Visual overview
Jobs wait in a queue and are scheduled onto an associated compute environment.
Technical reference
Batch separates job definition, queueing, scheduling, and compute supply.
Defines container image, resources, roles, environment, retries, and timeouts as applicable.
Submitted jobs wait until they can be placed.
Runs jobs on supported EC2/Fargate or EKS-backed environments.
Associated environments and queue priority influence placement.
Batch work is different from a long-running service
Batch workloads consist of finite jobs that can be queued, scheduled and executed when suitable compute is available. Examples include scientific processing, media transformation, data processing, simulation, rendering and other compute-intensive work that does not need to sit permanently behind a request endpoint. AWS Batch provides managed orchestration for this model and can scale compute capacity on Amazon EC2 instances or AWS Fargate resources.
Instead of maintaining a fixed fleet of worker servers and building a custom scheduler, you submit jobs with resource requirements. AWS Batch manages queues, scheduling and the relationship with configured compute environments. This lets capacity expand for a backlog and contract when work is complete, subject to the compute model and configuration selected.
Job definitions, queues and compute environments
A job definition describes how a job should run, including its container image, resource requirements and other runtime settings. A job queue holds submitted work until the scheduler can place it. Compute environments describe the capacity onto which eligible jobs can be scheduled. These layers allow application teams to express work separately from the exact machines that eventually execute it.
Queues can represent priorities or workload classes, while the scheduler decides when eligible jobs run according to configured policies and available capacity. This separation is valuable for throughput-oriented systems because the same compute environment can serve changing backlogs without every job needing to know which host will execute it.
EC2, Spot and Fargate choices
EC2-backed compute environments offer broad instance-type flexibility and can take advantage of Spot capacity for interruption-tolerant jobs. This is useful when batch workloads need specialized instances, accelerators, large fleets or cost optimization through diversified capacity. The trade-off is that EC2-based environments expose more fleet and capacity decisions.
Fargate-backed compute environments remove the need to manage server clusters and launch right-sized container compute for jobs. They are often simpler to operate when jobs fit Fargate's supported resource and platform constraints. AWS recommends evaluating Fargate for many general scenarios and EC2 when workload requirements need capabilities or scale characteristics that Fargate does not provide.
Reliability and cost thinking
Batch jobs should be designed for retry and should store durable inputs and outputs outside the lifetime of a worker. When using Spot-backed capacity, interruption tolerance becomes even more important: jobs that can checkpoint progress or safely restart are much better candidates than work that loses hours of state whenever a worker is reclaimed.
Cost optimization comes from matching the job's actual resource requirements to appropriate compute and avoiding a permanently idle fleet. Queue depth, startup latency, throughput targets, retry behavior and resource efficiency all matter. Batch is therefore not simply 'EC2 that starts later'; it is a managed scheduling and fleet-orchestration layer designed around queued work.
Key takeaways
- 01
AWS Batch manages queues, scheduling and compute scaling for finite batch jobs.
- 02
Job definitions describe work, job queues hold it, and compute environments provide capacity.
- 03
EC2 offers broader infrastructure choices; Fargate removes server-fleet management for compatible jobs.
- 04
Reliable batch workloads externalize durable state and tolerate retries or interruptions when their capacity model requires it.
Official AWS sources
Use these primary AWS resources for the source material behind this article and for deeper reference.