Efficient storage of data in memory

When dealing with big data, minimizing the amount of memory used is critical to avoid having to use disk-based access, which can be 100,000 times slower than random access. This notebook deals with ways to minimizee data storage for several common use cases:

  • Large arrays of homogenous data (often numbers)
  • Large string collections
  • Counting distinct values
  • Yes/No responses to queries

Methods covered range from the mundane (use numpy arrays rather than lists), to classic but less well-known data structures (e.g. prefix trees or tries) to algorithmically ingenious probabilistic data structures (e.g. bloom filter and hyperloglog).

Selective retrieval from disk-based storage

We have alrady seen that there are many ways to retrieve only the parts of the data we need now into memory at this particular moment. Options include

  • generators (e.g. to read a file a line at a time)
  • numpy.memmap
  • HDF5 via h5py
  • Key-value stores (e.g. redis)
  • SQL and NoSQL databases (e.g. sqlite3)

Storing numbers

Less memory is used when storing numbers in numpy arrays rather than lists.


Using only the precision needed can also save memory


Storing strings

def flatmap(func, items):
    return it.chain.from_iterable(map(func, items))
def flatten(xss):
    return (x for xs in xss for x in xs)
with open('data/Ulysses.txt') as f:
    word_list = list(flatten(line.split() for line in f))
target = 'WARRANTIES'
%timeit -r1 -n1 word_list.index(target)
1 loops, best of 1: 7.97 ms per loop
import bisect
%timeit -r1 -n1 bisect.bisect(word_list, target)
1 loops, best of 1: 31.8 µs per loop
word_set = set(word_list)
%timeit -r1 -n1 target in word_set
1 loops, best of 1: 2.25 µs per loop
! pip install hat_trie
%load_ext memory_profiler
from hat_trie import Trie
%memit word_trie = Trie(word_list)
peak memory: 116.97 MiB, increment: 0.39 MiB
%timeit -r1 -n1 target in word_trie
1 loops, best of 1: 5.86 µs per loop

Data Sketches

A sketch is a probabilistic algorithm or data structure that approximates some statistic of interest, typically using very little memory and processing time. Often they are applied to streaming data, and so must be able to incrementally process data. Many data sketches make use of hash functions to distribute data into buckets uniformly. Typically, data sketches have the following desirable properties

  • sub-linear in space
  • single scan
  • can be parallelized
  • can be combined (merge)

Some statistics that sketches have been used to estimate include

  • indicator variables (event detection)
  • counts
  • quantiles
  • moments
  • entropy

Packages for data sketches in Python are relatively immmature, and if you are interested, you could make a large contribution by creating a comprehensive open source library of data sketches in Python.

The Morris counter is used as a simple illustration of a probabilistic data structure, with the standard trade-off of using less memory in return for less accuracy. The algorithm is extremely simple - keep a counter \(c\) that represents the exponent - that is, when the Morris counter is \(c\), the estimated count is \(2^c\). The probabilistic part comes from the way that the counter is incremented by comparing a uniform random variate to \(1/2^c\).

from random import random

class MorrisCounter:
    def __init__(self, c=0):
        self.c = c

    def __len__(self):
        return 2 ** self.c

    def add(self, item):
        self.c += random() < 1/(2**self.c)
mc = MorrisCounter()
print('True\t\tMorris\t\tRel Error')
for i, word in enumerate(word_list):
    if i%int(.2e5)==0:
        print('%8d\t%8d\t%.2f' % (i, len(mc), 0 if i==0 else abs(i - len(mc))/i))
True                Morris          Rel Error
       0           2        0.00
   20000        8192        0.59
   40000       32768        0.18
   60000       65536        0.09
   80000       65536        0.18
  100000      131072        0.31
  120000      131072        0.09
  140000      131072        0.06
  160000      131072        0.18
  180000      131072        0.27
  200000      131072        0.34
  220000      262144        0.19
  240000      262144        0.09
  260000      524288        1.02

A simple way to increase the accuracy is to have multiple Morris counters and take the average. These two ideas of using a probabilistic calculation and multiple samples to improve precision are the basis for the more useful probabilisitc data structures described below.

