pycedar

Python binding of cedar (implementation of efficiently-updatable double-array trie) using Cython

日本語版の README は README.ja.md にあります。

Official URL of cedar: http://www.tkl.iis.u-tokyo.ac.jp/~ynaga/cedar/

Requirements

• Python 3.9 or newer
• A 64-bit platform: Windows, or a POSIX-compatible one (Linux, macOS)
• A C++ compiler, when building from source (wheels need none)

The extension is tested with CPython 3.9 through 3.15 on Linux, macOS and Windows, including the free-threaded builds of 3.14 and 3.15 (3.14t, 3.15t). Wheels cover Linux (x86_64, aarch64), macOS (x86_64, arm64) and Windows (AMD64, ARM64). The module is declared free-threading compatible, so it imports and works on a free-threaded interpreter without a ModuleNotFoundError; concurrency across distinct trie objects is supported (see Limitations).

Why a double-array trie?

A trie encodes the keys into its arrays, so it does not keep a separate string object per key, and it answers prefix queries that a hash map cannot.

200000 random lowercase keys of 4 to 16 characters, mapped to integers, measured with benchmarks/bench.py on CPython 3.13 (Linux, x86_64):

The trade is real and worth stating plainly: a point lookup costs roughly twice what a dict lookup does (30 ns against 16 ns on the same data; the numbers vary by machine). Reach for pycedar when the key set is large enough that memory matters, or when you need prefix search; reach for a dict when you only need point lookups.

Run python benchmarks/bench.py --help to reproduce these numbers on your own data.

Installation

install from PyPI release

$ pip install --user pycedar

install from GitHub master

$ pip install --user https://github.com/akivajp/pycedar/archive/master.zip

Type checking

Type stubs ship as the pycedar-stubs package (PEP 561). Type checkers and IDEs pick them up automatically once pycedar is installed — no extra step.

Usage

using python-like dict class based on double array trie

>>> import pycedar

>>> d = pycedar.dict()
>>> len(d)
0
>>> bool(d)
False
>>> list(d)
[]

>>> d['nineteen'] = 19
>>> d.set('twenty', 20)
20
>>> d['twenty one'] = 21
>>> d['twenty two'] = 22
>>> d['twenty three'] = 23
>>> d['twenty four'] = 24

>>> len(d)
6
>>> bool(d)
True
>>> list(d)
['nineteen', 'twenty', 'twenty four', 'twenty one', 'twenty three', 'twenty two']
>>> list(d.keys())
['nineteen', 'twenty', 'twenty four', 'twenty one', 'twenty three', 'twenty two']
>>> list(d.values())
[19, 20, 24, 21, 23, 22]
>>> list(d.items())
[('nineteen', 19), ('twenty', 20), ('twenty four', 24), ('twenty one', 21), ('twenty three', 23), ('twenty two', 22)]
>>> d['twenty four']
24
>>> 'twenty four' in d
True
>>> del d['twenty four']
>>> 'twenty four' in d
False
>>> d['twenty four']
Traceback (most recent call last):
    ...
KeyError: 'twenty four'
>>> d.get('twenty three')
23
>>> d.get('twenty four')       # the default is base_trie.NO_VALUE
-1
>>> d.get('twenty four', None) is None
True

Prefix queries return generators:

>>> list(d.find(''))
[('nineteen', 19), ('twenty', 20), ('twenty one', 21), ('twenty three', 23), ('twenty two', 22)]
>>> list(d.find('tw'))
[('twenty', 20), ('twenty one', 21), ('twenty three', 23), ('twenty two', 22)]
>>> list(d.find('twenty t'))
[('twenty three', 23), ('twenty two', 22)]
>>> list(d.find_keys('twenty'))
['twenty', 'twenty one', 'twenty three', 'twenty two']
>>> list(d.find_values('twenty'))
[20, 21, 23, 22]

Nodes let you keep a position in the trie and search relative to it:

>>> n = d.get_node('twenty')
>>> n.key()
'twenty'
>>> n.value()
20
>>> [child.key() for child in n.find_nodes(' t')]
[' three', ' two']

>>> d.get_node('twenty ') is None    # a path without a value
True

Saving and loading:

>>> d.save('test.dat')
0
>>> d2 = pycedar.dict()
>>> d2.setdefault('eighteen', 18)
18
>>> list(d2.items())
[('eighteen', 18)]
>>> d2.load('test.dat')              # replaces the whole trie image
0
>>> list(d2.items())
[('nineteen', 19), ('twenty', 20), ('twenty one', 21), ('twenty three', 23), ('twenty two', 22)]
>>> d2.setdefault('eighteen', 18)
18
>>> list(d2.items())
[('eighteen', 18), ('nineteen', 19), ('twenty', 20), ('twenty one', 21), ('twenty three', 23), ('twenty two', 22)]

