Fix rare ERR_IPC_CHANNEL_CLOSED bug in cluster mode (#244)

* Fix rare ERR_IPC_CHANNEL_CLOSED bug in cluster mode

A cluster worker can still be in the list of workers even if its IPC channel has closed (e.g. when the worker exits abruptly). Check that its IPC channel is open before sending a message.

* Remove a Node.js pre-v6 compatibility bit

CHANGELOG.md

@@ -11,6 +11,9 @@ project adheres to [Semantic Versioning](http://semver.org/).
 
 ### Changed
 
+- Check that cluster worker is still connected before attempting to query it for
+  metrics. (#244)
+
 ### Added
 
 - Added a `remove()` method on each metric type, based on [Prometheus "Writing Client Libraries" section on labels](https://prometheus.io/docs/instrumenting/writing_clientlibs/#labels)

lib/cluster.js

@@ -38,8 +38,6 @@ class AggregatorRegistry extends Registry {
 		const requestId = requestCtr++;
 
 		return new Promise((resolve, reject) => {
-			const nWorkers = Object.keys(cluster.workers).length;
-
 			function done(err, result) {
 				// Don't resolve/reject the promise if a callback is provided
 				if (typeof callback === 'function') {
@@ -50,13 +48,9 @@ class AggregatorRegistry extends Registry {
 				}
 			}
 
-			if (nWorkers === 0) {
-				return process.nextTick(() => done(null, ''));
-			}
-
 			const request = {
 				responses: [],
-				pending: nWorkers,
+				pending: 0,
 				done,
 				errorTimeout: setTimeout(() => {
 					request.failed = true;
@@ -71,7 +65,21 @@ class AggregatorRegistry extends Registry {
 				type: GET_METRICS_REQ,
 				requestId
 			};
-			for (const id in cluster.workers) cluster.workers[id].send(message);
+
+			for (const id in cluster.workers) {
+				// If the worker exits abruptly, it may still be in the workers
+				// list but not able to communicate.
+				if (cluster.workers[id].isConnected()) {
+					cluster.workers[id].send(message);
+					request.pending++;
+				}
+			}
+
+			if (request.pending === 0) {
+				// No workers were up
+				clearTimeout(request.errorTimeout);
+				process.nextTick(() => done(null, ''));
+			}
 		});
 	}
 
@@ -147,13 +155,7 @@ function addListeners() {
 
 	if (cluster.isMaster) {
 		// Listen for worker responses to requests for local metrics
-		cluster.on('message', function(worker, message) {
-			if (arguments.length === 2) {
-				// pre-Node.js v6.0
-				message = worker;
-				worker = undefined;
-			}
-
+		cluster.on('message', (worker, message) => {
 			if (message.type === GET_METRICS_RES) {
 				const request = requests.get(message.requestId);
 				message.metrics.forEach(registry => request.responses.push(registry));