Skip to content

Instantly share code, notes, and snippets.



Last active May 25, 2021
What would you like to do?
zero-downtime node.js app runner
// This script will boot app.js with the number of workers
// specified in WORKER_COUNT.
// The master will respond to SIGHUP, which will trigger
// restarting all the workers and reloading the app.
var cluster = require('cluster');
var workerCount = process.env.WORKER_COUNT || 2;
// Defines what each worker needs to run
// In this case, it's app.js a simple node http app
cluster.setupMaster({ exec: 'app.js' });
// Gets the count of active workers
function numWorkers() { return Object.keys(cluster.workers).length; }
var stopping = false;
// Forks off the workers unless the server is stopping
function forkNewWorkers() {
if (!stopping) {
for (var i = numWorkers(); i < workerCount; i++) { cluster.fork(); }
// A list of workers queued for a restart
var workersToStop = [];
// Stops a single worker
// Gives 60 seconds after disconnect before SIGTERM
function stopWorker(worker) {
var killTimer = setTimeout(function() {
}, 60000);
// Ensure we don't stay up just for this setTimeout
// Tell the next worker queued to restart to disconnect
// This will allow the process to finish it's work
// for 60 seconds before sending SIGTERM
function stopNextWorker() {
var i = workersToStop.pop();
var worker = cluster.workers[i];
if (worker) stopWorker(worker);
// Stops all the works at once
function stopAllWorkers() {
stopping = true;
console.log('stopping all workers');
for (var id in cluster.workers) {
// Worker is now listening on a port
// Once it is ready, we can signal the next worker to restart
cluster.on('listening', stopNextWorker);
// A worker has disconnected either because the process was killed
// or we are processing the workersToStop array restarting each process
// In either case, we will fork any workers needed
cluster.on('disconnect', forkNewWorkers);
// HUP signal sent to the master process to start restarting all the workers sequentially
process.on('SIGHUP', function() {
console.log('restarting all workers');
workersToStop = Object.keys(cluster.workers);
// Kill all the workers at once
process.on('SIGTERM', stopAllWorkers);
// Fork off the initial workers
console.log('app master',, 'booted');

This comment has been minimized.

Copy link

@deyhle deyhle commented Sep 17, 2014

Great script, thank you!
However, when having more than two workers and restarting it can take quite some time until all workers are restarted when they are having open connections. Before the next worker is restarted, it waits for 60 seconds. Wouldn't it be better if the script would try to already restart the next worker instead of waiting or would this make it possible that no worker is available during restart for a brief period?


This comment has been minimized.

Copy link

@thelinuxlich thelinuxlich commented Apr 16, 2015

Workers are not receiving message on disconnect, what can it be?


This comment has been minimized.

Copy link

@leonardjaviniar leonardjaviniar commented May 26, 2015

Starting and stopping by about 3 - 5 times causes the error below when connecting to a database:

     throw er; // Unhandled 'error' event
Error: write ENOTSUP - cannot write to IPC channel.  
    at errnoException (child_process.js:1001:11)  
    at (child_process.js:465:16)  
    at Worker.send (cluster.js:406:21)  
    at sendInternalMessage (cluster.js:399:10)  
    at handleResponse (cluster.js:177:5)  
    at respond (cluster.js:192:5)  
    at Object.messageHandler.queryServer (cluster.js:247:5)  
    at handleMessage (cluster.js:197:32)  
    at ChildProcess.emit (events.js:117:20)  
    at handleMessage (child_process.js:322:10)  

I am using MSSQL database connector for Node.js


This comment has been minimized.

Copy link

@amit-handa amit-handa commented Jul 14, 2015

facing issue in the script.
scenario: restarting workers
For 2 workers (say), it stops first worker. cluster.on('disconnect', ...) is called immediately (before worker.kill() is called through a timer). It is a bug.
Ideally, that function should be scheduled on cluster.on( 'exit', ...)
Hope I am correct. Doing this makes your script fly. would appreciate your inputs.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment