Clustering library with support for zero-downtime reloading
If server.js is your regular http server (e.g. express), create cluster.js and add:
var recluster = require('recluster'),
path = require('path');
var cluster = recluster(path.join(__dirname, 'server.js'));
cluster.run();
process.on('SIGUSR2', function() {
console.log('Got SIGUSR2, reloading cluster...');
cluster.reload();
});
console.log("spawned cluster, kill -s SIGUSR2", process.pid, "to reload");
then run it
node cluster.js
To hot-reload the server, simply run
kill -s SIGUSR2 <cluster_pid>
To find out which of the N (= number of cores by default) worker instances you're running from inside server.js, you can use
process.env.WORKER_ID
which is zero-based i.e. 0 <= WORKER_ID < N
var cluster = recluster(file, opt)
where
Absolute path to the module that defines the server
Number of active workers (default = cores)
Timeout to kill old workers after reload (seconds).
Defaults to 1 second in development, 1 hour in production.
Minimum time between worker respawns when workers die (seconds)
Maximum respawn time (reached via exponential backoff). Set to 0 or undefined to disable exponential backoff.
Use 'listening'
for servers (e.g. for express/connect http servers)
and 'started'
for workers that are immediately ready.
If you want to manually tell recluster when the worker is ready to replace
older workers you can use {readyWhen: 'ready'}
. Then, to signal readiness
from the worker use process.send({cmd: 'ready'})
Array of arguments to pass to the worker
Log various events to stdout. Currently only 'respawns' is supported.
Default: {respawns: true}
Which logger to use. Requires a console-compatible log method
Default: console
The returned object has the following methods:
Starts the cluster by running child processes
Hot-reloads new code. some of the children will remain active
for opt.timeout
seconds after reload
Terminates the entire cluster and removes all listeners.
Returns a hash of all worker slots (0 <= WORKER_ID < N). If a worker isn't available at that slot, the value in the hash is null or undefined. Otherwise, the value will be a worker object that is ready to serve requests.
Returns an array of all the workers, including those that are not yet ready or those that will be replaced.
A server worker can gracefully exit by cleaning up in the 'close' event of its server:
server.on('close', function() {
// cleanup
});
Non-server workers can listen for the disconnect command and shut down gracefully before the kill timeout:
process.on('message', function(m) {
if (m.cmd == 'disconnect') {
// cleanup
}
})
If you need sticky sessions e.g. for socket.io you can use the experimental companion module sticky-listen, which implements an alternate balancer that distributes the sockets based on the client IP (instead of the regular round-robin one)