Amazon ECS Execution Architecture and Selection Criteria – Part 1: Compute and Execution Types

Separation of Declaration and Execution: Capacity Provider-Centric Operations

In Amazon ECS operations, the launch type is merely a parameter declaring a compatible execution environment, while capacity providers handle actual compute capacity provisioning and scaling. AWS strongly recommends setting capacity providers as the default instead of launch types when creating new services. This avoids the issue of being unable to change the launch type when switching from EC2 to Fargate and enables flexible capacity management.

An Alternative Between Fargate and EC2: ECS Managed Instances

ECS Managed Instances have emerged for customers who want both the management convenience of Fargate and the hardware control of EC2. This option creates instances in the customer's account, while AWS automatically performs OS patching, scaling, and maintenance. It is suitable for AI inference servers and high-spec workloads that require GPU or specialized hardware but want to reduce the burden of OS management.

Characteristics and Constraints by Execution Type

• Fargate: Infrastructure ownership lies with AWS, and only vCPU and memory combinations can be specified. GPU usage is not possible.
• ECS Managed Instances: Instance type selection and GPU usage are possible, with a management fee added to EC2 costs. The Managed daemons feature allows for centralized agent management.
• Express Mode: A simplified mode that deploys services immediately with just three inputs: container definition, URL, and load balancer. Support for custom Task definitions has been expanded.
• RunTask and EventBridge: One-off tasks that do not require long-running execution are handled via the RunTask API, while scheduled executions are handled by EventBridge Scheduler. RunTask allows launching up to 10 Tasks per call.