Warner Bros. Discovery accelerates ML workflow delivery and reduces costs with Flyte
Challenge
Warner Bros. Discovery needed scalable ML orchestration across diverse data science teams.
Warner Bros. Discovery (WBD), serving 95.8 million subscribers, relies on machine learning to personalize recommendations, reduce churn, forecast revenue, and optimize messaging. Each workflow may include up to 500 features sourced from viewership, subscription, and metadata—resulting in terabytes of data processed daily.
Machine Learning Platform Engineer Frank Shen supports multiple data science groups, each using different environments (Databricks, Jupyter, SageMaker), creating fragmentation across development, testing, and production.
WBD’s reliance on Airflow created several barriers:
- Python-heavy ML code was difficult to integrate
- Notebooks weren’t compatible, forcing copy-and-paste workflows
- Local debugging wasn’t possible
- Workflow reuse was limited
- Scaling required dedicated DevOps support
WBD needed a platform that unified development workflows, reduced duplication, and supported large-scale orchestration without infrastructure overhead.
“You can compose workflows; one workflow can call other workflows, and you can chain them together and reuse them.”
Frank Shen
ML Platform Engineer at Warner Bros. Discovery
Solution
Flyte enabled Python-native development, workflow reuse, and cloud-scale execution.
Flyte became WBD’s orchestration layer for standardizing development and accelerating deployment.
Researchers can now develop in pure Python locally and annotate functions to convert them into Flyte tasks—eliminating the notebook-to-Airflow rewrite cycle. Flyte’s workflow composition capabilities allow teams to reuse pipelines, chain workflows, and reduce duplication across the organization.
Kubernetes-backed execution provides auto-scaling and resource management without requiring a dedicated DevOps team. WBD also benefited from Flyte’s active community, which quickly answered questions and addressed feature gaps.
Reusable workflows—such as shared XGBoost pipelines—further reduced engineering effort across teams.
efficiency gains across ML workflow delivery
models productized within months, not years
reduced duplication via shared workflows and components
Results
Flyte accelerated productionization and improved ML efficiency.
Within six months, WBD productized several ML models using Flyte and began migrating all Airflow training pipelines. A small ML engineering team now supports a large community of data scientists by publishing reusable workflows that eliminate repetitive work.
“A couple of us are sufficient to support a whole bunch of data scientists because we actually develop reusable Flyte workflows.” — Frank Shen
Shen also observed significant operational savings:
“My personal experience says there’s been at least 30% cost savings in efficiency.”
Flyte now serves as WBD’s AI development infrastructure—reducing operational overhead, improving developer productivity, and supporting personalized viewing experiences for tens of millions of subscribers.