mcs = [MorrisCounter() for i in range(10)]
print('True\t\tMorris\t\tRel Error')
for i, word in enumerate(word_list):
    for j in range(10):
    estimate = np.mean([len(m) for m in mcs])
    if i%int(.2e5)==0:
        print('%8d\t%8d\t%.2f' % (i, estimate, 0 if i==0 else abs(i - estimate)/i))
True                Morris          Rel Error
       0           2        0.00
   20000       20889        0.04
   40000       26214        0.34
   60000       36044        0.40
   80000       49152        0.39
  100000       72089        0.28
  120000       75366        0.37
  140000       95027        0.32
  160000      111411        0.30
  180000      131072        0.27
  200000      137625        0.31
  220000      137625        0.37
  240000      144179        0.40
  260000      170393        0.34

The Morris counter is less useful because the degree of memory saved as compared to counting the number of elements exactly is not much unless the numbers are staggeringly huge. In contrast, counting the number of distinct elements exactly requires storage of all distinct elements (e.g. in a set) and hence grows with the cardinality \(n\). Probabilistic data structures known as Distinct Value Sketches can do this with a tiny and fixed memory size.

Examples where counting distinct values is useful:

  • number of unique users in a Twitter stream
  • number of distinct records to be fetched by a databse query
  • number of unique IP addresses accessing a website
  • number of distinct queries submitted to a search engine
  • number of distinct DNA motifs in genomics data sets (e.g. microbiome)

A hash function takes data of arbitrary size and converts it into a number in a fixed range. Ideally, given an arbitrary set of data items, the hash function generates numbers that follow a uniform distribution within the fixed range. Hash functions are immensely useful throughout computer science (for example - they power Python sets and dictionaries), and especially for the generation of probabilistic data structures.

A simple hash function mapping

Note the collisions. If not handled, there is a loss of information. Commonly, practical hash functions return a 32 or 64 bit integer. Also note that there are an arbitrary number of hash functions that can return numbers within a given range.

Note also that because the hash function is deterministic, the same item will always map to the same bin.

def string_hash(word, n):
    return sum(ord(char) for char in word) % n
sentence = "The quick brown fox jumps over the lazy dog."
for word in sentence.split():
    print(word, string_hash(word, 10))
The 9
quick 1
brown 2
fox 3
jumps 9
over 4
the 1
lazy 8
dog. 0

Built-in Python hash function

Help on built-in function hash in module builtins:

hash(obj, /)
    Return the hash value for the given object.

    Two objects that compare equal must also have the same hash value, but the
    reverse is not necessarily true.
for word in sentence.split():
    print('{:<10s} {:24}'.format(word, hash(word)))
The             -184990475008303844
quick           3616884800889772302
brown           6377133929055916905
fox             -611579958660588990
jumps          -5271806623556898369
over            7546130948312823661
the             6678103606492090842
lazy           -1515512778017190090
dog.            3069897472948403276

Using a hash function from the MurmurHash3 library

Note that the hash function accepts a seed, allowing the creation of multiple hash functions. We also display the hash result as a 32-bit binary string.

import mmh3

for word in sentence.split():
    print('{:<10} {:+032b} {:+032b}'.format(word.ljust(10), mmh3.hash(word, seed=1234),
          mmh3.hash(word, seed=4321)))
The        +0001000011111110001001110101100 +1110110100100101010111100011010
quick      -0101111111011110110101100101000 +1000100001101010110000101101100
brown      +1000101010000110110010001110101 -1101101110000000010001100010100
fox        -1000000010010010000111001111011 +0111011111000011001001001110111
jumps      +0000010111000011010000100101010 +0010010001111110100010010110011
over       -0110101101111001001101011111011 -1101110111110010000101101000100
the        -1000000101110000000110011111001 +0001000111100111011000011100101
lazy       -1101011000111111110011111001100 +0010101110101100001000101110000
dog.       +0100110101101111101011110111111 -0101111000110000001011110001011

LogLog family

The binary digits in a (say) 32-bit hash are effectively random, and equivalent to a sequence of fair coin tosses. Hence the probability that we see a run of 5 zeros in the smallest hash so far suggests that we have added \(2^5\) unique items so far. This is the intuition behind the loglog family of Distinct Value Sketches. Note that the biggest count we can track with 32 bits is \(2^{32} = 4294967296\).

