@harperfast/harper 5.3.0-alpha.1 → 5.3.0-beta.2

package/server/threads/manageThreads.js

@@ -64,10 +64,12 @@ let threadTerminationTimeout = process.env.DEV_MODE === 'true' || process.env.DE
 const RESTART_TYPE = 'restart';
 const RESTART_PROGRESS_HEARTBEAT_MS = 15000;
 const REQUEST_THREAD_INFO = 'request_thread_info';
+const REQUEST_RUNNING_ISOLATED_APPLICATIONS = 'request-running-isolated-applications';
 const RESOURCE_REPORT = 'resource_report';
 const OS_THREAD_ID = 'os-thread-id';
 const STUCK_WORKER_REPORT = 'stuck-worker-report';
 const THREAD_INFO = 'thread_info';
+const RUNNING_ISOLATED_APPLICATIONS = 'running-isolated-applications';
 const ADDED_PORT = 'added-port';
 const ACKNOWLEDGEMENT = 'ack';
 const REMOVE_PORT = 'remove-port';
@@ -85,6 +87,7 @@ const AWAIT_PROCESS_GROUP_TERMINATION = 'await-process-group-termination';
 const PROCESS_GROUP_TERMINATION_CONFIRMED = 'process-group-termination-confirmed';
 const THREAD_INFO_REQUEST_TIMEOUT_MS = 1000;
 let getThreadInfo;
+let getRunningIsolatedApplications;
 let awaitProcessGroupTermination;
 // Worker-side backstop that force-exits if the graceful shutdown sequence doesn't finish in time.
 let selfExitTimer;
@@ -151,6 +154,19 @@ module.exports = {
 	workers,
 	setMonitorListener,
 	onMessageFromWorkers,
+	setIsolatedWorkerReconciler,
+	setRunningIsolatedApplicationsGetter,
+	isApplicationPrimaryWorker,
+	applicationWorkerIndex,
+	runsApplicationCodeSingletons,
+	isDedicatedWorker,
+	markBranchStorePath,
+	ownsStoreMaintenance,
+	ownsStoreExpiration,
+	workersForApplication,
+	stopWorker,
+	encodeRestartScope,
+	decodeRestartScope,
 	onMessageByType,
 	broadcast,
 	broadcastWithAcknowledgement,
@@ -173,6 +189,11 @@ module.exports = {
 	isThreadRunning,
 	waitUntilConfirmedGone,
 	restartNumber: workerData?.restartNumber || 1,
+	// Identifies this process incarnation, where the PID cannot: a container reuses PID 1. Minted once
+	// on the main thread and carried to workers, so live siblings agree on it — one derived per thread
+	// would not. `undefined` on a worker started without it; consumers must fall back, not treat that
+	// as a mismatch.
+	processIncarnation: workerData ? workerData.processIncarnation : randomBytes(8).toString('hex'),
 };
 
 connectedPorts.onMessageByType = onMessageByType;
@@ -191,6 +212,9 @@ connectedPorts.sendToThread = function (threadId, message) {
 		throw err;
 	}
 };
+// Direct thread-to-thread send, so a worker can reach a sibling (e.g. the record lock owner worker)
+// without a hop through main. Returns false when no port for the thread is connected.
+module.exports.sendToThread = connectedPorts.sendToThread;
 module.exports.whenThreadsStarted = new Promise((resolve) => {
 	module.exports.threadsHaveStarted = resolve;
 });
@@ -206,6 +230,101 @@ function setTerminateTimeout(newTimeout) {
 function getWorkerIndex() {
 	return workerData ? workerData.workerIndex : isMainWorker ? 0 : undefined;
 }
+/**
+ * The worker that owns `applicationName`'s per-application singletons (its scheduled jobs, its data
+ * files): pool worker 0 for a shared application and for root-level plugins (which pass no name), the
+ * dedicated worker for an isolated one -- never index 0, so node-wide duties stay on the pool, and
+ * never a dedicated worker for anything but its own application.
