How Netflix Built the Cosmos Platform to Power Massive Media Workflows

Introduction

Cosmos is a compute platform that blends the best traits of microservices, asynchronous workflows, and serverless to support resource‑intensive, long‑running algorithms that can span minutes to years, handling both high‑throughput and latency‑sensitive workloads.

Background

Netflix’s Media Cloud Engineering team operated a system that evolved from a 2007 streaming service to a third‑generation platform called Reloaded, which became difficult to scale and maintain as the team and use‑cases grew, leading to a monolithic architecture that hindered productivity.

To address these challenges, Netflix built Cosmos, a media‑centric, workflow‑driven microservice platform with goals of observability, modularity, developer productivity, and automated delivery.

Overview

Cosmos services retain strong contracts and isolated data like microservices but add multi‑step workflows and compute‑heavy, asynchronous serverless functions packaged as Docker images. Requests flow through an API layer (Optimus), a workflow engine (Plato), and a serverless layer (Stratum), scaling across thousands of containers.

Separation of Concerns

Cosmos separates logic into API, workflow, and serverless functions, and also separates application concerns from platform concerns. The platform API abstracts media‑specific details while hiding distributed computing complexities.

Key subsystems:

Optimus – API layer mapping external requests to internal models.

Plato – Workflow layer for business rule modeling.

Stratum – Serverless layer for stateless, compute‑intensive functions.

All subsystems communicate asynchronously via Timestone, a large‑scale low‑latency priority queue, enabling independent deployment through managed continuous delivery.

Workflow Rules

Developers define workflows using Emirax, a Groovy‑based DSL, with four parts: match, action, reaction, and error. Rules trigger Stratum functions, record state changes, and handle errors.

Latency‑Sensitive Applications

Services like Sagan require low latency; Stratum manages delay by using resource pools, warm capacity, micro‑batches, and priority scheduling to balance cost and responsiveness.

Throughput‑Sensitive Applications

Services such as Tapas prioritize high throughput, consuming millions of CPU hours and focusing on daily task volume rather than individual request latency. Stratum’s serverless layer on the Titus container platform enables opportunistic resource scheduling.

Strangler Fig Migration

To replace the legacy Reloaded system, Netflix adopted the Strangler Fig pattern, gradually extending new functionality around the old system until full migration.

Lessons Learned

Key insights include the importance of Netflix’s engineering culture, the power of a microservice + workflow + serverless model, and the need for a platform mindset that balances flexibility for application teams with consistency and reliability from the platform team.

Future Plans

In 2021, Netflix will continue moving workloads to Cosmos, improving the programming model, and enhancing usability, resilience, speed, and efficiency.