Skip to content

armgit / alook

Public Imported from GitHub
0

Multi-Agent Collaboration for Team and Friends

fix: resolve D1 overload by optimizing sweep queries

Gener870de2d6670b
diff --git a/src/shared/src/db/queries/task.ts b/src/shared/src/db/queries/task.ts
index e4b6ee05f..c6c457c79 100644
--- a/src/shared/src/db/queries/task.ts
+++ b/src/shared/src/db/queries/task.ts
@@ -616,28 +616,37 @@ const DEFAULT_STALE_RUNNING_SECONDS = Number(process.env.ALOOK_STALE_RUNNING_TIM
export async function failStaleRunningTasks(db: Database, workspaceId: string, staleSeconds = DEFAULT_STALE_RUNNING_SECONDS) {
const threshold = new Date(Date.now() - staleSeconds * 1000).toISOString();
- const lastMsg = db
- .select({
- taskId: taskMessage.taskId,
- lastMessageAt: sql<string>`max(${taskMessage.createdAt})`.as("last_message_at"),
- })
- .from(taskMessage)
- .groupBy(taskMessage.taskId)
- .as("last_msg");
-
- const staleTasks = await db
- .select({ id: agentTaskQueue.id })
+ const runningTasks = await db
+ .select({ id: agentTaskQueue.id, startedAt: agentTaskQueue.startedAt })
.from(agentTaskQueue)
- .leftJoin(lastMsg, eq(lastMsg.taskId, agentTaskQueue.id))
.where(
and(
eq(agentTaskQueue.workspaceId, workspaceId),
eq(agentTaskQueue.status, "running"),
- lt(sql`coalesce(${lastMsg.lastMessageAt}, ${agentTaskQueue.startedAt})`, threshold)
)
);
- if (staleTasks.length === 0) return [];
+ if (runningTasks.length === 0) return [];
+
+ const taskIds = runningTasks.map(t => t.id);
+ const lastMessages = await db
+ .select({
+ taskId: taskMessage.taskId,
+ lastMessageAt: sql<string>`max(${taskMessage.createdAt})`.as("last_message_at"),
+ })
+ .from(taskMessage)
+ .where(inArray(taskMessage.taskId, taskIds))
+ .groupBy(taskMessage.taskId);
+
+ const lastMsgMap = new Map(lastMessages.map(m => [m.taskId, m.lastMessageAt]));
+ const staleIds = runningTasks
+ .filter(t => {
+ const lastActivity = lastMsgMap.get(t.id) ?? t.startedAt;
+ return lastActivity != null && lastActivity < threshold;
+ })
+ .map(t => t.id);
+
+ if (staleIds.length === 0) return [];
const rows = await db
.update(agentTaskQueue)
@@ -646,7 +655,7 @@ export async function failStaleRunningTasks(db: Database, workspaceId: string, s
completedAt: new Date().toISOString(),
error: `timed out in running state (no message activity for ${Math.round(staleSeconds / 60)} minutes)`,
})
- .where(and(inArray(agentTaskQueue.id, staleTasks.map(t => t.id)), eq(agentTaskQueue.status, "running")))
+ .where(and(inArray(agentTaskQueue.id, staleIds), eq(agentTaskQueue.status, "running")))
.returning({ agentId: agentTaskQueue.agentId, workspaceId: agentTaskQueue.workspaceId, conversationId: agentTaskQueue.conversationId });
return rows;
}
diff --git a/src/web/src/lib/services/sweep.test.ts b/src/web/src/lib/services/sweep.test.ts
index 89ee78c2d..8974d7f03 100644
--- a/src/web/src/lib/services/sweep.test.ts
+++ b/src/web/src/lib/services/sweep.test.ts
@@ -50,13 +50,14 @@ vi.mock("@/lib/api/responses", () => ({
taskToResponse: (t: unknown) => t,
}));
-import { sweepStaleState } from "./sweep";
+import { sweepStaleState, _resetSweepThrottle } from "./sweep";
const db = {} as any;
describe("sweepStaleState", () => {
beforeEach(() => {
vi.clearAllMocks();
+ _resetSweepThrottle();
mockFailStaleKillTasks.mockResolvedValue([]);
mockFailStaleRunningTasks.mockResolvedValue([]);
mockActivateNextBufferedMessage.mockResolvedValue(null);
diff --git a/src/web/src/lib/services/sweep.ts b/src/web/src/lib/services/sweep.ts
index 8ac266d34..c7373323f 100644
--- a/src/web/src/lib/services/sweep.ts
+++ b/src/web/src/lib/services/sweep.ts
@@ -2,11 +2,25 @@ import type { Database } from "@alook/shared";
import { queries } from "@alook/shared";
import { TaskService } from "./task";
+const SWEEP_INTERVAL_MS = 30_000;
+const lastSweepAt = new Map<string, number>();
+
+/** @internal test-only */
+export function _resetSweepThrottle() {
+ lastSweepAt.clear();
+}
+
/**
* Unified workspace housekeeping. Any code path that wants to ensure
* stale state is cleaned up just calls this one function.
+ * Rate-limited to once per 30s per workspace within the same isolate.
*/
export async function sweepStaleState(db: Database, workspaceId: string) {
+ const now = Date.now();
+ const last = lastSweepAt.get(workspaceId) ?? 0;
+ if (now - last < SWEEP_INTERVAL_MS) return;
+ lastSweepAt.set(workspaceId, now);
+
// 1. Fail tasks stuck in "dispatched" for >20s (daemon crashed between claim and start)
const staleDispatched = await queries.task.failStaleDispatchedTasks(db, workspaceId);