+ */
+function isApplicationPrimaryWorker(applicationName) {
+	if (workerData?.isolatedApplication !== undefined) return applicationName === workerData.isolatedApplication;
+	return getWorkerIndex() === 0;
+}
+/**
+ * Stores that exist only on this thread: the branch stores an isolated application opened in its
+ * dedicated worker. Registered by openBranchDatabase, so the ownership predicates below can tell them
+ * from the shared stores every thread has open.
+ */
+const branchStorePaths = new Set();
+function markBranchStorePath(path, isBranch = true) {
+	if (isBranch) branchStorePaths.add(path);
+	else branchStorePaths.delete(path);
+}
+/**
+ * Whether this thread runs the "last worker" per-store maintenance for `storePath` (TTL scans, storage
+ * reclamation, audit cleanup). Shared stores belong to the last pool worker; a dedicated worker
+ * maintains only its own branch stores, which no other thread has open.
+ */
+function ownsStoreMaintenance(storePath) {
+	if (workerData?.isolatedApplication !== undefined) return branchStorePaths.has(storePath);
+	return getWorkerIndex() === getWorkerCount() - 1;
+}
+/** The "worker 0" counterpart: expiration eviction for `storePath`. */
+function ownsStoreExpiration(storePath) {
+	if (workerData?.isolatedApplication !== undefined) return branchStorePaths.has(storePath);
+	return getWorkerIndex() === 0;
+}
+/**
+ * Whether a singleton set up by APPLICATION CODE runs here (a caching table's `sourcedFrom`
+ * subscription): pool worker 0, or a dedicated worker -- the only code that runs on one is its own
+ * application's, so whatever it sets up is that application's to run. Not for root-level plugins,
+ * which load on every thread and must use isApplicationPrimaryWorker(name).
+ */
+function runsApplicationCodeSingletons() {
+	return getWorkerIndex() === 0 || workerData?.isolatedApplication !== undefined;
+}
+/** Whether this thread is a worker dedicated to one isolated application. */
+function isDedicatedWorker() {
+	return workerData?.isolatedApplication !== undefined;
+}
+/** The worker index as the application sees it: a dedicated worker is its application's worker 0. */
+function applicationWorkerIndex() {
+	return workerData?.isolatedApplication !== undefined ? 0 : getWorkerIndex();
+}
+/** Every started worker dedicated to `application`, including a replacement still booting. */
+function workersForApplication(application) {
+	return workers.filter((worker) => worker.application === application);
+}
+/**
+ * Stop one worker for good: shut it down, honour the drain extension it may ask for, force it after
+ * the same backstop the rolling restart uses (FORCE_EXIT on Bun, where terminate() segfaults), and
+ * resolve once it has exited. Marked as shut down first so its exit does not start a replacement.
+ */
+function stopWorker(worker) {
+	worker.wasShutdown = true;
+	return new Promise((resolve) => {
+		const armTerminate = (delay) =>
+			setTimeout(() => {
+				harperLogger.warn('Thread did not voluntarily terminate, terminating from the outside', worker.threadId);
+				if (isBun) {
+					try {
+						worker.postMessage({ type: FORCE_EXIT });
+					} catch {}
+				} else {
+					worker.terminate();
+				}
+			}, delay).unref();
+		let timeout = armTerminate(threadTerminationTimeout * 2);
+		worker.extendTerminateDeadline = (deadlineMs) => {
+			clearTimeout(timeout);
+			const { boundedTerminateDelay, getShutdownDrainCeilingMs } = require('../../components/shutdownDrain.ts');
+			timeout = armTerminate(
+				boundedTerminateDelay(deadlineMs, Date.now(), threadTerminationTimeout * 2, getShutdownDrainCeilingMs())
+			);
+		};
+		worker.on('exit', () => {
+			clearTimeout(timeout);
+			worker.extendTerminateDeadline = undefined;
+			resolve();
+		});
+		try {
+			worker.postMessage({ restartNumber: module.exports.restartNumber, type: hdbTerms.ITC_EVENT_TYPES.SHUTDOWN });
+		} catch {
+			clearTimeout(timeout);
+			resolve(); // already gone
+		}
+	});
+}
 function getWorkerCount() {
 	return workerData ? workerData.workerCount : isMainWorker ? 1 : undefined;
 }
@@ -239,8 +358,10 @@ const RESERVED_WORKER_DATA_KEYS = [
 	'workerCount',
 	'name',
 	'restartNumber',
+	'processIncarnation',
 	'ticketKeys',
 	'noServerStart',
+	'isolatedApplication',
 	'__proto__', // never a legitimate payload name; spread would define it as an own property
 ];
 const workerDataProviders = new Map();
@@ -294,20 +415,31 @@ function getTicketKeys() {
 	ticketKeys = isMainThread ? randomBytes(48) : workerData.ticketKeys;
 	return ticketKeys;
 }
+// What application code sees: a dedicated worker is its application's only worker, index 0 of 1, so
+// the documented partitioning idiom (`hash % server.workerCount === server.workerIndex`) covers
+// everything there. Node-wide duties keep using the raw getWorkerIndex/getWorkerCount.
