DAWG 0.5

DAWG and CompletionDAWG

DAWG and CompletionDAWG are useful when you need fast & memory efficient simple string storage. These classes does not support assigning values to keys.

DAWG and CompletionDAWG constructors accept an iterable with keys:

>>> import dawg
>>> words = [u'foo', u'bar', u'foobar', u'foö', u'bör']
>>> base_dawg = dawg.DAWG(words)
>>> completion_dawg = dawg.CompletionDAWG(words)

It is then possible to check if the key is in a DAWG:

>>> u'foo' in base_dawg
True
>>> u'baz' in completion_dawg
False

It is possible to find all keys that starts with a given prefix in a CompletionDAWG:

>>> completion_dawg.keys(u'foo')
>>> [u'foo', u'foobar']

and to find all prefixes of a given key:

>>> base_dawg.prefixes(u'foobarz')
[u'foo', u'foobar']

Iterator versions are also available:

>>> for key in completion_dawg.iterkeys(u'foo'):
...     print(key)
foo
foobar
>>> for prefix in base_dawg.iterprefixes(u'foobarz'):
...     print(prefix)
foo
foobar

It is possible to find all keys similar to a given key (using a one-way char translation table):

>>> replaces = dawg.DAWG.compile_replaces({u'o': u'ö'})
>>> base_dawg.similar_keys(u'foo', replaces)
[u'foo', u'foö']
>>> base_dawg.similar_keys(u'foö', replaces)
[u'foö']
>>> base_dawg.similar_keys(u'bor', replaces)
[u'bör']

BytesDAWG

BytesDAWG is a CompletionDAWG subclass that can store binary data for each key.

BytesDAWG constructor accepts an iterable with (unicode_key, bytes_value) tuples:

>>> data = [(u'key1', b'value1'), (u'key2', b'value2'), (u'key1', b'value3')]
>>> bytes_dawg = dawg.BytesDAWG(data)

There can be duplicate keys; all unique values are stored in this case:

>>> bytes_dawg[u'key1']
[b'value1, b'value3']

For unique keys a list with a single value is returned for consistency:

>>> bytes_dawg[u'key2']
[b'value2']

KeyError is raised for missing keys; use get method if you need a default value instead:

>>> bytes_dawg.get(u'foo', None)
None

BytesDAWG support keys, items, iterkeys and iteritems methods (they all accept optional key prefix). There is also support for similar_keys, similar_items and similar_item_values methods.

RecordDAWG

RecordDAWG is a BytesDAWG subclass that automatically packs & unpacks the binary data from/to Python objects using struct module from the standard library.

First, you have to define a format of the data. Consult Python docs (http://docs.python.org/library/struct.html#format-strings) for the format string specification.

For example, let’s store 3 short unsigned numbers (in a Big-Endian byte order) as values:

>>> format = ">HHH"

RecordDAWG constructor accepts an iterable with (unicode_key, value_tuple). Let’s create such iterable using zip function:

>>> keys = [u'foo', u'bar', u'foobar', u'foo']
>>> values = [(1, 2, 3), (2, 1, 0), (3, 3, 3), (2, 1, 5)]
>>> data = zip(keys, values)
>>> record_dawg = RecordDAWG(format, data)

As with BytesDAWG, there can be several values for the same key:

>>> record_dawg['foo']
[(1, 2, 3), (2, 1, 5)]
>>> record_dawg['foobar']
[(3, 3, 3)]

BytesDAWG and RecordDAWG implementation details

BytesDAWG and RecordDAWG stores data at the end of the keys:

<utf8-encoded key><separator><base64-encoded data>

Data is encoded to base64 because dawgdic C++ library doesn’t allow zero bytes in keys (it uses null-terminated strings) and such keys are very likely in binary data.

In DAWG versions prior to 0.5 <separator> was chr(255) byte. It was chosen because keys are stored as UTF8-encoded strings and chr(255) is guaranteed not to appear in valid UTF8, so the end of text part of the key is not ambiguous.

But chr(255) was proven to be problematic: it changes the order of the keys. Keys are naturally returned in lexicographical order by DAWG. But if chr(255) appears at the end of each text part of a key then the visible order would change. Imagine 'foo' key with some payload and 'foobar' key with some payload. 'foo' key would be greater than 'foobar' key: values compared would be 'foo<sep>' and 'foobar<sep>' and ord(<sep>)==255 is greater than ord(<any other character>).

So now the default <separator> is chr(1). This is the lowest allowed character and so it preserves the alphabetical order.

It is not strictly correct to use chr(1) as a separator because chr(1) is a valid UTF8 character. But I think in practice this won’t be an issue: such control character is very unlikely in text keys, and binary keys are not supported anyway because dawgdic doesn’t support keys containing chr(0).

