Worker Resource Auto-Tuning

4 posts · Jan 2022 – Mar 2026

What this trail covers

Worker Auto-Tuning automatically adjusts worker slots based on CPU and memory usage, maximizing efficiency and preventing out-of-memory errors with no manual tuning required. Worker Scaling provides real-time data for making informed scaling decisions for worker fleets, optimizing the balance between workflow execution speed and fleet size/cost.

All Temporal trails

Mar 2026When running your own orchestrator isn’t worth it anymore — and why EvenUp chose TemporalThis post details how EvenUp migrated their existing workloads to Temporal. They specifically highlight their use of Temporal's metrics for scaling workers, employing both slot-based scaling using `temporal_worker_task_slots_used` and latency-based fallbacks with `temporal_workflow_task_schedule_to_start_latency` and `temporal_activity_schedule_to_start_latency`. Additionally, they describe the implementation of a Kubernetes liveness probe using metrics like `temporal_num_pollers`, `temporal_worker_task_slots_available`, and `temporal_request{operation="RespondWorkflowTaskCompleted"}` to automatically restart unhealthy workers, resolving issues with stuck workers and periodic processing slowdowns.
Apr 2025Announcing KEDA-based auto-scaling for Temporal WorkersIntroduces the integration of KEDA with Temporal Workers, allowing for event-driven autoscaling based on Task Queue backlog. This enables scaling workers up or down dynamically, including scaling to zero, to optimize resource utilization and cost efficiency in Kubernetes environments. A demo application showcases this integration with an Order Management System.
Nov 2024Resource-based auto-tuning for Temporal WorkersIntroduces resource-based auto-tuning for Temporal Workers, allowing them to scale up to the maximum bounds of CPU and Memory of the underlying compute node. This feature aims to simplify Worker management by automatically adjusting available Task Slots based on target CPU and memory usage (`targetMemoryUsage`, `targetCpuUsage`). It also details optional granular controls like `minimumSlots`, `maximumSlots`, and `rampThrottleMs`, and an alternative static `numSlots` option. The post clarifies that resource-based tuning cannot be used concurrently with `maxConcurrentWorkflowTaskExecutionSize` or `maxConcurrentActivityExecutionSize`.
Sep 2024Product Updates from Replay 2024This post announces the general availability of Worker Auto-Tuning, which automatically adjusts worker slots based on CPU and memory usage. It also highlights the Worker Scaling feature, which uses real-time data from task queue metrics to inform scaling decisions for worker fleets.
Oct 2023An introduction to Worker tuningThis post introduces fundamental concepts of Temporal Workers and Task Queues, explaining how Workers process Workflow and Activity Tasks. It details key metrics (`workflow_task_schedule_to_start_latency`, `worker_task_slots_available`, and poll success rate) and configuration options (`MaxConcurrentActivityExecutionSize`, `MaxConcurrentWorkflowTaskExecutionSize`) for tuning Worker performance and scaling. It also highlights the importance of having more than one worker for resilience and discusses the implications of having too many workers.
Advertisement
Jan 2022Productionizing Workers in TemporalThis post provides a comprehensive guide to productionizing Temporal Workers, covering four key areas: Operational Metrics (monitoring activity/workflow health, scaling based on latency, using Tally/Prometheus/Datadog), Incident Response (distinguishing application vs. server failures, using logs, execution history, and replay), Upgrading and Versioning (using `getVersion` API, feature flagging by cohort, task queues for short-lived workflows), and Testing Failure Paths (testing for backlogs by shutting down workers, capacity planning, and testing activity timeouts and retry policies). It references existing documentation and tutorials for deeper dives.