FOSS Talks 3.0

Temporal workers are deceptively hard to scale on Kubernetes. They long-poll task queues, have wildly variable resource profiles depending on workflow type, and naive autoscaling strategies lead to either over-provisioned clusters or thrashing workers that drop tasks under load.

In this talk, I’ll share how we built a production autoscaling stack for Temporal workers using three complementary tools: KEDA for event-driven horizontal scaling driven by task queue backlog metrics, VPA for continuous right-sizing of worker pod resources, and Temporal’s Slot Supplier API to give each worker fine-grained control over its own concurrency ceiling.