Netflix Automates Container Right-Sizing for Serverless Media Processing Platform

Stratum: Netflix's Media Processing Platform

Stratum is Netflix's internal serverless media processing platform, built on over a decade of experience with large-scale distributed systems. It handles various media workloads, including encoding, packaging, inspection, and quality scoring, supporting playback, advertising, studio workflows, and AI embeddings.

Engineers develop Stratum functions by packaging application code and OS dependencies into OCI images. These containers are then scheduled as batch jobs on Netflix’s Titus compute platform, with upstream services triggering function executions. Stratum manages hundreds of thousands of containers daily for bursty, high-throughput, and latency-sensitive workloads, scaling to accommodate new content formats like podcasts, live events, and regional ingests.

The Challenge of Manual Resource Management

Function owners specify CPU, memory, disk, and network limits for their containers using YAML configurations. Manually maintaining optimal resource values is challenging due to evolving Netflix infrastructure, EC2 instance upgrades, and application-level changes. This often results in configurations being over-provisioned initially or drifting from ideal values over time.

Mis-sized configurations are a common issue in large-scale serverless platforms, increasing cognitive load for function builders and operational costs for platform owners. Engineers typically copy existing resource YAMLs or size for worst-case scenarios, rarely revisiting configurations until a failure occurs, which often leads to further over-provisioning out of caution.

Introducing the Stratum Resource Tuner (SRT)

To address these challenges, Netflix developed the Stratum Resource Tuner (SRT), an automated system designed to safely right-size running containers in production without human intervention. This system aims to overcome the limitations of manual resource tuning, which include high operational friction due to the time-consuming process of metric monitoring and trial-and-error testing for downsizing.