A fast, simple and lightweight
Bloom filter library for
Python, implemented in Rust. It's designed to be as pythonic as possible,
mimicking the built-in set
type where it can, and works with any
hashable object. While it's a new library (this project was started in
2023), it's currently the fastest option for Python by
a long shot (see the section
Benchmarks). Releases are published on
PyPI.
This library defines only one class, which can be used as follows:
>>> from rbloom import Bloom
>>> bf = Bloom(200, 0.01) # 200 items max, false positive rate of 1%
>>> bf.add("hello")
>>> "hello" in bf
True
>>> "world" in bf
False
>>> bf.update(["hello", "world"]) # "hello" and "world" now in bf
>>> other_bf = Bloom(200, 0.01)
### add some items to other_bf
>>> third_bf = bf | other_bf # third_bf now contains all items in
# bf and other_bf
>>> third_bf = bf.copy()
... third_bf.update(other_bf) # same as above
>>> bf.issubset(third_bf) # bf <= third_bf also works
True
For the full API, see the section Documentation.
On almost all platforms, simply run:
pip install rbloom
If you're on an uncommon platform, this may cause pip to build the library
from source, which requires the Rust
toolchain. You can also build
rbloom
by cloning this repository and running
maturin:
maturin build --release
This will create a wheel in the target/wheels/
directory, which can
subsequently also be passed to pip.
Why should you use this library instead of one of the other Bloom filter libraries on PyPI?
- Simple: Almost all important methods work exactly like their
counterparts in the built-in
set
type. - Fast:
rbloom
is implemented in Rust, which makes it blazingly fast. See section Benchmarks for more information. - Lightweight:
rbloom
has no dependencies of its own. - Maintainable: This library is very concise, and it's written
in idiomatic Rust. Even if I were to stop maintaining
rbloom
(which I don't intend to), it would be trivially easy for you to fork it and keep it working for you.
I started rbloom
because I was looking for a simple Bloom filter
dependency for a project, and the only sufficiently fast option
(pybloomfiltermmap3
) was segfaulting on recent Python versions. rbloom
ended up being twice as fast and has grown to encompass a more complete
API (e.g. with set comparisons like issubset
). Do note that it doesn't
use mmapped files, however. This shouldn't be an issue in most cases, as
the random access heavy nature of a Bloom filter negates the benefits of
mmap after very few operations, but it is something to keep in mind for
edge cases.
The following simple benchmark was implemented in the respective API of each library (see the comparison benchmarks):
bf = Bloom(10_000_000, 0.01)
for i in range(10_000_000):
bf.add(i + 0.5) # floats because ints are hashed as themselves
for i in range(10_000_000):
assert i + 0.5 in bf
This resulted in the following average runtimes on an M1 Pro (confirmed to be proportional to runtimes on an Intel machine):
Library | Time | Notes |
---|---|---|
rBloom | 2.52s | works out-of-the-box |
pybloomfiltermmap3 | 4.78s | unreliable [1] |
pybloom3 | 46.76s | works out-of-the-box |
Flor | 76.94s | doesn't work on arbitrary objects [2] |
bloom-filter2 | 165.54s | doesn't work on arbitrary objects [2] |
[1] The official package failed to install on Python 3.11 and kept segfaulting on 3.10 (Linux, January 2023). It seems to be fine for now (October 2024). [2] I was forced to convert to a byte representation, which is bad default behavior as it presents the problems mentioned below in the section "Cryptographic security".
Also note that rbloom
is compiled against a stable ABI for
portability, and that you can get a small but measurable speedup by
removing the "abi3-py37"
flag from Cargo.toml
and building
it yourself.