 Object.defineProperty(server, 'workerIndex', {
 	get() {
-		return getWorkerIndex();
+		return applicationWorkerIndex();
 	},
 });
 Object.defineProperty(server, 'workerCount', {
 	get() {
-		return getWorkerCount();
+		return workerData?.isolatedApplication !== undefined ? 1 : getWorkerCount();
 	},
 });
 if (!parentPort) {
 	onMessageByType(REQUEST_THREAD_INFO, (message, worker) => {
 		if (worker) sendThreadInfo(worker);
 	});
+	onMessageByType(REQUEST_RUNNING_ISOLATED_APPLICATIONS, (message, worker) => {
+		if (worker)
+			worker.postMessage({
+				type: RUNNING_ISOLATED_APPLICATIONS,
+				requestId: message.requestId,
+				applications: runningIsolatedApplicationsGetter(),
+			});
+	});
 	onMessageByType(RESOURCE_REPORT, (message, worker) => {
 		if (worker) recordResourceReport(worker, message);
 	});
@@ -342,11 +474,12 @@ listenersByType.set(hdbTerms.ITC_EVENT_TYPES.MIDDLEWARE_CHAINS_RESPONSE, null);
 listenersByType.set(hdbTerms.ITC_EVENT_TYPES.OPERATION_REGISTERED, null);
 listenersByType.set(hdbTerms.ITC_EVENT_TYPES.OPERATION_EXECUTE_REQUEST, null);
 listenersByType.set(hdbTerms.ITC_EVENT_TYPES.OPERATION_EXECUTE_RESPONSE, null);
-// getThreadInfo/awaitProcessGroupTermination register their own one-shot parentPort listener per
+// These request/response functions register their own one-shot parentPort listener per
 // call rather than going through onMessageByType, so without this, every reply would also reach
 // addPort's permanent dispatcher as an "unregistered" type: notifyMessageListeners would warn and
 // queue each one in messagesQueuedByType forever, and a lock-wait polls every 50ms.
 listenersByType.set(THREAD_INFO, null);
+listenersByType.set(RUNNING_ISOLATED_APPLICATIONS, null);
 listenersByType.set(PROCESS_GROUP_TERMINATION_CONFIRMED, null);
 
 function startWorker(path, options = {}) {
@@ -370,7 +503,7 @@ function startWorker(path, options = {}) {
 	availableMemory = Math.min(availableMemory, totalmem(), 20000 * MB);
 	const maxOldMemory =
 		resolveThreadHeapMemoryMb(envMgr.get(hdbTerms.CONFIG_PARAMS.THREADS_MAXHEAPMEMORY)) ??
-		Math.max(Math.floor(availableMemory / MB / (10 + (options.threadCount || 1) / 4)), 512);
+		Math.max(Math.floor(availableMemory / MB / (10 + (options.heapShareCount || options.threadCount || 1) / 4)), 512);
 	// Max young memory space (semi-space for scavenger) is 1/128 of max memory (limited to 16-64). For most of our m5
 	// machines this will be 64MB (less for t3's). This is based on recommendations from:
 	// https://www.alibabacloud.com/blog/node-js-application-troubleshooting-manual---comprehensive-gc-problems-and-optimization594965
@@ -416,6 +549,9 @@ function startWorker(path, options = {}) {
 		for (const requirePath of getRequireModules()) execArgv.push('--require', requirePath);
 	}
 
+	// Only a start that declares the serving topology may write it; see DESIGN.md on the two workerCounts.
