HAT-Trie for Python
C Python C++
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.



HAT-Trie structure for Python (2.x and 3.x).

This package is a Python wrapper for hat-trie C library.



pip install hat-trie


Create a new trie:

>>> from hat_trie import Trie
>>> trie = Trie()

trie variable is a dict-like object that support unicode keys and can have any Python object as a value. For keys that share prefixes it usually uses less memory than Python dict.

There is also hat_trie.IntTrie which only supports positive integers as values. It can be more efficient when you don't need arbitrary objects as values. For example, if you need to store float values then storing them in an array (either numpy or stdlib's array.array) and using IntTrie values as indices could be more memory efficient than storing Python float objects directly in hat_trie.Trie.

Another way to store float values is to use hat_trie.FloatTrie(). In this case precision is limited to float32.

Currently implemented methods are:

  • __getitem__()
  • __setitem__()
  • __contains__()
  • __len__()
  • get()
  • setdefault()
  • keys()
  • iterkeys()

Other methods are not implemented - contributions are welcome!


Performance is measured for hat_trie.Trie against Python's dict with 100k unique unicode words (English and Russian) as keys and '1' numbers as values.

Benchmark results for Python 3.3 (intel i5 1.8GHz, "1.000M ops/sec" == "1 000 000 operations per second"):

dict __getitem__ (hits)      6.874M ops/sec
trie __getitem__ (hits)      3.754M ops/sec
dict __contains__ (hits)     7.035M ops/sec
trie __contains__ (hits)     3.772M ops/sec
dict __contains__ (misses)   5.356M ops/sec
trie __contains__ (misses)   3.364M ops/sec
dict __len__                 785958.286 ops/sec
trie __len__                 574164.704 ops/sec
dict __setitem__ (updates)   6.830M ops/sec
trie __setitem__ (updates)   3.472M ops/sec
dict __setitem__ (inserts)   6.774M ops/sec
trie __setitem__ (inserts)   2.460M ops/sec
dict setdefault (updates)    3.522M ops/sec
trie setdefault (updates)    2.680M ops/sec
dict setdefault (inserts)    4.062M ops/sec
trie setdefault (inserts)    1.866M ops/sec
dict keys()                  189.564 ops/sec
trie keys()                  16.067 ops/sec

HAT-Trie is about 1.5x faster that datrie on all supported operations; it also supports fast inserts unlike datrie. On the other hand, datrie has more features (e.g. better iteration support and richer API); datrie is also more memory efficient.

If you need a memory efficient data structure and don't need inserts then marisa-trie or DAWG should work better.


Development happens at github:

Feel free to submit ideas, bugs, pull requests or regular patches.

Please don't commit changes to generated C files; I will rebuild them myself.

Running tests and benchmarks

Make sure tox is installed and run

$ ./update_c.sh
$ tox

from the source checkout. You will need Cython to do that.

Tests should pass under python 2.7 and 3.3+.

$ tox -c bench.ini

runs benchmarks.

Authors & Contributors

This module wraps hat-trie C library by Daniel Jones & contributors.


Licensed under MIT License.