Skip to content

Commit

Permalink
fill out README
Browse files Browse the repository at this point in the history
  • Loading branch information
josephruscio committed Sep 12, 2009
1 parent 6626079 commit b762303
Showing 1 changed file with 168 additions and 1 deletion.
169 changes: 168 additions & 1 deletion README
Original file line number Original file line Diff line number Diff line change
@@ -1,2 +1,169 @@
Aggregate is a ruby implementation of a statistics aggregator including histogram support 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/)


Aggregates are easy to instantiate, populate with sample data, and examine
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

Perhaps more importantly than the basic aggregate statistics detailed above
Aggregate also maintains a histogram of samples. Good explanation of why
its important: http://37signals.com/svn/posts/1836-the-problem-with-averages

The histogram is maintained 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. By default this is 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.

Linear histograms are specified with the three values low, high, and width.
Low and high specifiy 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! Any values added to the
aggregate that fall outside of the histogram range are recorded 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

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

Finally Aggregate contains sophisticated pretty-printing support that for
any given number of columns >= 80 (defaults to 80) and sample distribution
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)

The following code populates both a binary and linear histogram with the same
set of 65536 values generated by rand to produce two histograms:

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

** OUTPUT **
** 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

** 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.

** 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.

0 comments on commit b762303

Please sign in to comment.