Professor Oak on Bloom filters.

In this post, we will briefly go over the probabilistic data structure referred to as a Bloom filter. We’ll be using Pokémon to help us in understanding the general concept of how to make use of such a data structure. The Jupyter notebook corresponding to this post may be found on my GitHub page here:

If you prefer, there is also an accompanying video that goes along with this post here:

TL;DR

Bloom filters: “Lightweight” version of a hash table. Both hash tables and Bloom filters support efficient insertions and lookups. Bloom filters are more space efficient than hash tables, but this comes at the cost of having “false positives” for entry lookup. That is, Bloom filters can say with certainty that an element has not been inserted (no possibility of false negatives), but may indicate an element has been inserted when it has in fact not been (false positive).

When should I use a Bloom filter?

“I want a data structure that allows for fast lookups and insertions. I care about how much space the data structure uses. I don’t care if the data structure sometimes indicates an item is present when in fact it is not.”

Example:

I run a website and want to keep track of IP addresses that are blocked and I don’t particularly care if a blocked IP address is occasionally able to access my website, but I do care if someone not on the blocked list is unable to access the site.

More examples found on Wikipedia.

Bloom Filter: Toy Example in Python

In order to illustrate how a Bloom filter works let’s consider a toy example. We start with a bit vector; a vector whose elements are $0$ or $1$. To start, we initialize the bit vector to all zeros. For the purposes of this toy example, we will restrict our attention to a bit vector of size $20$.

bit_vector = [0] * 20
print(bit_vector)
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

The next ingredient we require is the use of a couple hash functions, that is, a function that maps data of arbitrary size to data of a fixed size. The types of hash functions used in Bloom filters are generally not of the “cryptographic variety”, for example, one usually wouldn’t use something like MD5. Non-cryptographic hash functions like Murmur and FNV are mostly used, primarily for their speed over most cryptographic hash functions.

There is a nice module in Python called pyhash that consists solely of non-cryptographic hashes.

import pyhash

Let’s combine the bit vector and non-cryptographic hash functions to put together a toy example of a Bloom filter. In our example, let’s say we’re using our Bloom filter as a Pokédex; a device to keep track of the Pokémon we have caught. Each time we catch a Pokémon, we update our Pokédex by running the name of the Pokémon through two hash functions. The output of the hashes indicates which bits to flip in our bit vector.

Hashing the strings “Pikachu” and “Charmander” using the FNV hash algorithm mod 20 (since 20 is the size of our bit vector in this example) results in 13 and 5. Likewise, hashing the same strings using the Murmur hashing algorithm mod 20 results in 10 and 9, respectively.

Feeding the Pokemon strings into the hash functions.

We use the outputs of the above hash algorithms to flip the bits located at the respective indices. For instance:

Resulting bit vector once we've hashed the Pokemon strings.

The following Python code achieves what we described above:

# Define FNV and Murmur hash functions from Pyhash.
fnv_hasher = pyhash.fnv1_32()
murmur_hasher = pyhash.murmur3_32()

# Calculate output of FNV and Murmur hash for Pikachu and Charmander.
fnv_hash_pikachu = fnv_hasher("Pikachu") % 20
murmur_hash_pikachu = murmur_hasher("Pikachu") % 20

fnv_hash_charmander = fnv_hasher("Charmander") % 20
murmur_hash_charmander = murmur_hasher("Charmander") % 20

# Print the output of FNV and Murmur hashes.
print("FNV hash output for Pikachu: " + str(fnv_hash_pikachu))
print("Murmur hash output for Pikachu: " + str(murmur_hash_pikachu))

print("FNV hash output for Charmander: " + str(fnv_hash_charmander))
print("Murmur hash output for Charmander: " + str(murmur_hash_charmander))

# Flip the bits of bit_vector in the corresponding locations from above hashes.
bit_vector[fnv_hash_pikachu] = 1
bit_vector[murmur_hash_pikachu] = 1

bit_vector[fnv_hash_charmander] = 1
bit_vector[murmur_hash_charmander] = 1

print(bit_vector)
FNV hash output for Pikachu: 13
Murmur hash output for Pikachu: 10
FNV hash output for Charmander: 5
Murmur hash output for Charmander: 9
[0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0]

Walking in the tall grass, a wild Bulbasaur appears!

Bulbasaur.

Let’s consult our Bloom filter Pokédex to see if we’ve already captured Bulbasaur or not

# Calculate output of FNV and Murmur hash for Bulbasaur.
fnv_hash_bulbasaur = fnv_hasher("Bulbasaur") % 20
murmur_hash_bulbasaur = murmur_hasher("Bulbasaur") % 20

# Print the FNV and Murmur hashes of Bulbasaur.
print("FNV hash output for Bulbasaur: " + str(fnv_hash_bulbasaur))
print("Murmur hash output for Bulbasaur: " + str(murmur_hash_bulbasaur))
FNV hash output for Bulbasaur: 11
Murmur hash output for Bulbasaur: 8

Adding Bulbasaur to our Bloom filter.

