Switch branches/tags
Find file
Fetching contributors…
Cannot retrieve contributors at this time
158 lines (118 sloc) 5.19 KB


Metrics provides an instrumentation toolkit to measure the behavior of your critical systems while they're running in production.



Counters count things.

They implement: inc, dec, clear

They expose: count

var counter = new metrics.Counter
counter.count // 2
counter.count // 0


Histograms sample a dataset to get a sense of its distribution. These are particularly useful for breaking down the quantiles of how longs things take (requests, method calls, etc). Metrics supports uniform distributions and exponentially decaying samples. Sample sizes and parameters of the distribution are all highly configurable, but the defaults will probably suit your needs. Exponential decay histograms favor more recent data, which is typically what you want

They implement: update, mean, stdDev, percentiles, clear

They expose: count, min, max, sum

var hist1 = new metrics.createExponentialDecayHistogram()
  , hist2 = new metrics.createUniformHistogram();
  hist2.mean(); // 2


A meter tracks how often things happen. It exposes a 1 minute rate, a 5 minute rate, and a 15 minute rate using exponentially weighted moving averages (the same strategy that unix load average takes).

They implement: mark, oneMinuteRate, fiveMinuteRate, fifteenMinuteRate, meanRate

var meter = new metrics.Meter
meter.meanRate(); // depends on how fast you called mark()


A Timer is a combination of a meter and a histogram. It samples timing data and rate the data is coming in. Everything you could possibly want!

They implement: update, mark, clear, count, min, max, mean, stdDev, percentiles, oneMinuteRate, fiveMinuteRate, fifteenMinuteRate, meanRate

var  timer = new metrics.Timer;

How to Use

Import Metrics

metrics = require('metrics');

Start a metrics Server

var metricsServer = new metrics.Server(config.metricsPort || 9091);

Servers are only one way to report your metrics. It's actually a thin layer on top of metrics.Report, which you could use to build other reporting mechanisms.

Add the metrics to the server

metricsServer.addMetric('', counter);
metricsServer.addMetric('', hist1);
metricsServer.addMetric('', hist2);
metricsServer.addMetric('', meter);
metricsServer.addMetric('', timer);

Advanced Usage

Typical production deployments have multiple node processes per server. Rather than each process exposing metrics on different ports, it makes more sense to expose the metrics from the "master" process. Writing a thin wrapper around this api to perform the process communication is trivial, with a message passing setup, the client processes could look something like this:

var Metric = exports = module.exports = function Metrics(messagePasser, eventType) {
  this.messagePasser = messagePasser;
  this.eventType = eventType;

Metric.prototype.newMetric = function(type, eventType) {
    method: 'createMetric'
    , type: type
    , eventType: eventType
Metric.prototype.forwardMessage = function(method, args) {
    method: 'updateMetric'
    , metricMethod: method
    , metricArgs: args
    , eventType: this.eventType

Metric.prototype.update = function(val) { return this.forwardMessage('update', [val]); }
Metric.prototype.mark = function(n) { return this.forwardMessage('mark', [n]); } = function(n) { return this.forwardMessage('inc', [n]); }
Metric.prototype.dec = function(n) { return this.forwardMessage('dec', [n]); }
Metric.prototype.clear = function() { return this.forwardMessage('clear'); }

And the server side that receives the createMetric and updateMetric rpcs could look something like this:

  createMetric: function(msg) {
      if (metricsServer) {
        msg.type = msg.type[0].toUpperCase() + msg.type.substring(1)
        metricsServer.addMetric(msg.eventType, new metrics[msg.type]);
  updateMetric: function(msg) {
    if (metricsServer) {
    var namespaces = msg.eventType.split('.')
      , event = namespaces.pop()
      , namespace = namespaces.join('.');
    var metric = metricsServer.trackedMetrics[namespace][event];
    metric[msg.metricMethod].apply(metric, msg.metricArgs);

For multiple server deployments, you have more options, but the best approach will be highly application dependent. Best of luck, and always be tracking!

How to Collect

Using the metrics server you can hit the server on your configured port and you'll get a json representation of your metrics. You should collect these periodically to generate timeseries to monitor the longterm health of your application. The metrics.Reporting object would let you write to a log periodically or however else you'd like to expose your metrics.