The image can also be kept in memory as bytes, and the dict supports pickling and copying:

>>> import pickle, copy

>>> image = d.dumps()
>>> restored = pickle.loads(pickle.dumps(d))
>>> restored['twenty']
20
>>> clone = copy.deepcopy(d)             # copy.copy works the same way
>>> clone['twenty'] = 999
>>> d['twenty']
20

>>> d.pop('twenty two')                  # like dict.pop
22
>>> d.pop('twenty two', 'gone')
'gone'
>>> d.popitem()                          # removes the first key in sorted order
('nineteen', 19)

popitem() returns the lexicographically smallest remaining key, because a trie keeps no insertion order — this is the one place pycedar.dict intentionally differs from dict.

using bytes keys

pycedar.dict is parameterised by its key type, which is enforced strictly:

>>> b = pycedar.dict(bytes)
>>> b[b'cedar'] = 1
>>> b[b'cedarpp'] = 2
>>> list(b.items())
[(b'cedar', 1), (b'cedarpp', 2)]
>>> b['str key'] = 3
Traceback (most recent call last):
    ...
TypeError: Argument 'key' has incorrect type (expected bytes, got str)

str keys are encoded as UTF-8 before they reach cedar, so lengths reported by the low level API are byte lengths, not character counts.

using more primitive data structures

pycedar.dict is a thin convenience layer over the trie classes. You can use them directly when you want cedar's raw semantics.

>>> t = pycedar.str_trie()
>>> t.set('apple', 1)
1
>>> t.set('applet', 2)
2
>>> t.set('apply', 3)
3

>>> t.exact_match_search('apple')      # (value, length, node id)
(1, 5, 259)
>>> t.exact_match_search('app')[0]     # a prefix carries no value
-1

>>> t.common_prefix_search('applet')   # every key that prefixes the query
[('apple', 1, 259), ('applet', 2, 368)]
>>> [key for key, value, node_id in t.common_prefix_predict('app')]
['le', 'let', 'ly']

>>> t.erase('apply')
0
>>> t.erase('apply')                   # already gone
-1
>>> t.num_keys()
2

Note the difference between the two prefix queries:

• common_prefix_search(key) returns the complete keys that are prefixes of key.
• common_prefix_predict(key) returns the remaining suffixes of every key that starts with key.

Both also have lazy variants that yield one triple at a time instead of building a list, which matters when a query can match a large part of the trie or when the first hits are all you need:

>>> [key for key, value, node_id in t.icommon_prefix_predict('app')]   # 'apply' was erased above
['le', 'let']
>>> list(t.icommon_prefix_search('applet'))
[('apple', 1, 259), ('applet', 2, 368)]

The lazy variants return the same triples in the same order as the list versions; results are not guaranteed if the trie is modified while iterating. Whether laziness pays depends on how early you stop. Measured with benchmarks/bench.py on the 20000-key set from the table above, taking only the first hit of a 2-character predict costs roughly a tenth of the full list (332 ns against 2892 ns per prefix), while draining the whole generator costs about 10% more than the list version; the search variants, whose results are capped at one per key byte, are cheap either way.

Enumerating a trie (or a subtree) uses begin / next:

>>> result, from_id, pos = t.traverse('app')   # locate the subtree
>>> value, node_id, length = t.begin(from_id, pos)
>>> while value != pycedar.base_trie.NO_PATH:
...     print(t.suffix(node_id, length), value)
...     value, node_id, length = t.next(node_id, length, from_id)
apple 1
applet 2

API reference

pycedar.base_trie

Base class for every trie. Not meant to be instantiated directly.

pycedar.str_trie / pycedar.bytes_trie / pycedar.unicode_trie

Specialisations of base_trie for str keys, bytes keys, and unicode keys respectively. On Python 3, unicode is str, so unicode_trie behaves identically to str_trie and is kept only for backward compatibility.

In addition to the base_trie members:

max_size caps the number of returned results; -1 means "no limit". from_id scopes the query to a subtree.

pycedar.node

A cursor into a trie. Obtained from base_trie.root, dict.root, dict.get_node(), dict.nodes() or node.find_nodes().

pycedar.dict

A dict-like façade over a trie. pycedar.dict(key_type) accepts str (the default) or bytes.

pycedar.dict (and the trie classes) support pickle, copy.copy() and copy.deepcopy() through the same serialized image.

pycedar.__version__ exposes the installed package version.