Looks like the outputs of hashing “Bulbasaur” (mod 20) result in an output of $11$ and $8$ for the FNV and Murmur hashes respectively. Consulting our Bloom filter bit vector, we check whether the bits are flipped “on” or “off” at these indices:

print(bit_vector[fnv_hash_bulbasaur])
print(bit_vector[murmur_hash_bulbasaur])
0
0

Since the entries for both of the hash functions for Bulbasaur result in $0$, Bulbasaur is not in our Pokédex. Let’s capture Bulbasaur and appropriately modify out Bloom filter Pokédex to reflect this.

# Flip the bits in the Bloom filter to indicate that we now have captured Bulbasaur.
bit_vector[fnv_hash_bulbasaur] = 1
bit_vector[murmur_hash_bulbasaur] = 1
print(bit_vector)
[0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0]

BULBASAUR was caught!

How big does our Bloom filter need to be?

In our toy example, the size of the Bloom filter consists of 20 entries. For our Pokedexing needs, the length of the filter is most likely too small to be useful. To see why, let’s assume we’ve caught 150 of the 151 of the Pokémon.

# 150 of the 151 Pokemon (excluding Pidgey)
caught_pokemon = ["Bulbasaur", "Ivysaur", "Venusaur", "Charmander", "Charmeleon", "Charizard", 
                  "Squirtle", "Wartortle", "Blastoise", "Caterpie", "Metapod", "Butterfree", "Weedle", 
                  "Kakuna", "Beedrill", "Pidgeotto", "Pidgeot", "Rattata", "Raticate", "Spearow", "Fearow", 
                  "Ekans", "Arbok", "Pikachu", "Raichu", "Sandshrew", "Sandslash", "Nidoran (female)", "Nidorina",
                  "Nidoqueen", "Nidoran (male)", "Nidorino", "Nidoking", "Clefairy", "Clefable", "Vulpix", 
                  "Ninetales", "Jigglypuff", "Wigglytuff", "Zubat", "Golbat", "Oddish", "Gloom", "Vileplume", 
                  "Paras", "Parasect", "Venonat", "Venomoth", "Diglet", "Dugtrio", "Meowth", "Persian", "Psyduck",
                  "Golduck", "Mankey", "Primeape", "Growlithe", "Arcanine", "Poliwag", "Poliwhirl", "Poliwrath",
                  "Abra", "Kadabra", "Alakazam", "Machop", "Machoke", "Machamp", "Bellsprout", "Weepinbell", 
                  "Victreebel", "Tentacool", "Tentacruel", "Geodude", "Graveler", "Golem", "Ponyta", "Rapidash",
                  "Slowpoke", "Slowbro", "Magnemite", "Magneton", "Farfetc'd", "Doduo", "Dodrio", "Seel",
                  "Dewgong", "Grimer", "Muk", "Shellder", "Cloyster", "Gastly", "Haunter", "Gengar", "Onyx",
                  "Drowzee", "Hypno", "Krabby", "Kingler", "Voltorb", "Electrode", "Exeggcute", "Exeggutor", 
                  "Cubone", "Marowak", "Hitmonlee", "Hitmonchan", "Lickitung", "Koffing", "Weezing", "Rhyhorn", 
                  "Rhydon", "Chansey", "Tangela", "Kangaskhan", "Horsea", "Seadra", "Goldeen", "Seaking", 
                  "Staryu", "Starmie", "Mr. Mime", "Scyther", "Jynx", "Electabuzz", "Magmar", "Pinsir", "Tauros",
                  "Magikarp", "Gyrados", "Lapras", "Ditto", "Eevee", "Vaporeon", "Jolteon", "Flareon", "Porygon", 
                  "Omanyte", "Omastar", "Kabuto", "Kabutops", "Aerodactyl", "Snorlax", "Articuno", "Zapdos",
                  "Moltres", "Dratini", "Dragonair", "Dragonite", "Mewtwo", "Mew" ]
pokedex_bloom_filter = [0] * 20

# Update the Bloom filter positions of the bit vector.
for pokemon in caught_pokemon:
    fnv_hash = fnv_hasher(pokemon) % 20
    murmur_hash = murmur_hasher(pokemon) % 20
    
    pokedex_bloom_filter[fnv_hash] = 1
    pokedex_bloom_filter[murmur_hash] = 1
    
# The Pokedex Bloom filter.    
print(pokedex_bloom_filter)
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]

Unbeknown to our Pokémon trainer, the last Pokemon they have yet to catch is the elusive Pidgey. As luck should have it, the trainer walks into some tall grass and a wild Pidgey appears!

Pidgey

Before deciding to catch Pidgey, the trainor pulls out their Pokédex to see if they’re already caught it.

fnv_hash_pidgey = fnv_hasher("Pidgey") % 20
murmur_hash_pidgey = murmur_hasher("Pidgey") % 20

print(pokedex_bloom_filter[fnv_hash_pidgey])
print(pokedex_bloom_filter[murmur_hash_pidgey])
1
1

Hmm, okay, well the entries in the Bloom filter indicate that Pidget may be in the Pokédex already. As mentioned before, Bloom filters will only give false positives, never false negatives. That is to say that our Pokédex will never say that Pidgey is not in our Pokédex if it actually is, but it may say that Pidgey is in our Pokédex, even if it actually isn’t.