+	if (typeof options.threadCount === 'number') workerCount = options.threadCount;
+
 	const worker = new Worker(isAbsolute(path) ? path : join(PACKAGE_ROOT, path), {
 		resourceLimits: {
 			maxOldGenerationSizeMb: maxOldMemory,
@@ -430,9 +566,11 @@ function startWorker(path, options = {}) {
 			addThreadIds: channelsToConnect.map((channel) => channel.existingPort.threadId),
 			addPortIsJobWorkers: channelsToConnect.map((channel) => channel.existingPort.isJobWorker === true),
 			workerIndex: options.workerIndex,
-			workerCount: (workerCount = options.threadCount),
+			workerCount: options.threadCount,
 			name: options.name,
+			isolatedApplication: options.application,
 			restartNumber: module.exports.restartNumber,
+			processIncarnation: module.exports.processIncarnation,
 			ticketKeys: getTicketKeys(),
 		},
 		transferList: portsToSend,
@@ -466,7 +604,12 @@ function startWorker(path, options = {}) {
 	});
 	worker.on('exit', (_code) => {
 		workers.splice(workers.indexOf(worker), 1);
-		if (!processShuttingDown && !worker.wasShutdown && options.autoRestart !== false) {
+		if (
+			!processShuttingDown &&
+			!worker.wasShutdown &&
+			options.autoRestart !== false &&
+			options.shouldAutoRestart?.(worker) !== false
+		) {
 			// if this wasn't an intentional shutdown, restart now (unless we have tried too many times)
 			if (worker.unexpectedRestarts < MAX_UNEXPECTED_RESTARTS) {
 				options.unexpectedRestarts = worker.unexpectedRestarts + 1;
@@ -481,6 +624,7 @@ function startWorker(path, options = {}) {
 	startMonitoring();
 	if (options.onStarted) options.onStarted(worker); // notify that it is ready
 	worker.name = options.name;
+	worker.application = options.application; // the isolated application this worker is dedicated to, if any
 	return worker;
 }
 
@@ -502,12 +646,19 @@ const OVERLAPPING_RESTART_TYPES = [hdbTerms.THREAD_TYPES.HTTP];
  * from a worker, nothing — the restart is handed to the main thread.
  */
 
+/**
+ * `application` selects which workers of the type restart: omitted preserves the historic all-worker
+ * behavior, explicit undefined restarts the shared pool only, a name restarts only that application's
+ * dedicated worker, and '*' restarts all.
+ */
 async function restartWorkers(
 	name = null,
 	maxWorkersDown = Math.max(Math.floor(workerCount / 8), 1), // restart 1/8 of the threads at a time, but at least 1
 	startReplacementThreads = true,
-	onProgress = null
+	onProgress = null,
+	application = undefined
 ) {
+	if (arguments.length < 5) application = '*';
 	if (isMainThread) {
 		// Declining is not the same as delegating: a caller reporting on the restart must not read this
 		// as "another thread is completing it".
@@ -520,9 +671,19 @@ async function restartWorkers(
 		} catch (e) {
 			harperLogger.error('Unable to reestablish current working directory', e);
 		}
+		const freshlyStarted = new Set(); // dedicated workers the reconcile just started: already on the new code
 		// problematic cyclic dependency, bind late
 		const { resetRestartNeeded } = require('../../components/requestRestart.ts');
-		resetRestartNeeded();
+		// One process-wide bit cannot say WHICH application is pending, so it may only be cleared by a
+		// restart that demonstrably covers every worker the bit could stand for. A restart scoped to one
+		// application leaves the whole pool on its old code; a pool restart replaces every pool worker and
+		// reconciles dedicated slots, but leaves an already-running dedicated worker on its old modules.
+		// Clearing from either would report restartRequired: false while a deployed component is still
+		// loaded nowhere. So: all workers always, the pool only while no dedicated worker is running --
+		// which is every node that uses no isolated application, i.e. the historic behavior unchanged.