Limitations

Concurrent access is safe only across distinct tries

The module is declared free-threading compatible (PEP 703) and works on free-threaded CPython 3.14 and 3.15. That covers using separate trie objects from several threads at once, which the test suite exercises.

It does not make an individual trie thread-safe: concurrent operations on the same trie race, exactly as they would under the GIL, and can corrupt it. Guard a shared trie with a lock as you would any other mutable object.

Values are C int sized, and two of them are reserved

Values are stored as C int, so they must fit in -2**31 .. 2**31-1; anything larger raises OverflowError.

-1 and -2 are cedar's sentinels — base_trie.NO_VALUE and base_trie.NO_PATH — and every writer rejects them:

>>> limited = pycedar.dict()
>>> limited['key'] = -1
Traceback (most recent call last):
    ...
ValueError: -1 is reserved and cannot be stored: ...

update() checks the resulting value rather than the delta, because a perfectly ordinary delta can still land on a sentinel. When it does, the delta is rolled back and the stored value is left untouched:

>>> limited['counter'] = 1
>>> limited.update('counter', -3)     # 1 + (-3) == -2
Traceback (most recent call last):
    ...
ValueError: -2 is reserved and cannot be stored: ...
>>> limited['counter']
1

Every other value round trips, negative ones included.

Before 0.3.0 these two were accepted and silently corrupted the trie: -1 made a key invisible to in, get() and d[key] while leaving it visible to iteration, and -2 ended every traversal early, hiding each key that came after it. If you are upgrading and were storing either, the values were not being read back correctly in the first place.

Keys cannot contain a NUL byte

cedar keeps short key suffixes in a NUL terminated array, so a key carrying a NUL of its own breaks that invariant. Such keys are rejected:

>>> keyed = pycedar.dict()
>>> keyed['a\x00b'] = 1
Traceback (most recent call last):
    ...
ValueError: key contains a NUL byte, ...

Before 0.4.0 the write was accepted, and it did more than read back wrong: inserting such a key and then inserting one that shared its prefix corrupted memory and crashed the interpreter.

The serialization format is platform-dependent and unauthenticated

The native cedar .dat format depends on the pointer size and byte order of the machine that wrote it, and it carries no integrity checks. Only load files produced by pycedar on a compatible platform and obtained from a trusted source.

The same applies to dumps() / loads(), which use the identical layout, and to pickle — pickling serializes this image, so pickles are only portable between machines of the same platform and integrity is not verified.

I/O failures are reported through return codes

save() / load() / open() return 0 on success and -1 on failure instead of raising; check the return value.

Running out of memory is the exception to that rule: save() and load() raise MemoryError rather than returning -1. A failed load() leaves the trie empty, because the previous contents are released before the new ones are allocated. The instance stays valid and can be reused. See pycedar/core/cedar/README.md.

Development

$ python -m pip install --upgrade pip
$ python -m pip install . pytest
$ pytest

To rebuild in place while iterating on pycedar.pyx:

$ python -m pip install "Cython>=3.1,<4" setuptools
$ python setup.py build_ext --inplace

./clean.sh removes build artifacts.

Benchmarks

benchmarks/bench.py times the operations pycedar is used for. Run it against two builds to check whether a change actually paid off:

$ python benchmarks/bench.py --label before
$ python benchmarks/bench.py --label after

It uses rich for the table when that is installed, and plain text otherwise. --help lists the knobs.

Releasing

The version lives in a single place, pycedar/VERSION.

1. Update pycedar/VERSION and CHANGELOG.md.
2. Commit and push to master.
3. Push a matching tag, e.g. git tag v0.2.0 && git push origin v0.2.0.

The release workflow verifies that the tag matches pycedar/VERSION, builds the sdist and the Linux/macOS wheels with cibuildwheel, and publishes them to PyPI via Trusted Publishing.

License

pycedar is distributed under the same terms as cedar itself: GPLv2, LGPLv2.1 and BSD-2-Clause. See the license files bundled under pycedar/core/cedar/.

Credits

cedar is written by Naoki Yoshinaga. The copy vendored under pycedar/core/cedar/ carries local modifications. pycedar/core/cedar/README.md records what was changed and why, why the 2022 upstream tarball was deliberately not re-vendored, and what has to be re-applied if anyone syncs with a newer release.

See CHANGELOG.md for the list of contributors to each release.