Scaling Kubernetes workloads on custom metrics

The 2025 State of Containers and Serverless report highlights that while 64% of organizations use Kubernetes Horizontal Pod Autoscaler (HPA) to manage workload capacity, only 20% use custom metrics for scaling, relying instead on resource metrics like CPU and memory. However, these resource metrics often fail to accurately reflect real-time demand, leading to delayed scaling and potential performance degradation. The report suggests that custom metrics, such as queue depth, request rate, and tail latency, provide more effective scaling signals for specific workload patterns like event-driven workers, high-throughput APIs, latency-sensitive services, and database-backed services. Datadog Kubernetes Autoscaling (DKA) is presented as a solution that integrates metrics, observability, and autoscaling into one platform, allowing for more responsive scaling based on application metrics. DKA enables teams to define scaling signals using Datadog's query language, providing tools to tune scaling behavior and safeguard against custom metric delays by reverting to local CPU-based scaling when necessary. The article emphasizes the importance of aligning scaling strategies with business logic, encouraging the use of custom metrics to avoid overprovisioning and ensure high performance.