AZ-305 Practice Question 38: Parallel Batch Compute
Microsoft Azure Solutions Architect Expert (AZ-305) · Test 4 · question 8 of 10
A hedge fund runs a Monte Carlo risk simulation every night. The workload is a command-line executable invoked roughly 60,000 times with different inputs; the invocations are independent and do not communicate. A run consumes about 20,000 core-hours and must finish inside a six-hour window, and an individual invocation that is interrupted can simply be run again. The fund wants to minimize cost and does not want to install or operate scheduler software.
Which compute solution should you recommend?
Show answer and explanation
Answer:
Azure Batch is described by Microsoft as the service for large-scale parallel and HPC batch jobs, and Monte Carlo risk modeling is the first example it lists of an intrinsically parallel workload. Batch creates and manages the pool of compute nodes, installs the application, and schedules the tasks (there is no cluster or job scheduler software to install, manage, or scale), and the documentation points to Azure Spot VMs and automatic scaling as the way to manage the cost and size of a pool. Because an interrupted invocation can simply be rerun, Spot eviction is an acceptable trade for the discount, and Batch itself carries no extra charge beyond the underlying resources.
Option A would work and adds a platform the fund does not need. It takes on cluster upgrades, node pool management, and Kubernetes expertise to run a workload that is a queue of independent executables, not a set of long-lived services.
Option B is the "build your own Batch" answer. Scale sets provide the virtual machines; the work distribution, task retry, node health handling, and result collection are all code the fund writes and maintains, which is exactly the scheduler software it wants to avoid.
Option C fits the trigger model but not the workload. Functions on the Consumption plan are tuned for short, event-driven executions with an execution time limit, and tens of thousands of long, CPU-bound simulation runs on a serverless plan is both an awkward fit and more expensive per core-hour than a Spot pool.
Reference: Azure Batch runs large parallel jobs in the cloud