+		const coversEveryWorker =
+			application === '*' || (application === undefined && !workers.some((worker) => worker.application));
+		if (coversEveryWorker) resetRestartNeeded();
 		// This is here to prevent circular dependencies
 		if (startReplacementThreads) {
 			const { loadRootComponents } = require('../loadRootComponents.js');
@@ -532,6 +693,13 @@ async function restartWorkers(
 			const loading = setInterval(() => onProgress?.(), RESTART_PROGRESS_HEARTBEAT_MS).unref();
 			try {
 				await loadRootComponents();
+				// isolated applications added or removed by the reload get their dedicated worker started or
+				// stopped; a crash-looping newcomer can take a while, so the heartbeat runs through this too
+				try {
+					for (const application of (await isolatedWorkerReconciler?.()) ?? []) freshlyStarted.add(application);
+				} catch (error) {
+					harperLogger.error('Could not reconcile isolated application workers', error);
+				}
 			} finally {
 				clearInterval(loading);
 			}
@@ -539,7 +707,10 @@ async function restartWorkers(
 		}
 
 		module.exports.restartNumber++;
-		if (maxWorkersDown < 1) {
+		// `Infinity` is shutdownWorkers' "all at once" sentinel and must survive.
+		if (typeof maxWorkersDown !== 'number' || Number.isNaN(maxWorkersDown)) {
+			maxWorkersDown = 1;
+		} else if (maxWorkersDown < 1) {
 			// we accept a ratio of workers, and compute absolute maximum being down at a time from the total number of
 			// threads
 			maxWorkersDown = maxWorkersDown * workers.length;
@@ -566,14 +737,17 @@ async function restartWorkers(
 		// replacement — an unavoidable brief gap, but only for worker-owned listeners, since the main
 		// thread keeps serving the HTTP ports throughout. This ordering is also what lets
 		// listenOnPorts() treat a dedicated listener's EADDRINUSE as an external conflict.
-		const canPreStartReplacement = process.platform !== 'win32' && process.platform !== 'darwin' && !isBun;
+		const platformCanPreStartReplacement = process.platform !== 'win32' && process.platform !== 'darwin' && !isBun;
 		const restarting = workers.slice(0);
 		for (let index = 0; index < restarting.length; index++) {
 			const worker = restarting[index];
 			// Terminal shutdown: stop replacing workers mid-loop — the guard for every replacement start below.
 			if (processShuttingDown && startReplacementThreads) break;
 			if ((name && worker.name !== name) || worker.wasShutdown) continue; // filter by type, if specified
+			if (application !== '*' && worker.application !== application) continue; // and by isolated application
+			if (worker.application && freshlyStarted.has(worker.application)) continue;
 			const overlapping = OVERLAPPING_RESTART_TYPES.indexOf(worker.name) > -1;
+			const canPreStartReplacement = platformCanPreStartReplacement && !worker.application;
 			if (overlapping && startReplacementThreads && canPreStartReplacement) {
 				// Overlapping restart: start the replacement and wait until it is accepting connections
 				// *before* shutting down the worker it replaces. The replacement joins the (SO_REUSEPORT)
@@ -754,6 +928,7 @@ async function restartWorkers(
 		parentPort.postMessage({
 			type: RESTART_TYPE,
 			workerType: name,
+			scope: encodeRestartScope(application),
 		});
 	}
 }
@@ -803,7 +978,7 @@ function whenWorkerStarted(newWorker) {
 	});
 }
 function shutdownWorkers(name) {
-	return restartWorkers(name, Infinity, false);
+	return restartWorkers(name, Infinity, false, null, '*');
 }
 function beginProcessShutdown() {
 	processShuttingDown = true;
@@ -823,6 +998,28 @@ async function shutdownWorkersNow(name) {
 	}
 }
 
+/**
+ * The restart scope on the wire: a worker cannot post `undefined` and have it mean "the pool" (a
+ * message with no scope at all means "everything", as every pre-existing sender intends).
