feat(concurrency): bullmq based concurrency control system

[icecrasher321]

Summary

• BullMQ based concurrency control system for executions currently running in line [manual execs excluded]. Can tune limits based on resources.

• Overall admin gates to prevent rate limiting services based crashes.

Type of Change

• New feature
• Other: Reliability

Testing

Tested manually under different configurations, added extensive test suite

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
docs	Skipped		Mar 26, 2026 9:29pm

[cursor]

PR Summary

High Risk
High risk because it rewires core workflow/webhook/schedule execution queuing and status reporting to a new BullMQ + workspace-dispatch pipeline with new capacity limits and fallback behavior; mistakes could drop/strand jobs or incorrectly reject traffic.

Overview
Introduces a BullMQ-backed workspace dispatch layer that queues execution jobs per workspace/lane, enforces global/workspace queue caps, leases concurrency slots, and reconciles BullMQ state back into dispatch records (with in-memory or Redis-backed dispatch storage).

Updates workflow execution, scheduled execution, and webhook trigger APIs to enqueue through enqueueWorkspaceDispatch when BullMQ is enabled, adds explicit capacity responses (429 admission gate, 503 dispatch-queue-full with Retry-After), and changes /api/jobs/[jobId] to be dispatcher-aware via presentDispatchOrJobStatus.

Adds billing-driven workspace concurrency limits (plan defaults + enterprise metadata override) with caching, extends enterprise billing metadata parsing with zod (including concurrency limit), and documents/markets the new concurrency limits in docs and pricing UI.

Replaces the async-jobs Redis backend with a BullMQ backend and adjusts inline-execution semantics so only the database fallback executes inline.

^{Written by Cursor Bugbot for commit 53733e4. This will update automatically on new commits. Configure here.}

[greptile-apps]

Greptile Summary

This PR introduces a comprehensive BullMQ-based queuing and concurrency control system for workflow, webhook, and schedule executions. It replaces the previous Redis/database job queue backends with BullMQ queues backed by Redis, adds a per-workspace fairness dispatcher (with a Lua-script-based atomic claim mechanism), a lease-based concurrency limit (keyed to billing plan), an in-process admission gate for external API requests, and a standalone worker process to consume jobs. It also adds a buffered SSE stream for non-manual executions that now run asynchronously through the dispatch queue. The change is substantial (~5900 lines added) and represents a foundational infrastructure improvement for reliability and rate-limiting protection.

Key changes and issues found:

• Misleading variable name in schedule route (apps/sim/app/api/schedules/execute/route.ts, lines 118–123): workspaceId is assigned the getWorkflowById function, not a workspace ID string. The code works at runtime but is extremely confusing and should be renamed (e.g. getWorkflowByIdFn).

• GLOBAL_DEPTH_KEY double-decrement race (apps/sim/lib/core/workspace-dispatch/redis-store.ts): markDispatchJobCompleted / markDispatchJobFailed unconditionally call DECR on GLOBAL_DEPTH_KEY. If the reconciler and the BullMQ worker both process the same job's completion concurrently, the counter is decremented twice. The Math.max(0, …) guard and periodic reconcileGlobalQueueDepth mitigate the visible impact, but a guard that skips the decrement when the record is already terminal would be more correct.

• Non-atomic BullMQ add + dispatch record update in finalizeAdmittedJob (apps/sim/lib/core/workspace-dispatch/planner.ts): The BullMQ job is added before the dispatch record is marked admitted. If markDispatchJobAdmitted fails, an orphaned BullMQ job exists with a released lease, creating a window where the job could be executed while the reconciler simultaneously restores the dispatch record to waiting. Reversing the order (mark admitted first, then add to BullMQ) makes the failure mode fully safe for the reconciler to recover.

• Long-polling HTTP handler (apps/sim/app/api/workflows/[id]/execute/route.ts): waitForDispatchJob can block the HTTP handler for up to 5.5 minutes. This is safe on dedicated servers but will be silently dropped by many load balancers/proxies with shorter idle timeouts. Infrastructure timeout requirements should be documented.

• Per-process admission gate (apps/sim/lib/core/admission/gate.ts): The inflight counter is module-local. In a horizontally scaled deployment the effective global limit is n_pods × ADMISSION_GATE_MAX_INFLIGHT. Operators should be guided to set the env-var to global_limit / n_pods.

Confidence Score: 3/5

• PR introduces valuable infrastructure but has a few logic issues around counter correctness, partial-failure atomicity, and infrastructure timeout assumptions that should be resolved before production rollout.
• The core design (Lua-script atomic claim, lease-based concurrency, BullMQ workers) is well-thought-out and the test suite is comprehensive. However, the non-atomic finalizeAdmittedJob step creates a real (if narrow) window for duplicate execution, the GLOBAL_DEPTH_KEY double-decrement weakens the global gate, and the long-polling HTTP handler will silently fail for users behind standard load balancers. These issues are fixable but should be addressed before heavy production traffic.
• Pay close attention to apps/sim/lib/core/workspace-dispatch/planner.ts (operation ordering), apps/sim/lib/core/workspace-dispatch/redis-store.ts (double-decrement guard), and apps/sim/app/api/workflows/[id]/execute/route.ts (long-polling timeout).

