Estée Lauder Companies has adopted Cloud Run worker pools to enhance its application architecture, particularly for its consumer-facing AI offerings. This shift allows the company to efficiently manage workloads that require continuous background execution, especially during high-traffic periods like the holiday shopping season.
The company’s Rostrum platform, initially designed for internal tools, faced challenges in scaling for public use. To address this, Estée Lauder migrated to a producer-consumer model utilizing Cloud Run worker pools, which provide a reliable environment for processing AI prompts from numerous simultaneous users.
Key Achievements
By implementing this architecture, Estée Lauder Companies has realized several benefits:
- Message Durability: Utilizing Pub/Sub as a buffer ensures no user messages are lost during peak traffic.
- Improved UI Performance: The decoupling of server-side rendering from message processing enhances user experience.
- Reduced Operational Overhead: Minimal server management allows the team to focus on enhancing user experience.
Serverless Solutions for Complex Workloads
Cloud Run worker pools cater to workloads that require non-HTTP protocols, enabling high-performance operations. This capability is crucial for hosting services that were previously incompatible with Google Cloud's serverless platform.
With the general availability of worker pools, Cloud Run now supports:
- Pull-Based Workloads: Reliable scaling for tasks that continuously pull messages from various queues.
- Distributed AI/ML Workloads: Support for advanced training and fine-tuning of large language models.
Cost Efficiency
Worker pools are approximately 40% more cost-effective than traditional request-driven services for long-running tasks, making them an attractive option for organizations looking to optimize their cloud spending.
Scaling with CREMA
To facilitate scaling, Estée Lauder Companies has also implemented the Cloud Run External Metrics Autoscaler (CREMA). This tool automatically adjusts the number of instances based on external metrics from sources like Kafka and GitHub, ensuring optimal performance during traffic fluctuations.
To deploy CREMA, users need to define scaling logic in a YAML configuration file, allowing for seamless integration with existing workflows.
Getting Started
Organizations interested in deploying worker pools can refer to the official documentation for guidance. For advanced scaling options, exploring the CREMA open-source repository can provide valuable insights.
