TaskCluster Metrics and Monitoring Library

Build Status npm License

A convenient library to wrap up all of the pieces needed for a Taskcluster service to record metrics with Statsum and report errors with Sentry. By default it will report any errors that cause the process to exit, and report as warnings any errors that cause stats writing to not work. To disable any of these, you can see the Options and Defaults section below.

Process monitoring can be turned on by using the monitor.resources(<process name>) function where <process name> will generally end up being something like web or worker.

Taskcluster has some generic concepts that are able to be monitored easily using utility functions in this package. The Usage section lists these cases and shows how to use this package to measure them.

Changelog

View the changelog on the releases page.

Requirements

This is tested on and should run on any of node {8, 9, 10}.

Usage

This library must be provided with Taskcluster credentials that have the following scopes:

  • auth:sentry:<name of project>
  • auth:statsum:<name of project>

First, create a monitor. This is most often done in a taskcluster-lib-loader component.

let load = loader({
  monitor: {
    requires: ['process', 'profile', 'cfg'],
    setup: ({process, profile, cfg}) => monitor({
      rootUrl: 'https://taskcluster.example.com',
      credentials: cfg.taskcluster.credentials,
      projectName: 'taskcluster-foo',
      mock: cfg.monitor.mock,  // false in production, true in testing
      process,
    }),
  },
});

The available options are:

  • rootUrl - the rootUrl for this Taskcluster instance; used with credentials to fetch statsum and sentry keys
  • credentials: {clientId: '...', accessToken: '...'} - Taskcluster credentials (no default - must be provided)
  • projectName - The project that will be written under to Statsum and Sentry.
  • patchGlobal - If true (the default), any uncaught errors in the service will be reported to Sentry.
  • reportStatsumErrors - If true (the default), any errors reporting to Statsum will be reported to Sentry.
  • process - If set to a string that identifies this process, cpu and memory usage of the process will be reported on an interval. Note: This can also be turned on by monitor.resources(...) later if wanted. That allows for gracefully stopping as well.
  • statsumToken - a function that will return a Statsum token (async (projectName) => {token, expires, baseUrl}); the default value uses credentials to fetch a token from the Auth service.
  • sentryDSN - a function that will return a Sentry DSN (async (projectName) => {dsn: {secret: '...'}, expires}); the default value uses credentials to fetch a DSN from the Auth service.
  • sentryOptions:options given to the raven.Client constructor
  • mock - If true, the monitoring object will be a fake that stores data for testing but does not report it (for testing).
  • enable - If false, the monitoring object will only report to the console (but not store data; for deployments without monitoring)
  • aws - If provided, these should be of the form {credentials: {accessKeyId: '...', secretAccessKey: '...'}, region: '...'}
  • logName - If provided, this should be the name of a AWS Firehose deliveryStream that can be written to with the aws creds
  • gitVersion - git version (for correlating errors); or..
  • gitVersionFile - file containing git version (relative to app root)

Measuring and Counting Things

More details on the usage of measure and count can be found at the Statsum client.

To record a current measurement of a named value:

monitor.measure('foo', 10);

To increment the count of some value:

monitor.count('bar');
monitor.count('bar', 4); // increment by 4

To construct an object capable of measuring and counting, but which adds a prefix to the measured and counted names, use

let thingMonitor = monitor.prefix('thing');
thing.measure('foo', 11);
thing.count('bar');

The monitor will automatically flush its statistics to statsum periodically, but you can force a flush with

await monitor.flush();

Reporting Errors

There are lots of options for reporting errors:

// Report error as a string, without a stacktrace
monitor.reportError('Something went wrong!');
// Report error (from catch-block or something like that)
monitor.reportError(new Error("..."));
// Report error as a warning
monitor.reportError(new Error("..."), 'warning');
// Report error as info
monitor.reportError(new Error("..."), 'info');
// Report error as debug
monitor.reportError(new Error("..."), 'debug');
// Report an error with tags
monitor.reportError(new Error("..."), {foo: 'bar'});
// Report a warningr with tags
monitor.reportError(new Error("..."), 'warning', {foo: 'bar'});

Monitoring CPU & Memory

// Begin monitoring CPU & Memory
let stopMonitor = monitor.resources('web');

you can later call stopMonitor() to gracefully shut down the monitoring.

