Netflix Deploys Stratum Resource Tuner for Automated Container Right-Sizing

Netflix has introduced the Stratum Resource Tuner (SRT), an automated system designed to right-size containers within its Stratum serverless media processing platform. The platform manages hundreds of thousands of containers daily for encoding, packaging, and AI embeddings, where manual resource configuration previously led to significant over-provisioning and operational friction.

Tuning Strategy and Safety Mechanisms

SRT targets high-throughput workloads that account for 70-80% of resource usage, applying distinct strategies for different resource types:

• CPU and Network: Tuned aggressively to the p99 usage level with no headroom, as these are compressible resources where temporary limits only slightly slow execution.
• Memory and Disk: Tuned conservatively to the p99.9 level with a safety buffer to prevent OOM (Out of Memory) or OOD (Out of Disk) crashes. Changes are clamped to a maximum 15% reduction per cycle to ensure stability.

The system operates through a state machine with stages including CREATED, EVALUATING, APPLIED, REJECTED, and MONITORED. Recommendations are generated daily based on the past month of usage data stored in Iceberg tables, accessed via Keystone streaming. Function owners must manually opt in to automatic right-sizing on a per-resource-type basis.