SAP-C02 Practice Question 54: GPU Batch Pipeline Compute
AWS Certified Solutions Architect – Professional (SAP-C02) · Test 6 · question 4 of 10
A genomics company runs a nightly analysis pipeline. Each run submits about 40,000 containerized jobs. Individual jobs take between 20 and 90 minutes. The alignment stage needs GPU instances, the variant-calling stage that follows needs only CPU and memory, and the second stage must not begin for a sample until that sample's first stage has succeeded.
Every job is idempotent and safe to restart, and the pipeline has an eight-hour window to finish, so an interrupted job simply runs again. Cost is the dominant design constraint. The platform team is four people and will not staff a cluster scheduler, cluster autoscaler, or capacity planning process.
Which design should the solutions architect choose?
Show answer and explanation
Answer:
AWS Batch exists to run batch computing workloads without the team owning the scheduler or the capacity. AWS describes it as removing the undifferentiated heavy lifting of configuring and managing the infrastructure, provisioning compute resources efficiently in response to submitted jobs to eliminate capacity constraints and reduce compute costs. It scales capacity across EC2 instances, Fargate, and container orchestration underneath, so the GPU stage and the CPU stage get their own compute environments sized to different instance types and the right hardware for each. Running them on Spot capacity is what makes cost the winner here, and it is safe because every job is idempotent and restartable inside a generous eight-hour window. Job dependencies express "sample N's variant-calling waits for sample N's alignment" natively, so no external orchestration is needed for the ordering.
Option A fails on runtime. Lambda functions are capped at 15 minutes, and the jobs run for 20 to 90 minutes, so the work cannot complete regardless of concurrency settings.
Option B produces the right technical outcome but hands the four-person team exactly what they said they would not staff: a cluster to upgrade, a node group and autoscaler to tune, and a queueing controller to maintain. A permanently provisioned cluster also pays for idle capacity outside the nightly window.
Option D breaks on the GPU stage; AWS Fargate does not offer GPU-backed tasks, so the alignment work has nowhere to run. It also forgoes the Spot savings that dominate the cost of 40,000 long-running jobs.
Reference: What is AWS Batch?