Probabilistic data structures. Part 2. Cardinality

tech talk @ ferret
Andrii Gakhov
PROBABILISTIC DATA STRUCTURES
ALL YOU WANTED TO KNOW BUT WERE AFRAID TO ASK
PART 2: CARDINALITY

CARDINALITY
Agenda:
▸ Linear Counting
▸ LogLog, SuperLogLog, HyperLogLog, HyperLogLog++

• To determine the number of distinct elements, also
called the cardinality, of a large set of elements
where duplicates are present
Calculating the exact cardinality of a multiset requires an amount of memory
proportional to the cardinality, which is impractical for very large data sets.
THE PROBLEM

LINEAR COUNTING: ALGORITHM
• Linear counter is a bit map (hash table) of size m (all
elements set to 0 at the beginning).
• Algorithm consists of a few steps:
• for every element calculate hash function and set the
appropriate bit to 1
• calculate the fraction V of empty bits in the structure  
(divide the number of empty bits by the bit map size m )
• estimate cardinality as n ≈ -m ln V

LINEAR COUNTING: EXAMPLE
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
• Consider linear counter with 16 bits (m=16)
• Consider MurmurHash3 as the hash function h 
(to calculate the appropriate index, we divide result by mod 16)
• Set of 10 elements: “bernau”, “bernau”, “bernau”,
“berlin”, “kiev”, “kiev”, “new york”, “germany”, “ukraine”,
“europe” (NOTE: the real cardinality n = 7)
h(“bernau”) = 4, h(“berlin”) = 4, h(“kiev”) = 6, h(“new york”) = 6,
h(“germany”) = 14, h(“ukraine”) = 7, h(“europe”) = 9
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0 0 0 0 1 0 1 1 0 1 0 0 0 0 1 0

LINEAR COUNTING: EXAMPLE
number of empty bits: 11
m = 16
V = 11 / 16 = 0.6875
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0 0 0 0 1 0 1 1 0 1 0 0 0 0 1 0
• Cardinality estimation is
n ≈ - 16 * ln (0.6875) = 5.995

LINEAR COUNTING: READ MORE
• http://dblab.kaist.ac.kr/Prof/pdf/Whang1990(linear).pdf
• http://www.codeproject.com/Articles/569718/
CardinalityplusEstimationplusinplusLinearplusTimep

HYPERLOGLOG: INTUITION
• The cardinality of a multiset of uniformly distributed numbers can be estimated
by the maximum number of leading zeros in the binary representation of each
number. If such value is k, then the number of distinct elements in the set is 2k
P(rank=1) = 1/2 - probability to find a binary representation, that starts with 1
P(rank = 2) = 1/2
2
- probability to find a binary representation, that start with 01
…
P(rank=k) = 1/2
k
rank = number of leading zeros + 1, e.g. rank(f) = 3
0 0 1 0 0 0 1 0 0 0 1 0 0 0 0 0
0 1 2
leading zeros
3
f =
• Therefore, for 2k
binary representations we shell find at least one
representation with rank = k
• If we remember the maximal rank we’ve seen and it’s equal to k, then we can
use 2k
as the approximation of the number of elements

HYPERLOGLOG
• proposed by Flajolet et. al., 2007
• an extension of the Flajolet–Martin algorithm (1985)
• HyperLogLog is described by 2 parameters:
• p – number of bits that determine a bucket to use averaging 
(m = 2p
is the number of buckets/substreams)
• h - hash function, that produces uniform hash values
• The HyperLogLog algorithm is able to estimate cardinalities of
> 109
with a typical error rate of 2%, using 1.5kB of memory
(Flajolet, P. et al., 2007).

HYPERLOGLOG: ALGORITHM
• HyperLogLog uses randomization to approximate the
cardinality of a multiset.This randomization is achieved by
using hash function h
• Observe the maximum number of leading zeros that for all
hash values:
• If the bit pattern 0L−1 1 is observed at the beginning of a
hash value (so, rank = L), then a good estimation of the size
of the multiset is 2L.

HYPERLOGLOG: ALGORITHM
• Stochastic averaging is used to reduce the large variability:
• The input stream of data elements S is divided into m substreams Si using
the ﬁrst p bits of the hash values (m = 2p)
.
• In each substream, the rank (after the initial p bits that are used for
substreaming) is measured independently.
• These numbers are kept in an array of registers M, where M[i] stores the
maximum rank it seen for the substream with index i.
• The cardinality estimation is calculated computes as the normalized bias
corrected harmonic mean of the estimations on the substreams
DVHLL = const(m)⋅m2
⋅ 2
−M j
j=1
m
∑
⎛
⎝⎜
⎞
⎠⎟
−1

HYPERLOGLOG: EXAMPLE
• Consider L=8 bits hash function h
• Index elements “berlin” and “ferret”:
h(“berlin”) = 0110111 h(“ferret”) = 1100011
• Define buckets and calculate values to store: 
(use first p =3 bits for buckets and least L - p = 5 bits for ranks)
• bucket(“berlin”) = 011 = 3 value(“berlin”) = rank(0111) = 2
• bucket(“ferret”) = 110 = 6 value(“ferret”) = rank(0011) = 3
• Let’s use p=3 bits to define a bucket (then m=23
=8 buckets).
1 2 3 4 5 6 7
0 0 0 0 0 0 0 0
0
M
1 2 3 4 5 6 7
0 0 0 2 0 0 3 0
0
M

HYPERLOGLOG: EXAMPLE
• Estimate the cardinality be the HLL formula (C ≈ 0.66):
DVHLL ≈ 0.66 * 82
/ (2-2
+ 2-4
) = 0.66 * 204.8 ≈ 135≠3
• Index element “kharkov”:
• h(“kharkov”) = 1100001
• bucket(“kharkov”) = 110 = 6 value(“kharkov”) = rank(0001) = 4
• M[6] = max(M[6], 4) = max(3, 4) = 4
1 2 3 4 5 6 7
0 0 0 2 0 0 4 0
0
M
NOTE: For small cardinalities HLL has a strong bias!!!

