AWS re:Invent 2025 - 5-Star Customer Service: Duolingo's Path to Compute Savings (CMP349)

Duolingo's Path to Compute Savings with AWS Graviton

Leveraging Managed Services with Graviton

• Duolingo leveraged AWS managed services like Amazon Elasticache and Amazon RDS that support Graviton processors to achieve immediate cost savings.
• For Amazon Elasticache, Duolingo migrated from R5 instances to R6G Graviton-based instances, carefully planning the migration to minimize disruption:
  • Created a new Graviton-based cache and gradually shifted traffic from the old cache to mitigate risk.
  • Monitored cache usage and found most workloads only used a small fraction of the NVME storage, allowing a migration to less expensive EBS-backed instances.
• For Amazon RDS, Duolingo followed a similar approach, first testing the migration in a canary environment before rolling out to production:
  • Analyzed RDS instance utilization and found most workloads did not require the high-performance NVME storage.
  • Migrated RDS instances from NVME to EBS-backed, taking advantage of the lower interruption rates on the spot market for EBS instances.
• These managed service migrations resulted in over 75% of Duolingo's managed services running on Graviton, delivering a 10% reduction in RDS and Elasticache costs.

Expanding the EC2 Instance Pool

• Duolingo initially ran most of their EC2 workloads on NVME-backed instances due to high storage requirements, but as their observability and logging improved, they found they could leverage less expensive EBS-backed instances.
• To expand their instance pool and improve flexibility, Duolingo:
  • Analyzed instance utilization and found most workloads were not fully utilizing the NVME storage.
  • Leveraged the AWS Spot Instance Advisor to evaluate interruption rates for EBS vs. NVME instances, finding EBS instances had lower interruption rates.
  • Gradually rolled out EBS-backed instances, first in a canary environment, then expanding to dev, staging, and production environments.
• This instance pool optimization resulted in over 10% savings on Duolingo's EC2 costs without changing any application logic, while also improving their ability to adopt newer instance types more quickly.

Migrating Microservices to Graviton

• Duolingo started early with ARM-based development, having their engineers "dogfood" ARM-based laptops to identify compatibility issues.
• They selected their critical monolithic "Duolingo 2" application as the first target to migrate to Graviton, using a phased approach:
  • Built and tested the application locally on Graviton, addressing any package or code compatibility issues.
  • Deployed the Graviton-based version to a canary environment, closely monitoring for any runtime issues.
  • Gradually expanded the Graviton deployment to smaller production environments, then full production.
• The Graviton migration of the monolith resulted in a 10% reduction in task count and 20% faster deployment times.
• Duolingo then expanded their Graviton adoption, targeting their top 5 most resource-intensive microservices, which represented 25% of their EC2 spend.
  • These migrations were facilitated by improvements to their CI/CD tooling, allowing developers to easily choose between x86 or Graviton-based deployments.
  • In total, Duolingo now has over 30% of their microservices running on Graviton, resulting in a 20% reduction in their overall EC2 costs.

Looking Ahead: Future Compute Initiatives

• Duolingo is planning further initiatives to accelerate their Graviton adoption, with a goal of reaching 50% of their compute on Graviton by 2026.
• They are also exploring the use of Kubernetes (EKS) to improve flexibility and enable features like a service mesh and ephemeral staging environments.
• Additionally, Duolingo is investigating a multi-architecture strategy, allowing them to deploy the same application image on either x86 or Graviton instances, providing more options to optimize for cost and performance.

Key Takeaways

• Duolingo achieved significant cost savings and performance improvements by leveraging AWS Graviton processors across managed services, EC2 instances, and microservices.
• Careful planning, phased rollouts, and close collaboration with AWS technical account managers were critical to the success of Duolingo's Graviton adoption.
• Expanding the instance pool, optimizing for storage requirements, and improving CI/CD tooling enabled Duolingo to accelerate their Graviton migration.
• Duolingo's journey demonstrates how the right partnership and technical expertise can help organizations transform cloud constraints into opportunities for innovation and efficiency.