p">,</span> <span class="n">attr_keys</span><span class="p">)</span>
<span class="n">g</span><span class="o">.</span><span class="n">startElementNS</span><span class="p">((</span><span class="n">namespace</span><span class="p">,</span> <span class="n">name</span><span class="p">),</span> <span class="n">name</span><span class="p">,</span> <span class="n">attr2</span><span class="p">)</span>
<span class="k">if</span> <span class="n">body</span><span class="p">:</span>
    <span class="n">g</span><span class="o">.</span><span class="n">characters</span><span class="p">(</span><span class="n">body</span><span class="p">)</span>
def end_tag(name, namespace=None):
g.endElementNS((namespace, name), name)
def tag(name, attr={}, body=None, namespace=None):
start_tag(name, attr, body, namespace)
end_tag(name, namespace)

From there, usage looks like this:

print """<?xml version="1.0" encoding="utf-8'?>"""
start_tag(u'list', {u'id':10})
for item in some_list:
start_tag(u'item', {u'id': item[0]})
tag(u'title', body=item[1])
tag(u'desc', body=item[2])
end_tag(u'item')
end_tag(u'list')
g.endDocument()

The one issue I did run into (in my data) was some pagebreak characters floating around (^L aka 12 aka x0c) which were tweaking the XML encoder, but you can strip them out in a variety of places, for example by rewriting the main loop:

for item in some_list:
    item = [ x.replace('\x0c','') for x in item ]
    # etc

Really, the XMLGenerator just worked, even when dealing with a quite large file.

Performance

Although my script created a different mix of XML elements than the above example, it wasn't any more complex, and had fairly reasonable performance. Processing of the 6.4 gig CSV file into a 6.5 gig XML file took between 19 - 24 minutes, which means it was able to read-process-write about five megabytes per second.

In terms of raw speed, that isn't particularly epic, but performing a similar operation (was actually XML to XML rather than CSV to XML) with Perl's XML::Twig it took eight minutes to process a ~100 megabyte file, so I'm pretty pleased with the quality of the Python standard library and how it handles large files.

The breadth and depth of the standard library really makes Python a joy to work with for these simple one-shot scripts. If only it had Perl's easier to use regex syntax...