Important Files Changed

Filename	Overview
apps/sim/lib/core/workspace-dispatch/dispatcher.ts	New per-process dispatcher loop that coalesces wakeups and drives workspace admission; process-local state is intentional but should be noted for multi-pod deployments.
apps/sim/lib/core/workspace-dispatch/redis-store.ts	Redis-backed dispatch store using a Lua script for atomic job claiming; GLOBAL_DEPTH_KEY can be double-decremented in a race between the reconciler and worker completing the same job.
apps/sim/lib/core/workspace-dispatch/planner.ts	Claims next admissible workspace job via a Redis lock; BullMQ add and markDispatchJobAdmitted are not atomic — a partial failure leaves an orphaned BullMQ job that may be double-executed.
apps/sim/lib/core/workspace-dispatch/reconciler.ts	Periodic reconciliation of stranded/stuck dispatch jobs; correctly scoped to admitting/admitted/running states, but can race with the worker on markDispatchJobCompleted.
apps/sim/lib/core/workspace-dispatch/worker.ts	Lease heartbeat + job lifecycle hooks for BullMQ workers; solid implementation with proper cleanup in finally block.
apps/sim/lib/core/admission/gate.ts	Simple in-process admission gate limiting concurrent external API requests; effective global limit equals n_pods × MAX_INFLIGHT which should be documented for operators.
apps/sim/app/api/workflows/[id]/execute/route.ts	Refactored to route non-manual executions through workspace dispatch; long-polling waitForDispatchJob (up to 5.5 min) in the HTTP handler risks infrastructure timeout disconnects.
apps/sim/app/api/schedules/execute/route.ts	Updated to route scheduled and mothership jobs through workspace dispatch; contains a misleading variable named workspaceId that holds a function reference (getWorkflowById).
apps/sim/lib/billing/workspace-concurrency.ts	New module resolving per-workspace concurrency limits from billing plan, with Redis-backed and in-memory cache (60s TTL); clean implementation.
apps/sim/worker/index.ts	New standalone BullMQ worker process with configurable concurrency per queue, graceful shutdown, and periodic dispatcher wake/notification sweep intervals.

Sequence Diagram

sequenceDiagram
    participant Client
    participant RouteHandler as API Route Handler
    participant AdmissionGate as Admission Gate (in-process)
    participant Dispatcher as Workspace Dispatcher
    participant RedisStore as Redis Dispatch Store
    participant BullMQ as BullMQ Queue (Redis)
    participant Worker as BullMQ Worker Process
    participant DispatchWorker as Dispatch Worker (worker.ts)

    Client->>RouteHandler: POST /api/workflows/[id]/execute
    RouteHandler->>AdmissionGate: tryAdmit()
    alt At capacity (inflight >= MAX_INFLIGHT)
        AdmissionGate-->>RouteHandler: null
        RouteHandler-->>Client: 429 Too Many Requests
    else Admitted
        AdmissionGate-->>RouteHandler: ticket
        RouteHandler->>Dispatcher: enqueueWorkspaceDispatch(input)
        Dispatcher->>RedisStore: enqueueWorkspaceDispatchJob()
        RedisStore-->>Dispatcher: jobRecord (status=waiting)
        Dispatcher->>Dispatcher: runDispatcherLoop() [void]
        Dispatcher->>RedisStore: popNextWorkspaceId()
        Dispatcher->>RedisStore: claimWorkspaceJob() [Lua script]
        RedisStore-->>Dispatcher: {type: admitted, record, leaseId}
        Dispatcher->>BullMQ: queue.add(jobName, payload, {jobId})
        Dispatcher->>RedisStore: markDispatchJobAdmitted()
        Dispatcher-->>RouteHandler: dispatchJobId
        RouteHandler->>RouteHandler: waitForDispatchJob(id, timeout) [polls 250ms]
        Worker->>BullMQ: picks up job
        Worker->>DispatchWorker: runDispatchedJob(metadata, fn)
        DispatchWorker->>RedisStore: markDispatchJobRunning()
        DispatchWorker->>DispatchWorker: executeQueuedWorkflowJob() / executeWorkflowJob()
        DispatchWorker->>RedisStore: markDispatchJobCompleted(output)
        DispatchWorker->>RedisStore: releaseWorkspaceLease()
        DispatchWorker->>Dispatcher: wakeWorkspaceDispatcher()
        RedisStore-->>RouteHandler: record (status=completed) [via poll]
        RouteHandler->>AdmissionGate: ticket.release()
        RouteHandler-->>Client: 200 JSON result
    end

Loading

_{Last reviewed commit: be83c97}

[icecrasher321]

bugbot run

[icecrasher321]

bugbot run

[icecrasher321]

bugbot run

[icecrasher321]

bugbot run

[icecrasher321]

bugbot run

[icecrasher321]

bugbot run

[icecrasher321]

bugbot run

[cursor]

Cursor Bugbot has reviewed your changes and found 3 potential issues.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}