Using Python Iterparse For Large XML Files

I need to write a parser in Python that can process some extremely large files ( > 2 GB ) on a computer without much memory (only 2 GB). I wanted to use iterparse in lxml to do it.

My file is of the format:

<item>
  <title>Item 1</title>
  <desc>Description 1</desc>
</item>
<item>
  <title>Item 2</title>
  <desc>Description 2</desc>
</item>

and so far my solution is:

from lxml import etree

context = etree.iterparse( MYFILE, tag='item' )

for event, elem in context :
      print elem.xpath( 'description/text( )' )

del context

Unfortunately though, this solution is still eating up a lot of memory. I think the problem is that after dealing with each "ITEM" I need to do something to cleanup empty children. Can anyone offer some suggestions on what I might do after processing my data to properly cleanup?


Solution 1:

Try Liza Daly's fast_iter. After processing an element, elem, it calls elem.clear() to remove descendants and also removes preceding siblings.

def fast_iter(context, func, *args, **kwargs):
    """
    http://lxml.de/parsing.html#modifying-the-tree
    Based on Liza Daly's fast_iter
    http://www.ibm.com/developerworks/xml/library/x-hiperfparse/
    See also http://effbot.org/zone/element-iterparse.htm
    """
    for event, elem in context:
        func(elem, *args, **kwargs)
        # It's safe to call clear() here because no descendants will be
        # accessed
        elem.clear()
        # Also eliminate now-empty references from the root node to elem
        for ancestor in elem.xpath('ancestor-or-self::*'):
            while ancestor.getprevious() is not None:
                del ancestor.getparent()[0]
    del context


def process_element(elem):
    print elem.xpath( 'description/text( )' )

context = etree.iterparse( MYFILE, tag='item' )
fast_iter(context,process_element)

Daly's article is an excellent read, especially if you are processing large XML files.


Edit: The fast_iter posted above is a modified version of Daly's fast_iter. After processing an element, it is more aggressive at removing other elements that are no longer needed.

The script below shows the difference in behavior. Note in particular that orig_fast_iter does not delete the A1 element, while the mod_fast_iter does delete it, thus saving more memory.

import lxml.etree as ET
import textwrap
import io

def setup_ABC():
    content = textwrap.dedent('''\
      <root>
        <A1>
          <B1></B1>
          <C>1<D1></D1></C>
          <E1></E1>
        </A1>
        <A2>
          <B2></B2>
          <C>2<D></D></C>
          <E2></E2>
        </A2>
      </root>
        ''')
    return content


def study_fast_iter():
    def orig_fast_iter(context, func, *args, **kwargs):
        for event, elem in context:
            print('Processing {e}'.format(e=ET.tostring(elem)))
            func(elem, *args, **kwargs)
            print('Clearing {e}'.format(e=ET.tostring(elem)))
            elem.clear()
            while elem.getprevious() is not None:
                print('Deleting {p}'.format(
                    p=(elem.getparent()[0]).tag))
                del elem.getparent()[0]
        del context

    def mod_fast_iter(context, func, *args, **kwargs):
        """
        http://www.ibm.com/developerworks/xml/library/x-hiperfparse/
        Author: Liza Daly
        See also http://effbot.org/zone/element-iterparse.htm
        """
        for event, elem in context:
            print('Processing {e}'.format(e=ET.tostring(elem)))
            func(elem, *args, **kwargs)
            # It's safe to call clear() here because no descendants will be
            # accessed
            print('Clearing {e}'.format(e=ET.tostring(elem)))
            elem.clear()
            # Also eliminate now-empty references from the root node to elem
            for ancestor in elem.xpath('ancestor-or-self::*'):
                print('Checking ancestor: {a}'.format(a=ancestor.tag))
                while ancestor.getprevious() is not None:
                    print(
                        'Deleting {p}'.format(p=(ancestor.getparent()[0]).tag))
                    del ancestor.getparent()[0]
        del context

    content = setup_ABC()
    context = ET.iterparse(io.BytesIO(content), events=('end', ), tag='C')
    orig_fast_iter(context, lambda elem: None)
    # Processing <C>1<D1/></C>
    # Clearing <C>1<D1/></C>
    # Deleting B1
    # Processing <C>2<D/></C>
    # Clearing <C>2<D/></C>
    # Deleting B2

    print('-' * 80)
    """
    The improved fast_iter deletes A1. The original fast_iter does not.
    """
    content = setup_ABC()
    context = ET.iterparse(io.BytesIO(content), events=('end', ), tag='C')
    mod_fast_iter(context, lambda elem: None)
    # Processing <C>1<D1/></C>
    # Clearing <C>1<D1/></C>
    # Checking ancestor: root
    # Checking ancestor: A1
    # Checking ancestor: C
    # Deleting B1
    # Processing <C>2<D/></C>
    # Clearing <C>2<D/></C>
    # Checking ancestor: root
    # Checking ancestor: A2
    # Deleting A1
    # Checking ancestor: C
    # Deleting B2

study_fast_iter()

Solution 2:

iterparse() lets you do stuff while building the tree, that means that unless you remove what you don't need anymore, you'll still end up with the whole tree in the end.

For more information: read this by the author of the original ElementTree implementation (but it's also applicable to lxml)

Solution 3:

Why won't you use the "callback" approach of sax?

Solution 4:

In my experience, iterparse with or without element.clear (see F. Lundh and L. Daly) cannot always cope with very large XML files: It goes well for some time, suddenly the memory consumption goes through the roof and a memory error occurs or the system crashes. If you encounter the same problem, maybe you can use the same solution: the expat parser. See also F. Lundh or the following example using OP’s XML snippet (plus two umlaute for checking that there are no encoding issues):

import xml.parsers.expat
from collections import deque

def iter_xml(inpath: str, outpath: str) -> None:
    def handle_cdata_end():
        nonlocal in_cdata
        in_cdata = False

    def handle_cdata_start():
        nonlocal in_cdata
        in_cdata = True

    def handle_data(data: str):
        nonlocal in_cdata
        if not in_cdata and open_tags and open_tags[-1] == 'desc':
            data = data.replace('\\', '\\\\').replace('\n', '\\n')
            outfile.write(data + '\n')

    def handle_endtag(tag: str):
        while open_tags:
            open_tag = open_tags.pop()
            if open_tag == tag:
                break

    def handle_starttag(tag: str, attrs: 'Dict[str, str]'):
        open_tags.append(tag)

    open_tags = deque()
    in_cdata = False
    parser = xml.parsers.expat.ParserCreate()
    parser.CharacterDataHandler = handle_data
    parser.EndCdataSectionHandler = handle_cdata_end
    parser.EndElementHandler = handle_endtag
    parser.StartCdataSectionHandler = handle_cdata_start
    parser.StartElementHandler = handle_starttag
    with open(inpath, 'rb') as infile:
        with open(outpath, 'w', encoding = 'utf-8') as outfile:
            parser.ParseFile(infile)

iter_xml('input.xml', 'output.txt')

input.xml:

<root>
    <item>
    <title>Item 1</title>
    <desc>Description 1ä</desc>
    </item>
    <item>
    <title>Item 2</title>
    <desc>Description 2ü</desc>
    </item>
</root>

output.txt:

Description 1ä
Description 2ü