Batch

Batch workloads — genomics pipelines, financial risk runs, rendering, ETL, simulations — arrive as thousands of independent jobs that need a lot of compute for a while and none the rest of the time. Running a permanent cluster wastes money; hand-scripting Auto Scaling around a queue is fiddly. AWS Batch does it: you submit jobs (a container + vCPU/memory + command) to a job queue, and Batch provisions exactly the compute a compute environment needs (EC2, Spot, or Fargate), runs the jobs with priorities and dependencies, retries failures, and scales the fleet back to zero when the queue drains.

• Job queues, compute environments, and job definitions are Regional. A managed compute environment launches instances across the subnets/AZs you give it, so Batch can place jobs in whichever AZ has capacity (important for Spot).
• Compute environment types: managed (Batch owns capacity — EC2, EC2 Spot, or Fargate/Fargate Spot) or unmanaged (you attach your own ECS-registered capacity).
• For tightly-coupled multi-node parallel (MPI) jobs, Batch can place instances in a cluster placement group in one AZ for low-latency interconnect.

graph LR
SUB["Submit jobs (container + vCPU/mem + cmd)"] --> Q["Job Queue (priority)"]
Q --> BATCH["AWS Batch scheduler"]
BATCH --> CE
subgraph CE["Managed Compute Environment (multi-AZ)"]
  FG["Fargate / Fargate Spot"]
  EC2N["EC2 On-Demand"]
  SPOT["EC2 Spot fleet"]
end
CE -->|"runs"| JOBS["Containerized jobs (retries, array, dependsOn)"]
CE -.->|"queue empty"| ZERO(("Scale to 0"))
classDef q fill:#f3ecfb,stroke:#8b3de0,stroke-width:2px,color:#3a1a5c
classDef job fill:#e8f0fe,stroke:#3b82f6,stroke-width:2px,color:#173a70
class Q,BATCH q
class JOBS job

AWS Batch adds no charge. You pay only for the EC2 or Fargate capacity while jobs run, plus EBS and data transfer. The savings come from scale-to-zero between runs and from letting Batch use Spot (up to ~90% off) with automatic fallback and retry for interrupted jobs. Because the fleet is transient, a well-tuned queue costs a fraction of a standing HPC cluster.

• Trigger phrases: “run thousands of batch jobs,” “no cluster to manage,” “scale compute to zero when idle,” “cost-optimize large-scale batch / HPC / rendering / genomics” → AWS Batch (usually with Spot).
• Batch vs Lambda: Lambda caps at 15 minutes and 10 GB memory; Batch jobs can run for hours or days on large/GPU instances. Use Batch for long, heavy compute.
• Batch vs Step Functions: Step Functions orchestrates a workflow of steps/services; Batch executes a large fleet of similar compute jobs. They’re often combined (Step Functions triggers Batch jobs).
• Array jobs run N copies indexed by AWS_BATCH_JOB_ARRAY_INDEX; job dependencies (dependsOn) build simple DAGs; job priority orders queues sharing a compute environment.
• Use EC2 launch template / custom AMI for extra drivers (e.g. GPU), larger scratch disk, or an agent. Fargate compute environments are simplest but can’t do GPU or very large instances.
• Send job logs to CloudWatch Logs; use EventBridge on job state changes for notifications and downstream triggers.