Prepare and test your skills
Prepare and test your skills
Free sample question, answered and explained below. The Professional Cloud Architect practice bank has 2,335 more like it. Try 10 free: instant score, no account.
An enterprise machine learning team is optimizing its infrastructure on Google Cloud to train large deep learning models using specialized accelerators, such as NVIDIA A100 GPUs. The training jobs are batch-oriented, delay-tolerant, and require guaranteed runtime completion once started, without unexpected mid-run interruptions or preemptions.
The team wants to optimize accelerator acquisition and compute costs while ensuring efficient scheduling of high-demand GPU resources on Vertex AI.
Which compute scheduling architecture should the team implement for their Vertex AI custom training jobs?
NVIDIA Triton Inference Server on Vertex AI Endpoints is an optimized serving runtime tailored for high-concurrency, low-latency online model inference, supporting multi-framework model execution on GPU and CPU compute instances.
rawPredict or predict), whereas the requirement is to run asynchronous, batch deep learning model training.The organization needed to deploy a trained multi-framework model to a production endpoint to serve low-latency online inference requests from web clients.
While Triton is part of the broader AI acceleration stack, compute architecture must separate training pipelines (CustomJob) from serving endpoints (Endpoint).
Spot VMs are surplus Compute Engine compute instances available at significant discounts. However, they can be reclaimed by Google Cloud at any time with a 30-second preemption notice whenever the capacity is needed for standard on-demand workloads.
The workload consisted of short, stateless tasks or fault-tolerant hyperparameter searches that can tolerate frequent terminations and restarts.
A standard Vertex AI TrainingPipeline provisions compute nodes immediately on-demand using standard Google Cloud GPU quota without any queuing or specialized workload scheduling mechanisms.
RESOURCE_EXHAUSTED).custom_model_training_nvidia_a100_gpus quota rather than flexible scheduling pools.The training job were time-critical, had strict real-time starting requirements, and had guaranteed on-demand quota reservations.
For delay-tolerant batch training on constrained AI Hypercomputer infrastructure, relying on unmanaged on-demand scheduling leads to frequent job launch failures.
Dynamic Workload Scheduler (DWS) is an advanced compute orchestration feature in Google Cloud and Vertex AI designed to optimize accelerator allocation for large-scale AI/ML workloads. It queues training jobs until the requested compute accelerators (such as NVIDIA A100 or H100 GPUs) become fully available, running them with predictable performance and discounted economics.

FLEX_START strategy, jobs consume preemptible quota allocations rather than scarce on-demand quotas, offering cost-optimized pricing tiers while acquiring top-tier accelerator hardware.max_wait_duration (e.g., 1800s), the team defines how long Vertex AI will hold the job in queue waiting for GPU capacity before timing out.aiplatform.CustomJob using gca_custom_job_compat.Scheduling.Strategy.FLEX_START.DWS FLEX_START strikes the exact balance between cost optimization and workload stability for large AI models, providing discounted, non-preemptible accelerator runtime for batch training.
Answer them yourself and get your score the moment you finish. No account needed.
Full access includes
Keep the momentum going with these hand-picked practice scenarios
Last updated