Peloton Cycle's insanely fast integer dictionary

pip install intdict==0.0.1


Bloomin fast Bloomfilters from Peloton

peloton_bloomfilter.SharedMemoryBloomfilter is the easiest to use, fastest bloomfilter implementation for cPython.


Bloomfilters are probabilistic data structures supporting basic object membership testing. Objects are added to bloomfilters with the add method and tested for membership with the in operator. A bloomfilter reporting False to an in query is guaranteed to not have the object; a bloomfilter reporting True likely has the tested object subject to a tunable false positive rate.

peloton_bloomfilters implements three bloom-filter classes; BloomFilter is a plain old bloomfilterfor single threads or gevent apps; ThreadSafeBloomFilter releases the GIL and uses __atomic_or_fetch to prevent lost bits during writes and SharedMemoryBloomfilter supports the creation of bloomfilters that are shared between processes in real time using files and mmap

To create a bloomfilter object you merely import the module and call it with two or three parameters: the file name to hold the shared memory mmap object, the capacity of the bloomfilter and its false positive rate.

>>> from peloton_bloomfilter import *
>>> bf = BloomFilter(1000, 0.001)
>>> tsbf = ThreadSafeBloomFilter(1000, 0.001)
>>> smbf = SharedMemoryBloomfilter("/tmp/filter", 1000, 0.001)

Adding and testing membership against a bloomfilter works exactly like a set, except a bloomfilter cannot be enumerated.

>>> smbf.add(1)
>>> 1 in smbf
>>> 2 in smbf

Note that add returns False. SharedMemoryBloomfilter has a limited capacity; before each add the remaining capacity is tested an if its insufficient the bloom-filer will be cleared prior to performing the add and True is returned

len() reports the number of items stored in the bloomfilter since created or last cleared.

>>> len(smbf)

bloomfilters may be explicitly cleared.

>>> smbf.clear()
>>> 1 in smbf
>>> len(smbf)


peloton_bloomfilter.SharedMemoryBloomfilter is the fastest cPython bloomfilter implementation known to its authors. How fast? Here we benchmark peloton_bloomfilter against pybloomfiltermmap-0.3.14 in their ability to add and test membership for member and non-member objects against a 1,000,000 capacity bloomfilter for varying false error rates.

Both libraries were compiled with gcc-4.8.5, CFLAGS="-mtune=native -march=native" on Unbuntu 14.04 running on a Dell XPS 13 with 16Gb Ram and a dual core Intel(R) Core(TM) i7-6560U CPU @ 2.20GHz and cpupower frequency-set -g performance. Times are in nanoseconds.

1,000,000 adds

1/p     peloton   pybloommap
10        139       276
100       184       394 
10000     431       459
1000000   693       757

1,000,000 membership tests, existing items

1/p     peloton   pybloommap
10        87        224
100       114       352
10000     307       424
1000000   523       671

1,000,000 membership tests, absent items

1/p     peloton   pybloommap
10        81        209
100       82        222
10000     102       159
1000000   119       179