+ */
+function encodeRestartScope(application) {
+	return application === undefined ? '' : application; // '' is not a legal application name
+}
+function decodeRestartScope(message) {
+	if (message.scope === undefined) return '*';
+	return message.scope === '' ? undefined : message.scope;
+}
+
+let isolatedWorkerReconciler = null;
+let runningIsolatedApplicationsGetter = () => [];
+/** Registered by socketRouter: starts and stops dedicated workers to match the root config's isolated applications. */
+function setIsolatedWorkerReconciler(reconcile) {
+	isolatedWorkerReconciler = reconcile;
+}
+function setRunningIsolatedApplicationsGetter(getter) {
+	runningIsolatedApplicationsGetter = getter;
+}
+
 const messageListeners = [];
 function onMessageFromWorkers(listener) {
 	messageListeners.push(listener);
@@ -1089,6 +1286,7 @@ function getChildWorkerInfo() {
 	return workers.map((worker) => ({
 		threadId: worker.threadId,
 		name: worker.name,
+		application: worker.application,
 		heapTotal: worker.resources?.heapTotal,
 		heapUsed: worker.resources?.heapUsed,
 		externalMemory: worker.resources?.external,
@@ -1198,6 +1396,38 @@ if (parentPort && workerData?.addPorts) {
 				}, timeoutMs);
 			}
 		});
+	let nextRunningIsolatedApplicationsRequestId = 0;
+	getRunningIsolatedApplications = (timeoutMs) =>
+		new Promise((resolve, reject) => {
+			const requestId = ++nextRunningIsolatedApplicationsRequestId;
+			let timeout;
+			parentPort.on('message', receiveApplications);
+			try {
+				parentPort.postMessage({ type: REQUEST_RUNNING_ISOLATED_APPLICATIONS, requestId });
+			} catch (error) {
+				cleanup();
+				reject(error);
+				return;
+			}
+			function receiveApplications(message) {
+				if (message.type === RUNNING_ISOLATED_APPLICATIONS && message.requestId === requestId) {
+					cleanup();
+					resolve(message.applications);
+				}
+			}
+			function cleanup() {
+				if (timeout) clearTimeout(timeout);
+				parentPort.off('message', receiveApplications);
+			}
+			if (timeoutMs != null) {
+				timeout = setTimeout(() => {
+					cleanup();
+					const error = new Error(`Timed out waiting for isolated application topology after ${timeoutMs}ms`);
+					error.code = 'ERR_ISOLATED_APPLICATIONS_TIMEOUT';
+					reject(error);
+				}, timeoutMs);
+			}
+		});
 	let awaitTerminationRequestId = 0;
 	awaitProcessGroupTermination = (ownerThreadId, signal) =>
 		new Promise((resolve) => {
@@ -1222,10 +1452,18 @@ if (parentPort && workerData?.addPorts) {
 		});
 } else {
 	getThreadInfo = getChildWorkerInfo;
+	getRunningIsolatedApplications = isMainThread
+		? () => runningIsolatedApplicationsGetter()
+		: () => {
+				const error = new Error('No channel to the main thread for isolated application topology');
+				error.code = 'ERR_ISOLATED_APPLICATIONS_UNAVAILABLE';
+				return Promise.reject(error);
+			};
 	awaitProcessGroupTermination = (ownerThreadId) =>
 		pendingProcessGroupTerminations.get(ownerThreadId) ?? Promise.resolve();
 }
 module.exports.getThreadInfo = getThreadInfo;
+module.exports.getRunningIsolatedApplications = getRunningIsolatedApplications;
 
 // Listeners notified when a connected thread's port closes (worker exit/restart), so
 // modules holding per-thread state (e.g. registeredOperations' registry and in-flight
@@ -1720,7 +1958,7 @@ if (isMainThread) {
 					if (queuedRestart) clearTimeout(queuedRestart);
 					queuedRestart = setTimeout(async () => {
 						if (beforeRestart) await beforeRestart();
-						await restartWorkers();
+						await restartWorkers(undefined, undefined, true, null, '*');
 						console.log('Reloaded Harper components, changed files:', Array.from(changedFiles));
 						changedFiles.clear();
 					}, 100);
@@ -1742,3 +1980,7 @@ if (isMainThread) {
 		realExit(0);
 	});
 }
+
+// Required here, not from logging, which must never import this module; last in the file because it
+// calls onThreadExit.
+require('./logRotationTransport.ts');