This is precisely the issue we’re having at the moment. We, the audience, know that Pidgey is not in the Pokédex. Since the size of the Bloom filter is only of size $20$ and the total number of Pokémon is $151$, there will inevitably be some hash collisions leading to a situation like the one we’re encountering now.

Unfortunately, our Pokémon trainer decides that Pokéballs are expensive and misses their opportunity to complete their Pokédex. If Professor Oak gave our trainer a Bloom filter Pokédex with a larger number of entries, he could have reduced the probability of this unfortunate event.

As it happens, the rate of false positives is approximately

\((1 - e^{-kn/m})^k\),

where $k$ is the number of hash functions used, $n$ is the number of inserted elements, and $m$ is the total number of bits, or equivalently, the length of the bit vector used for the Bloom filter.

More information on the above equation can be found on the Wikipedia page where there is a more in-depth discussion on the probability of obtaining false positives.

So in order to reduce the probability of collisions, we need to tweak the parameters $k$ and $m$; the number of hash functions we use and the size of the bit vector, respectively. Assuming that the hash function selects an index in the bit vector completely at random, the probability that an element in the Bloom filter is not set to $1$ can be calculated by

\[1 - \frac{1}{m}.\]

By taking this to the $k^{th}$ power, we have the probability that the elements in the Bloom filter are not set to $1$ which can be calculated by

\[\left( 1 - \frac{1}{m} \right)^k.\]

If we have inserted $n$ Pokémon into our Bloom filter, the probability that a certain bit in the vector is still $0$ is given by

\[\left( 1 - \frac{1}{m} \right)^{kn}.\]

For instance, in our toy example, we had that the size of the bit vector for the Bloom filter was $m = 20$. We made use of $k = 2$ hash functions, FNV and Murmur. When we caught Pikachu and Charmander, this sets the value corresponding to the total number of Pokémon in our Pokédex to $n = 2$. Therefore the probability that a certain element in the bit vector is still $0$ before capturing Bulbasaur is given as

\[\left( 1 - \frac{1}{20} \right)^{4} \approx 0.81450.\]

Not terrible, but as we catch more Pokémon, this number drops quite a bit. Since there are 151 Pokémon in total, it doesn’t take long for the probability of any element in the bit vector to still be $0$ to be quite low. For instance, even when we’ve caught 20 Pokémon, there’s only a $0.1285$ chance of any element in the vector still being $0$

\[\left( 1 - \frac{1}{20} \right)^{40} \approx 0.1285.\]

One thing that Professor Oak must do then is to stop being so stingy and allocate more space for the Bloom filter. The other important thing to keep note of is the number of hash functions to use. If we have too many hash functions, then our Bloom filter will be quite slow, not to mention the entries in our Bloom filter will fill up faster with more hash functions in use. Alternatively, if we limit our hash functions, that increases the chance that we obtain a higher number of false positives.

Luckily, if we select both $m$ and $n$, the following formula is known to give us an optimal value for $k$, the number of hash functions to use

\[k = (m/n)\ln(2),\]

such that the probability that the Bloom filter erroneously claims that a Pokémon is in the Pokédex is minimized. Again, that function is given as

\[\left(1 - e^{-kn/m} \right)^k.\]

Using this, let’s try to come up with a Bloom filter that performs better than our toy example for the purposes of being used for a Pokédex.

First, (at least for Pokémon Red and Blue) the total number of Pokémon one can encounter is 151.

Set of Pokemon in the Red/Blue universe.

So we can pick $n = 151$. For $m$, we can pick something quite a bit larger than that, just to be safe. Let’s set $m = 1000$ as the number of entries in our bit vector for our Bloom filter. Using these parameters, we can compute $k$.

from __future__ import division
import math

# Total number of Pokemon in the Blue/Red universe.
n = 151
# Selecting parameter m to be larger than what we require. 
m = 1000
# Calculating the optimal k to determine how many hash functions we should use. 
k = (m//n)*math.log(2,math.e)
print(k)
4.15888308336

With these selected parameters, the probability that we encounter a false positive is

# How likely is it with the parameters n,m, and k that we encounter a false positive?
(1 - math.e**(-k*n/m))**k
0.04189505785699619

Selecting the parameters for a Bloom filter is a relatively indirect process. Are we okay with accepting this for our purposes? Perhaps for catching Pokémon, this probability is low enough to be just fine for our purposes. However, if we want to push it a bit lower, can select a different value for $m$, recompute $k$, and see if the probability is any smaller.

Hopefully, this post was useful to you in understanding what Bloom filters are, how they work, and optimistically achieved this in a relatively entertaining way. If you have any questions, comments, or corrections, please do not hesitate to comment below and let me know! Thanks again for taking the time to read!