A Bloom filter provides fast determination if a key is not in a set. It consists of a collection of N bits. Adding a key to the filter requires generating one more more bits (hashing) to represent the key, and those bits are then OR’ed together with existing bit values in the collection. Thus, the filter represents the aggregate bit mask of all of the keys added to it.
Below is a depiction of a Bloom filter as three keys are added to it. Note that once a bit in the filter is set (green), it stays that way forever.
To check if a key
K is in the filter, one just hashes the key to get the bit pattern and then checks of all of the bits are present in the filter. Since the combining operation is an OR, if a bit is not found in the collection, then the key was never added. However, the addition of two or more different keys could set the bits of key
K, thus making it appear that
K was added to the collection some time in the past. That is why the filter only guarantees being able to filter out unseen keys; false positives may happen due to bit patterns overlapping with other keys.
Here, we ask if a key that was not added to the filter is a member. Since the two 1 bits of the key are also set in the filter bit array, the filter can only say perhaps.
Even though the filter can not conclusively say a key is part of the collection, the Bloom filter is still very useful as a way to reduce the amount of work a system must perform when given a random set of keys. Take for instance a search engine that can preprocess the searchable material and generate bit patterns of all the material. The work performed by the search engine is now reduced by some amount due to the ability to definitively state that some queries will not return a value.
For a past project, my team used a Bloom filter to be able to quickly determine if there was any value in downloading an encrypted and compressed log file for search processing. Since downloading, decrypting, and uncompressing a log file are all expensive operations, anything that we could do to reduce their occurrence would be a big win.
The keys of the Bloom filter were SHA hashes of user IDs, and the hashing of each key resulted in 2 bits of information (out of a table size of N bits). As the cloud service generated log entries it also added the user ID hash to a bloom filter. When the service closed, compressed, encrypted, and saved the log on Azure, it also saved the bit array of the Bloom filter as metadata of the log blob.
Python Bloom Filter
Below is the
BloomFilter class I wrote for our log query operation. It takes in a list of byte values which will be used as the source for a bitarray object (little-endian). It also accepts a count of the number of hashes (bits) a key has.
from bitarray import bitarray import Logger class BloomFilter(object): '''Implementation of a read-only Bloom filter. Takes binary values from an Azure table entity. ''' def __init__(self, rawBytes, hashCount): '''Initialize our bitarray that returns true/false for indices from hash values rawBytes -- binary values from Azure entity that represents the Bloom Filter recording hashCount -- value from Azure entity that defines how many bits a key has ''' gLog.begin(len(rawBytes), hashCount) # Initialize bit array from raw bytes self.bits = bitarray(endian = 'little') self.bits.frombytes(rawBytes) self.size = self.bits.length() self.hashCount = hashCount gLog.end() def contains(self, hashInit): '''Determine if a given user ID hash pair is found in the filter. Implementation follows that of the C# code in the cloud service. hashInit -- the starting hash pair Returns True if the object represented by the hashInit is most likely in the filter (there can be false positives) ''' gLog.begin(hashInit) # Convert to indices into the table h1, h2 = [h % self.size for h in hashInit] # Check N times with different hashes to see if the Skype ID is found. There are false positives, but no # false negatives. for index in range(self.hashCount): gLog.debug('h1:', h1, 'h2:', h2) if self.bits[h1] == False: gLog.end(False) return False # Generate a new, unique hash pair h1 = (h1 + h2) % self.size h2 = (h2 + index) % self.size gLog.end(True) return True
contains method checks to see if the filter might contain a given value. The sole parameter is a 2-tuple of integer values which represent a hash of the user ID. The routine then iterates the required number of times to generate all of the bits for the key. However, if a bit does not exist in the filter’s bit array, the method stops and returns false.
Additional bits come from adding previous bit positions plus an incremental offset. If the original 2-tuple values are randomly distributed for user IDs, then the resulting bits should be as well.
The above class does not offer a way to update the filter’s bit array, but adding such a feature would be trivial to implement using the code in the