Timing Functions/Promises

Often we wish to measure how long time an operation takes, synchronous or asynchronous, this can done using the monitor.timer(key, funcOrPromise) method. It takes a key (as name of the metric) and a function or promise to measure the time of. If the function returns a promise, it'll include the time it takes for the promise to resolve.

The following examples are valid usages:

// Timing a synchronous operation
let root = monitor.timer('compute-sqrt', () => {
  return Math.sqrt(25);
})
assert(root === 5);

// Timing a single asynchronous function (promise)
let task = await monitor.timer('load-task', queue.task(taskId));
assert(task.workerType == '...'); // task is the task definition response

// Timing an asynchronous function
let task = await monitor.timer('poll-for-task', async () => {
  while (true) {
    try {
      return await queue.task(taskId);
    } catch (err) {
      // Ignore error and try again
      // In the real would you want a maximum time before you stop polling
      // And probably some sleeping between each polling attempt...
    }
  }
});
assert(task.workerType == '...'); // task is the task definition response

Rejected promises and errors will bubble up, and the time be will measured and recoded just like successful functions or promises.

Timing Handlers

A common pattern in Taskcluster projects is to have handler functions in a worker that take a message as an argument and perform some action. These can be timed (in milliseconds) by wrapping them with taskcluster-lib-monitor:

let monitor = await monitoring({
  projectName: 'tc-stats-collector',
  credentials: {clientId: 'test-client', accessToken: 'test'},
});

let listener = new taskcluster.PulseListener({
  credentials: {clientId: 'test-client', accessToken: 'test'},
  queueName: 'a-queue-name',
});

let handler = function(message) {
  console.log(message);
};

listener.on('message', monitor.timedHandler('logging-listener', handler));

Express Timing Middleware

Most Taskcluster services are Express services. We can easily time how long endpoints take to respond to requests by inserting taskcluster-lib-monitor as middleware:

let monitor = await monitoring({
  projectName: 'tc-stats-collector',
  credentials: {clientId: 'test-client', accessToken: 'test'},
});

// Express setup, etc.

middleware.push(monitor.expressMiddleware('name_of_function'));

This is already integrated in taskcluster-lib-api and probably doesn't need to be implemented in your service on its own.

Timing AWS SDK Calls

Oftentimes a lot of a service's time will be spent interacting with AWS services. These interactions can be measured as in the following example:

let aws = require('aws-sdk');
let ec2 = new aws.EC2({region: 'us-west-2'});
monitor.patchAWS(ec2);
await ec2.describeAvailabilityZones().promise().catch(err => {
  debug('Ignored ec2 error, we measure duration, not success, err: ', err);
});

Timing Arbitary Steps

If none of the above options are convenient for you, you can also just start and stop timers whenever you want. A timer may only be started and measured once. Any attempts over that will cause it to throw an Error.

let doodad = monitor.timeKeeper('metricName');
// Do some stuff here that takes some amount of time
// ...
// Keep doing stuff here however long you like
doodad.measure();

Audit Logs

For the time being, this is restricted to services that have use AWS credentials directly rather than via accessing via the auth service. Given a set of credentials that allow writing to a Kinesis Firehose and the name of that Firehose, this will allow writing arbitrary JSON blobs to that endpoint. The blobs will end up in S3 for permanent storage. We use this for things like audit logs that we want to keep for a long time. Records must be less than 1MB when stringified.

    let monitor = await monitoring({
      ...,
      aws: {credentials: {accessKeyId: 'foo', secretAccessKey: 'bar'}, region: 'us-east-1'},
      logName: 'audit-log-stream',
    });
    monitor.log({foo: 'bar', baz: 233}); // This will be submitted on a timed interval
    await monitor.flush(); // This will await all records being submitted

Testing

yarn install and yarn test. You can set DEBUG=taskcluster-lib-monitor,test if you want to see what's going on. There are no keys required to test this library.

Hacking

New releases should be tested on Travis to allow for all supported versions of Node to be tested. Once satisfied that it works, new versions should be created with yarn version rather than by manually editing package.json and tags should be pushed to Github. Make sure to update the changelog!

License

Mozilla Public License Version 2.0