aggregate_afurmanov

h1. Aggregate

By Joseph Ruscio

Aggregate is an intuitive ruby implementation of a statistics aggregator including both default and configurable histogram support. It does this without recording/storing any of the actual sample values, making it suitable for tracking statistics across millions/billions of sample without any impact on performance or memory footprint. Originally inspired by the Aggregate support in "SystemTap.":http://sourceware.org/systemtap

h2. Getting Started

Aggregates are easy to instantiate, populate with sample data, and then inspect for common aggregate statistics:

#After instantiation use the << operator to add a sample to the aggregate:
stats = Aggregate.new

loop do
  # Take some action that generates a sample measurement
  stats << sample
end

# The number of samples
stats.count

# The average
stats.mean

# Max sample value
stats.max

# Min sample value
stats.min

# The standard deviation
stats.std_dev

h2. Histograms

Perhaps more importantly than the basic aggregate statistics detailed above Aggregate also maintains a histogram of samples. For anything other than normally distributed data are insufficient at best and often downright misleading 37Signals recently posted a terse but effective "explanation":http://37signals.com/svn/posts/1836-the-problem-with-averages of the importance of histograms. Aggregates maintains its histogram internally as a set of "buckets". Each bucket represents a range of possible sample values. The set of all buckets represents the range of "normal" sample values.

h3. Binary Histograms

Without any configuration Aggregate instance maintains a binary histogram, where each bucket represents a range twice as large as the preceding bucket i.e. [1,1], [2,3], [4,5,6,7], [8,9,10,11,12,13,14,15]. The default binary histogram provides for 128 buckets, theoretically covering the range [1, (2^127) - 1] (See NOTES below for a discussion on the effects in practice of insufficient precision.)

Binary histograms are useful when we have little idea about what the sample distribution may look like as almost any positive value will fall into some bucket. After using binary histograms to determine the coarse-grained characteristics of your sample space you can configure a linear histogram to examine it in closer detail.

h3. Linear Histograms

Linear histograms are specified with the three values low, high, and width. Low and high specify a range [low, high) of values included in the histogram (all others are outliers). Width specifies the number of values represented by each bucket and therefore the number of buckets i.e. granularity of the histogram. The histogram range (high - low) must be a multiple of width:

#Want to track aggregate stats on response times in ms
response_stats = Aggregate.new(0, 2000, 50)

The example above creates a linear histogram that tracks the response times from 0 ms to 2000 ms in buckets of width 50 ms. Hopefully most of your samples fall in the first couple buckets!

h3. Histogram Outliers

An Aggregate records any samples that fall outside the histogram range as outliers:

# Number of samples that fall below the normal range
stats.outliers_low

# Number of samples that fall above the normal range
stats.outliers_high

h3. Histogram Iterators

Once a histogram is populated Aggregate provides iterator support for examining the contents of buckets. The iterators provide both the number of samples in the bucket, as well as its range:

#Examine every bucket
@stats.each do |bucket, count|
end

#Examine only buckets containing samples
@stats.each_nonzero do |bucket, count|
end

h3. Histogram Bar Chart

Finally Aggregate contains sophisticated pretty-printing support to generate ASCII bar charts. For any given number of columns >= 80 (defaults to 80) and sample distribution the to_s method properly sets a marker weight based on the samples per bucket and aligns all output. Empty buckets are skipped to conserve screen space.

# Generate and display an 80 column histogram
puts stats.to_s

# Generate and display a 120 column histogram
puts stats.to_s(120)

This code example populates both a binary and linear histogram with the same set of 65536 values generated by rand to produce the two histograms that follow it:

require 'rubygems'
require 'aggregate'

# Create an Aggregate instance
binary_aggregate = Aggregate.new
linear_aggregate = Aggregate.new(0, 65536, 8192)

65536.times do
  x = rand(65536)
  binary_aggregate << x
  linear_aggregate << x
end

puts binary_aggregate.to_s
puts linear_aggregate.to_s

h4. Binary Histogram

value |------------------------------------------------------------------| count
    1 |                                                                  |     3
    2 |                                                                  |     1
    4 |                                                                  |     5
    8 |                                                                  |     9
   16 |                                                                  |    15
   32 |                                                                  |    29
   64 |                                                                  |    62
  128 |                                                                  |   115
  256 |                                                                  |   267
  512 |@                                                                 |   523
 1024 |@                                                                 |   970
 2048 |@@@                                                               |  1987
 4096 |@@@@@@@@                                                          |  4075
 8192 |@@@@@@@@@@@@@@@@                                                  |  8108
16384 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@                                  | 16405
32768 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@| 32961
      ~
Total |------------------------------------------------------------------| 65535

h4. Linear (0, 65536, 4096) Histogram

value |------------------------------------------------------------------| count
    0 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@  |  4094
 4096 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|  4202
 8192 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@  |  4118
12288 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@   |  4059
16384 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@    |  3999
20480 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@  |  4083
24576 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@  |  4134
28672 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |  4143
32768 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |  4152
36864 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@   |  4033
40960 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@   |  4064
45056 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@   |  4012
49152 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@   |  4070
53248 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@  |  4090
57344 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@  |  4135
61440 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |  4144
Total |------------------------------------------------------------------| 65532

We can see from these histograms that Ruby's rand function does a relatively good job of distributing returned values in the requested range.

h2. Examples

Here's an example of a "handy timing benchmark":http://gist.github.com/187669 implemented with aggregate.

h2. NOTES

Ruby doesn't have a log2 function built into Math, so we approximate with log(x)/log(2). Theoretically log( 2^n - 1 )/ log(2) == n-1. Unfortunately due to precision limitations, once n reaches a certain size (somewhere > 32) this starts to return n. The larger the value of n, the more numbers i.e. (2^n - 2), (2^n - 3), etc fall trap to this errors. Could probably look into using something like BigDecimal, but for the current purposes of the binary histogram i.e. a simple coarse-grained view the current implementation is sufficient.