Prepare and test your skills
Prepare and test your skills
Free sample question, answered and explained below. The Professional Cloud DevOps Engineer practice bank has 1,455 more like it. Try 10 free: instant score, no account.
An AI platform engineering team is designing a provisioning workflow on Google Kubernetes Engine (GKE) for large-scale, distributed training jobs that require multi-host TPU/GPU slices. The workload has the following operational requirements and constraints:
Which accelerator provisioning strategy should the team implement?
Capacity Reservations provide guaranteed compute and accelerator availability in a specific zone. When combined with Committed Use Discounts (CUDs), they provide price predictability and reserve dedicated hardware for one-year or three-year terms.
The organization was hosting 24/7 production AI inference services or continuous high-priority training pipelines with strict SLAs requiring dedicated, zero-latency accelerator provisioning.
Capacity planning for batch AI workloads typically favors queue-based systems (like DWS) over static reservations to maximize hardware utilization across teams.
Spot VMs are spare Compute Engine capacity offered at substantial discounts (up to 60–91% off standard on-demand pricing). In exchange for these discounts, Spot instances are ephemeral and can be preempted by Google Cloud whenever capacity is needed elsewhere.
The workload consisted of embarrassingly parallel, short-lived tasks (such as individual video transcode segments or independent data processing tasks) that checkpoint state continuously and can recover instantly from node preemptions.
While Spot VMs are a core cost-optimization tool in GKE, they are generally unsuitable for large-scale synchronous distributed training where all nodes must remain online throughout the run.
Standard on-demand provisioning with the GKE Cluster Autoscaler dynamically scales node pools up and down based on pending Pod resource requests, paying standard on-demand list prices per second.
The workloads were latency-sensitive web services or API endpoints that needed immediate, non-queued scaling to handle unpredictable live user traffic bursts.
Dynamic Workload Scheduler was created specifically to overcome standard autoscaler limitations when dealing with highly constrained, expensive AI/ML hardware accelerators.
Dynamic Workload Scheduler (DWS) in Flex-start mode is a capacity provisioning mechanism designed specifically for AI/ML and batch computing workloads requiring scarce accelerators like GPUs and TPUs. Instead of failing immediately when resources are temporarily constrained, DWS queues resource requests and provisions the entire requested capacity simultaneously once available.
DWS Flex-start perfectly balances cost efficiency, batch queuing, and run stability, making it the recommended GCP architecture for high-performance accelerator workloads that need uninterrupted multi-day runs without fixed multi-year reservations.
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 Validated