The accuracy of the sketch can be improved by averaging results with multiple coin flippers. In practice, this is done by using the first \(k\) bit registers to identify \(2^k\) different coin flippers. Hence, the max count is now \(2 ** (32 - k)\). The hyperloglog algorithm uses the harmonic mean of the \(2^k\) flippers which reduces the effect of outliers and hence the variance of the estimate.

for i in range(1, 15):
    k = 2**i
    hashes = [''.join(map(str, np.random.randint(0,2,32))) for i in range(k)]
    print('%6d\t%s' % (k, min(hashes)))
    2      00010001101110111100011011110100
    4      00001111000110101110111110111011
    8      00011100101100110010101100011110
   16      01000001100111101011001100101001
   32      00001011000101110101000110101010
   64      00000000110100000010001101011101
  128      00000000100101100011100101100111
  256      00000000001101010111001001101010
  512      00000000001001011011001100001100
 1024      00000000001010000100011000011011
 2048      00000000010100111100001101001100
 4096      00000000000000110001101010011100
 8192      00000000000000001100110110000101
16384      00000000000000001101011110101001
pip install hyperloglog
from hyperloglog import HyperLogLog
hll = HyperLogLog(0.01) # accept 1% counting error
print('True\t\tHLL\t\tRel Error')
s = set([])
for i, word in enumerate(word_list):
    if i%int(.2e5)==0:
        print('%8d\t%8d\t\t%.2f' % (len(s), len(h1), 0 if i==0 else abs(len(s) - len(h1))/i))
True                HLL             Rel Error

NameError                                 Traceback (most recent call last)

<ipython-input-1-22150e0b1002> in <module>()
      1 print('True\t\tHLL\t\tRel Error')
      2 s = set([])
----> 3 for i, word in enumerate(word_list):
      4     s.add(word)
      5     hll.add(word)

NameError: name 'word_list' is not defined

Bloom filters

Bloom filters are designed to answer queries about whether a specific item is in a collection. If the answer is NO, then it is definitive. However, if the answer is yes, it might be a false positive. The possibility of a false positive makes the Bloom filter a probabilistic data structure.

A bloom filter consists of a bit vector of length \(k\) initially set to zero, and \(n\) different hash functions that return a hash value that will fall into one of the \(k\) bins. In the construction phase, for every item in the collection, \(n\) hash values are generated by the \(n\) hash functions, and every position indicated by a hash value is flipped to one. In the query phase, given an item, \(n\) hash values are calculated as before - if any of these \(n\) positions is a zero, then the item is definitely not in the collection. However, because of the possibility of hash collisions, even if all the positions are one, this could be a false positive. Clearly, the rate of false positives depends on the ratio of zero and one bits, and there are Bloom filter implementations that will dynamically bound the ratio and hence the false positive rate.

Possible uses of a Bloom filter include:

  • Does a particular sequence motif appear in a DNA string?
  • Has this book been recommended to this customer before?
  • Check if an element exists on disk before performing I/O
  • Check if URL is a potential malware site using in-browser Bloom filter to minimize network communication
  • As an alternative way to generate distinct value counts cheaply (only increment count if Bloom filter says NO)
pip install git+https://github.com/jaybaird/python-bloomfilter.git
from pybloom import ScalableBloomFilter

# The Scalable Bloom Filter grows as needed to keep the error rate small
# The default error_rate=0.001
sbf = ScalableBloomFilter()
for word in word_set:
test_words = ['banana', 'artist', 'Dublin', 'masochist', 'Obama']
for word in test_words:
    print(word, word in sbf)
banana True
artist True
Dublin True
masochist False
Obama False
### Chedck
for word in test_words:
    print(word, word in word_set)
banana True
artist True
Dublin True
masochist False
Obama False
%load_ext version_information
%version_information pybloom, hyperloglog, hat_trie
Python3.5.1 64bit [GCC 4.2.1 (Apple Inc. build 5577)]
OSDarwin 15.4.0 x86_64 i386 64bit
Thu Apr 14 16:01:59 2016 EDT