OpenClaw Signals: Scheduler Split, Resource Quotas, and the Lock Debate

Architecture

Proposal: Decouple Scheduler Into Queue, Dispatch, and Execution Layers

A new proposal splits the monolithic task scheduler into separate queue, dispatch, and execution layers with clean interfaces between each. This is a big architectural move that enables better testing, horizontal scaling of individual components, and clearer error boundaries. If it lands, it changes how you think about scaling OpenClaw workers independently.

Bug Fix

Fix: Metrics Collector Race Condition in Health Check Path

A race condition in the metrics collector was causing occasional stale data to be served to health check endpoints, which means load balancers could route traffic to unhealthy nodes. The fix adds proper locking around the metrics snapshot window. Production operators running OpenClaw behind any LB should pull this patch.

Proposal

Proposal: Resource Quotas and Per-Queue Resource Budgets

The discussion proposes defining explicit resource budgets per task queue so one misbehaving queue can't starve others of CPU or memory. This is table stakes for multi-tenant production use but OpenClaw lacks it today. Worth tracking if you're running mixed workloads on shared clusters.

Feature

Add Configurable Task TTL and Dead-Letter Routing

A new config option lets you set a max task age in the queue, after which stale tasks get dead-lettered instead of silently consuming worker capacity. Simple but effective for preventing queue bloat from stuck workflows. Quick win for anyone hitting mysterious queue growth in production.

Discussion

RFC: Should OpenClaw Bundle a Distributed Lock Primitive?

The community is debating whether OpenClaw should ship a built-in distributed lock service or keep relying on external providers like Redis or etcd. The thread surfaces real tradeoffs around complexity, dependency weight, and operational burden. If you've hit coordination issues across workers, this is the thread to watch.