If you can’t guarantee chr(1) is not a part of keys, lexicographical order is not important to you or there is a need to read a BytesDAWG/RecordDAWG created by DAWG < 0.5 then pass payload_separator argument to the constructor:

>>> BytesDAWG(payload_separator=b'\xff').load('old.dawg')

The storage scheme has one more implication: values of BytesDAWG and RecordDAWG are also sorted lexicographically.

For RecordDAWG there is a gotcha: in order to have meaningful ordering of numeric values store them in big-endian format:

>>> data = [('foo', (3, 2, 256)), ('foo', (3, 2, 1)), ('foo', (3, 2, 3))]
>>> d = RecordDAWG("3H", data)
>>> d.items()
[(u'foo', (3, 2, 256)), (u'foo', (3, 2, 1)), (u'foo', (3, 2, 3))]

>>> d2 = RecordDAWG(">3H", data)
>>> d2.items()
[(u'foo', (3, 2, 1)), (u'foo', (3, 2, 3)), (u'foo', (3, 2, 256))]

IntDAWG

IntDAWG is a {unicode -> int} mapping. It is possible to use RecordDAWG for this, but IntDAWG is natively supported by dawgdic C++ library and so __getitem__ is much faster.

Unlike BytesDAWG and RecordDAWG, IntDAWG doesn’t support having several values for the same key.

IntDAWG constructor accepts an iterable with (unicode_key, integer_value) tuples:

>>> data = [ (u'foo', 1), (u'bar', 2) ]
>>> int_dawg = dawg.IntDAWG(data)

It is then possible to get a value from the IntDAWG:

>>> int_dawg[u'foo']
1

Persistence

All DAWGs support saving/loading and pickling/unpickling.

Write DAWG to a stream:

>>> with open('words.dawg', 'wb') as f:
...     d.write(f)

Save DAWG to a file:

>>> d.save('words.dawg')

Load DAWG from a file:

>>> d = dawg.DAWG()
>>> d.load('words.dawg')

Read DAWG from a stream:

>>> d = dawg.RecordDAWG(format_string)
>>> with open('words.record-dawg', 'rb') as f:
...     d.read(f)

DAWG objects are picklable:

>>> import pickle
>>> data = pickle.dumps(d)
>>> d2 = pickle.loads(data)

How is source code organized

There are 4 folders in repository:

• bench - benchmarks & benchmark data;

• lib - original unmodified dawgdic C++ library and a customized version of libb64 library. They are bundled for easier distribution; if something is have to be fixed in these libraries consider fixing it in the original repositories;

• src - wrapper code; src/dawg.pyx is a wrapper implementation; src/*.pxd files are Cython headers for corresponding C++ headers; src/*.cpp files are the pre-built extension code and shouldn’t be modified directly (they should be updated via update_cpp.sh script).

• tests - the test suite.

Running tests and benchmarks

Make sure tox is installed and run

$ tox

from the source checkout. Tests should pass under python 2.6, 2.7, 3.2 and 3.3.

In order to run benchmarks, type

$ tox -c bench.ini

Authors & Contributors

• Mikhail Korobov <kmike84@gmail.com>;

• Dan Blanchard.

This module is based on dawgdic C++ library by Susumu Yata & contributors.

base64 decoder is based on libb64 by Chris Venter.

0.5 (2012-10-08)

The storage scheme of BytesDAWG and RecordDAWG is changed in this release in order to provide the alphabetical ordering of items.

This is a backwards-incompatible release. In order to read BytesDAWG or RecordDAWG created with previous versions of DAWG use payload_separator constructor argument:

>>> BytesDAWG(payload_separator=b'\xff').load('old.dawg')

0.4.1 (2012-10-01)

• Segfaults with empty DAWGs are fixed by updating dawgdic to latest svn.

0.4 (2012-09-26)

• iterkeys, iteritems and iterprefixes methods (thanks Dan Blanchard).

0.3.2 (2012-09-24)

• prefixes method for finding all prefixes of a given key.

0.3.1 (2012-09-20)

• bundled dawgdic C++ library is updated to the latest version.

0.3 (2012-09-13)

• similar_keys, similar_items and similar_item_values methods for more permissive lookups (they may be useful e.g. for umlaut handling);

• load method returns self;

• Python 3.3 support.

0.2 (2012-09-08)

Greatly improved memory usage for DAWGs loaded with load method.

There is currently a bug somewhere in a wrapper so DAWGs loaded with read() method or unpickled DAWGs uses 3x-4x memory compared to DAWGs loaded with load() method. load() is fixed in this release but other methods are not.

0.1 (2012-09-08)

Initial release.