RDBHistogram

RDBHistogram (Rolling Dynamic range Bucket Histogram) is an histogram algorithm that aims for high precision while preserving memory. Histograms are a great tool to quantify performance measurements (like latency) which tend to have a non-normal distribution with extremely large values.

Statistics captured:

• max
• min
• median
• 75th percentile
• 95th percentile
• 99th percentile
• 99.9th percentile

With the default configuration, the RDBHistogram tracks last minute statistics with 10% accuracy. Compared to other algorithms (the algorithm used by Metrics), the RDB histogram provides better accuracy, predictability and a smaller memory footprint.

Why?

Read the comparison with the popular Metrics Histogram algorithm .

Usage

Create the histogram object

var RDBHistogram = require('rdb-histogram');
var histogram = new RDBHistogram();

Put values into the histogram

histogram.update(value);

Get the statistics

histogram.toJSON();

The returned json has the form

{
  min: 182.0700962934643,
  max: 875.7033819006756,
  count: 40000,
  median: 510.0446645318259,
  p75: 572.0086002136738,
  p95: 659.6095291048899,
  p99: 720.9989285909948,
  p999: 783.4347930209091,
  numBuckets: 117
}

Where most values are self explanatory. The numBuckets field is an indication of the number of buckets used internally by the histogram. The memory used by the histogram is 3 numbers for each bucket - in the above example 117 buckets means 351 numbers are used internally.

patching node-measured

node-measured is a node.js implementation of the excellent Metrics library. Like the metrics library it has the failing of the Metrics histogram algorithm which is based on sampling and suffers from inherent inaccuracy, and more importantly, unpredictable inaccuracy - one cannot calculate what the inaccuracy will be given a certain input data or rate of input samples simply because the algorithm is preserves a set of samples that is probably to be representative but is not guaranteed to be so.

The RDBHistogram provides a higher precision algorithm with a lesser memory footprint and a predicable accuracy - accuracy is only a factor of the RDBHistogram configuration.

The RDBHistogram includes a patch function that patches node-measured, replacing it's Histogram algorithm with a compatible RDBHistogram. To parch node-measured use the following:

var measured = require('measured');
var RDBHistogram = require('rdb-histogram');
RDBHistogram.patchMeasured(measured);

then one can use node-measured histogram in a compatible way -

var histogram = new measured.Histogram();

or

var collection = measured.createCollection();
var histogram = collection.histogram('metric-name');

Configuration

The histogram accepts a single configuration object with the following properties:

• historyInterval - The length in mSec of a single time bucket. Defaults to 15000 - 15 seconds.
• historyLength - The number of time buckets to use. Defaults to 4 time buckets.
• minValue - The minimal value, that anything under this value is considered as part of a single minimum bucket. Defaults to 1.
• mainScale - The number of buckets used between each scale (1..10). The default value of 5 ensures 40% (10^(1/5)) accuracy (before considering subScale).
• subScale - The number of sub-buckets used to breakdown an interesting bucket (a bucket that has one of the percentiles in the output statistics). The default value of 5 ensures accuracy of about 10% (10^(1/25)).