Armin Friedl
d8d8436325
All checks were successful
continuous-integration/drone/push Build is passing
|
||
---|---|---|
bytetrie | ||
examples | ||
tests | ||
.drone.yml | ||
.gitignore | ||
LICENSE | ||
README.md | ||
setup.py | ||
TODO.md | ||
trie.png |
Bytetrie
A fast, dependency-free, self-compressing trie with radix 256 in pure python.
Bytetrie allows fast prefix search in a large corpus of keys. Each key can be associated with arbitrary data. It features fast lookup times at the cost of expensive insertion. A Bytetrie is best used if it can be pre-filled with data. However, due to its in-band compression it can be also used for on-the-fly updates.
Keys
Keys are byte strings. Therefore, each node in the trie can have up to 256 children (the radix). Keys do work well with utf-8 and other encodings as long as the encoding is consistent and deterministic. That is, grapheme clusters are always encoded to the same byte sequence -- even if the standard allows for ambiguity. Usually that's a non-issue as long as the same encoder is used for insertion and lookup.
Since prefix search in unicode strings is one of the most common use-cases of bytetrie, a unicode layer on top of bytetrie is planned.
Data
Bytetrie can associate arbitrary python objects with keys. Data (or rather a reference thereof) is kept in-tree. No further processing is done.
In addition, bytrie allows multi-valued tries. Every key is then associated with a sequence of arbitrary objects.
Performance
Despite being in pure python bytetrie is fast. Sifting through the full
geonames "allCountries" dataset for
places starting with Vienna
takes a mere 512µs. That's not even a
millisecond for searching through 12,041,359 places. For comparison, a warmed-up
ripgrep search through the same dataset takes three orders of magnitude (400ms)
longer on the same machine.
On the downside, building the trie takes about 20 minutes and considerable memory. Also, the performance is mostly trumped by the time it takes to collect terminal nodes. The higher up the trie the search ends (and hence the more results the prefix search yields) the longer it takes. There are several low-hanging fruits left and further performance improvements are in the pipeline.
Dependencies
None. That's the point.
Getting started
Install bytetrie via pip.
pip install -U bytetrie
The public interface is ByteTrie
with the two methods insert
and find
.
Find returns a list of Terminals
from which the key
and the value
of the
node can be retrieved.
from bytetrie import ByteTrie
t = ByteTrie(multi_value=True)
t.insert(b"Hallo", "Dutch")
t.insert(b"Hello", "English")
t.insert(b"Hug", "Gaelic")
t.insert(b"Hallo", "German")
t.insert("Hē".encode("utf-8"), "Hindi")
t.insert("Halló".encode("utf-8"), "Icelandic")
t.insert(b"Hej", "Polish")
t.insert(b"Hei", "Romanian")
t.insert(b"Hujambo", "Swahili")
t.insert(b"Hej", "Swedish")
t.insert(b"Helo", "Welsh")
print("Where to say 'Hi' with 'He'?")
print(f"{[(n.key(), n.value()) for n in t.find(b'He')]}")
print("Where to say 'Hi' with 'Ha'?")
print(f"{[(n.key().decode('utf-8'), n.value()) for n in t.find(b'Ha')]}")
print("Where to say 'Hi' with 'Hē'?")
print(f"Say 'Hi' with utf-8: {[(n.key().decode('utf-8'), n.value()) for n in t.find('Hē'.encode('utf-8'))]}")
Contribute
If you want to contribute to bytetrie
feel free to send patches to
dev[at]friedl[dot]net. Alternatviely, you can issue a pull request on GitHub
which will be cherry picked into my tree. If you plan significant long-term
contributions drop me a mail for access to the incubator repository.
Github Users
If you are visiting this repository on GitHub, you are on a mirror of https://git.friedl.net/incubator/bytetrie. This mirror is regularily updated with my other GitHub mirrors.
Like with my other incubator projects, once I consider bytetrie
reasonable
stable the main tree will move to GitHub.