HYPERLOGLOG: PROPERTIES
• Memory requirement doesn't grow linearly with L (unlike MinCount or
Linear Counting) - for hash function of L bits and precision p, required
memory:
• original HyperLogLog uses 32 bit hash codes, which requires 5 · 2
p
bits
• It’s not necessary to calculate the full hash code for the element
• ﬁrst p bits and number of leading zeros of the remaining bits are
enough
• There are no evidence that some of popular hash functions (MD5, Sha1,
Sha256, Murmur3) performs signiﬁcantly better than others.
log2 L +1− p( )⎡⎢ ⎤⎥⋅2p
bits

HYPERLOGLOG: PROPERTIES
• The standard error can be estimated as:
σ =
1.04
2p
so, if we use 16 bits (p=16) for bucket indices, we receive the
standard error in 0.40625%
• Algorithm has large error for small cardinalities.
• For instance, for n = 0 the algorithm always returns roughly 0.7m
• To achieve better estimates for small cardinalities, use
LinearCounting below a threshold of 5m/2

HYPERLOGLOG: APPLICATIONS
• PFCOUNT in Redis returns the approximated cardinality
computed by the HyperLogLog data structure  
(http://antirez.com/news/75)
• Redis implementation uses 12Kb per key to count with a standard
error of 0.81%, and there is no limit to the number of items you can
count, unless you approach 264 items

HYPERLOGLOG: READ MORE
• http://algo.inria.fr/flajolet/Publications/DuFl03-LNCS.pdf
• http://algo.inria.fr/flajolet/Publications/FlFuGaMe07.pdf
• https://stefanheule.com/papers/edbt13-hyperloglog.pdf
• https://highlyscalable.wordpress.com/2012/05/01/
probabilistic-structures-web-analytics-data-mining/
• https://hal.archives-ouvertes.fr/file/index/docid/465313/
filename/sliding_HyperLogLog.pdf
• http://stackoverflow.com/questions/12327004/how-does-
the-hyperloglog-algorithm-work

HYPERLOGLOG++
• proposed by Stefan Heule et. al., 2013 for Google PowerDrill
• an improved version of HyperLogLog (Flajolet et. al., 2007)
• HyperLogLog++ is described by 2 parameters:
• p – number of bits that determine a bucket to use averaging 
(m = 2p
is the number of buckets/substreams)
• h - hash function, that produces uniform hash values
• The HyperLogLog++ algorithm is able to estimate cardinalities
of ~ 7.9 · 10
9
with a typical error rate of 1.625%, using 2.56KB of
memory (Micha Gorelick and Ian Ozsvald, High Performance
Python, 2014).

HYPERLOGLOG++: IMPROVEMENTS
• use 64-bit hash function
• algorithm that only uses the hash code of the input values is limited by the
number of bits of the hash codes when it comes to accurately estimating
large cardinalities
• In particular, a hash function of L bits can at most distinguish 2L
different
values, and as the cardinality n approaches 2L
hash collisions become
more and more likely and accurate estimation gets impossible
• if the cardinality approaches 264 ≈ 1.8 · 1019, hash collisions become a problem
• bias correction
• original algorithm overestimates the real cardinality for small sets, but
most of the error is due to bias.
• storage efﬁciency
• uses different encoding strategies for hash values, variable length
encoding for integers, difference encoding

HYPERLOGLOG++ VS HYPERLOGLOG
• accuracy is signiﬁcantly better for large range of cardinalities
and equally good on the rest
• sparse representation allows for a more adaptive use of memory
• if the cardinality n is much smaller then m, then HyperLogLog++
requires signiﬁcantly less memory
• For cardinalities between 12000 and 61000, the bias correction
allows for a lower error and avoids a spike in the error when
switching between sub-algorithms.
• 64 bit hash codes allow the algorithm to estimate cardinalities well
beyond 1 billion

HYPERLOGLOG++: APPLICATIONS
• cardinality metric in Elasticsearch is based on the
HyperLogLog++ algorithm for big cardinalities (adaptive
counting)
• Apache DataFu, collection of libraries for working with
large-scale data in Hadoop, has an implementation of
HyperLogLog++ algorithm

HYPERLOGLOG++: READ MORE
• http://static.googleusercontent.com/media/
research.google.com/en//pubs/archive/40671.pdf
• https://research.neustar.biz/2013/01/24/hyperloglog-
googles-take-on-engineering-hll/

▸ @gakhov
▸ linkedin.com/in/gakhov
▸ www.datacrucis.com
THANK YOU

Probabilistic data structures. Part 2. Cardinality

More Related Content

What's hot

Viewers also liked

Similar to Probabilistic data structures. Part 2. Cardinality

More from Andrii Gakhov

Recently uploaded

Probabilistic data structures. Part 2. Cardinality