This library defines only one class, the signature of which should be
thought of as follows. Note that only the first few methods differ from
the built-in set
type:
class Bloom:
# expected_items: max number of items to be added to the filter
# false_positive_rate: max false positive rate of the filter
# hash_func: optional argument, see section "Cryptographic security"
def __init__(self, expected_items: int, false_positive_rate: float,
hash_func=__builtins__.hash)
@property
def size_in_bits(self) -> int # number of buckets in the filter
@property
def hash_func(self) -> Callable[[Any], int] # retrieve the hash_func
# given to __init__
@property
def approx_items(self) -> float # estimated number of items in
# the filter
# see section "Persistence" for more information on these four methods
@classmethod
def load(cls, filepath: str, hash_func) -> Bloom
def save(self, filepath: str)
@classmethod
def load_bytes(cls, data: bytes, hash_func) -> Bloom
def save_bytes(self) -> bytes
#####################################################################
# ALL SUBSEQUENT METHODS ARE #
# EQUIVALENT TO THE CORRESPONDING METHODS #
# OF THE BUILT-IN SET TYPE #
#####################################################################
def add(self, obj) # add obj to self
def __contains__(self, obj) -> bool # check if obj in self
def __bool__(self) -> bool # False if empty
def __repr__(self) -> str # basic info
def __or__(self, other: Bloom) -> Bloom # self | other
def __ior__(self, other: Bloom) # self |= other
def __and__(self, other: Bloom) -> Bloom # self & other
def __iand__(self, other: Bloom) # self &= other
# these extend the functionality of __or__, __ior__, __and__, __iand__
def union(self, *others: Union[Iterable, Bloom]) -> Bloom # __or__
def update(self, *others: Union[Iterable, Bloom]) # __ior__
def intersection(self, *others: Union[Iterable, Bloom]) -> Bloom # __and__
def intersection_update(self, *others: Union[Iterable, Bloom]) # __iand__
# these implement <, >, <=, >=, ==, !=
def __lt__, __gt__, __le__, __ge__, __eq__, __ne__(self,
other: Bloom) -> bool
def issubset(self, other: Bloom) -> bool # self <= other
def issuperset(self, other: Bloom) -> bool # self >= other
def clear(self) # remove all items
def copy(self) -> Bloom # duplicate self
To prevent death and destruction, the bitwise set operations only work on
filters where all parameters are equal (including the hash functions being
the exact same object). Because this is a Bloom filter, the __contains__
and approx_items
methods are probabilistic, as are all the methods that
compare two filters (such as __le__
and issubset
).
Python's built-in hash function is designed to be fast, not maximally collision-resistant, so if your program depends on the false positive rate being perfectly correct, you may want to supply your own hash function. This is especially the case when working with very large filters (more than a few tens of millions of items) or when false positives are very costly and could be exploited by an adversary. Just make sure that your hash function returns an integer between -2^127 and 2^127 - 1. Feel free to use the following example in your own code:
from rbloom import Bloom
from hashlib import sha256
from pickle import dumps
def hash_func(obj):
h = sha256(dumps(obj)).digest()
# use sys.byteorder instead of "big" for a small speedup when
# reproducibility across machines isn't a concern
return int.from_bytes(h[:16], "big", signed=True)
bf = Bloom(100_000_000, 0.01, hash_func)
When you throw away Python's built-in hash function and start hashing serialized representations of objects, however, you open up a breach into the scary realm of the unpythonic:
- Numbers like
1
,1.0
,1 + 0j
andTrue
will suddenly no longer be equal. - Instances of classes with custom hashing logic (e.g. to stop caches inside instances from affecting their hashes) will suddenly display undefined behavior.
- Objects that can't be serialized simply won't be hashable at all.
Making you supply your own hash function in this case is a deliberate design decision intended to show you what you're doing and prevent you from shooting yourself in the foot.
Also note that using a custom hash will incur a performance penalty over using the built-in hash.
The save
and load
methods, along with their byte-oriented counterparts
save_bytes
and load_bytes
, allow you to save and load filters to and
from disk/Python bytes
objects. However, as the built-in hash function's
salt changes between invocations of Python, they only work on filters with
custom hash functions. Note that it is your responsibility to ensure that
the hash function you supply to the loading functions is the same as the
one originally used by the filter you're loading!
bf = Bloom(10_000, 0.01, some_hash_func)
bf.add("hello")
bf.add("world")
# saving to a file
bf.save("bf.bloom")
# loading from a file
loaded_bf = Bloom.load("bf.bloom", some_hash_func)
assert loaded_bf == bf
# saving to bytes
bf_bytes = bf.save_bytes()
# loading from bytes
loaded_bf_from_bytes = Bloom.load_bytes(bf_bytes, some_hash_func)
assert loaded_bf_from_bytes == bf
The size of the file is bf.size_in_bits / 8 + 8
bytes.
Statement of attribution: Bloom filters were originally proposed in (Bloom, 1970). Furthermore, this implementation makes use of a constant recommended by (L'Ecuyer, 1999) for redistributing the entropy of a single hash over multiple integers using a linear congruential generator.