AWS Certified Solutions Architect - Associate SAA-C03 Practice Test 70
Question 690
A ticketing API receives a burst of one million asynchronous tasks in one minute, but downstream Lambda processing is limited to a safe fixed concurrency. The business can tolerate delayed processing but not data loss. Which buffer should it use?
- Invoke the Lambda worker directly from the API for all tasks and depend on client-side retries during throttling, without durable job admission.
- Place work in SQS and scale consumers within the concurrency constraint while tracking message age.
- Send each independent task through a Kinesis Data Streams shard workflow and manage ordering and retention as if ordered streaming were required.
- Create one Step Functions Standard execution per incoming small task and use execution history as the only backlog admission mechanism.
Explanation
Correct Answer
B:SQS can absorb a burst of independent asynchronous jobs while Lambda event-source mapping and reserved/max concurrency settings limit concurrent consumers. Visibility timeouts, retries, a dead-letter queue and message age monitoring protect reliability during gradual drain.
Incorrect Answers
A:Direct synchronous invocation makes the burst and throttling behavior part of the request path. It lacks the independent durable backlog and controlled drain that the workload needs.
C:Kinesis can buffer streaming records, but this task queue has no partition ordering or replay-throughput requirement. Shard and consumer management add complexity versus SQS under the stated needs.
D:Step Functions can orchestrate durable workflows but is not the most direct managed work-queue buffer for millions of unrelated short jobs awaiting bounded Lambda consumers.
Question 691
A game service needs low cold-start latency for its latency-sensitive Lambda API. The engineer sets reserved concurrency to 50 and assumes 50 function instances stay initialized at all times. What distinction matters?
- Raise reserved concurrency to 100 and treat the new value as a pool of one hundred preinitialized function environments.
- Enable SnapStart on the function, assuming it guarantees a chosen number of always-ready concurrent executions for every Lambda runtime.
- Keep reserved concurrency for execution isolation, but configure supported provisioned concurrency on a published version or alias to keep the required number of execution environments initialized.
- Lower the function timeout to one second, treating a shorter configured duration as equivalent to initializing execution environments before requests arrive.
Explanation
Correct Answer
C:Reserved concurrency governs available execution slots. Provisioned concurrency initializes and maintains configured execution environments for a qualified version or alias, addressing latency-sensitive cold starts.
Incorrect Answers
A:Increasing reserved concurrency changes allowed concurrent execution capacity, not the number of preinitialized runtime environments. Cold starts can still occur as the function scales.
B:SnapStart can reduce initialization latency on supported runtimes and configurations, but it does not reserve a specific number of prewarmed execution environments and has compatibility constraints.
D:A function timeout limits maximum invocation duration; it does not prepare a runtime before invocation and can instead terminate legitimate requests.
Question 692
An ECS task definition on Fargate specifies an arbitrary fraction of a vCPU combined with an unsupported memory size and is rejected at registration. How should the architect size tasks?
- Use any CPU and memory numbers because Fargate inherits the unlimited sizing flexibility of EC2 bare-metal instances.
- Increase the desired task count while keeping an invalid task definition unchanged.
- Add an ALB listener rule with a larger timeout to change the Fargate validation matrix.
- Choose a supported Fargate CPU-memory combination appropriate to measured container requirements.
Explanation
Correct Answer
D:Fargate enforces task-level resource configuration combinations rather than accepting arbitrary EC2 host shape ratios.
Incorrect Answers
A:The managed serverless compute platform constrains supported combinations.
B:An invalid resource specification prevents tasks from launching regardless of desired count.
C:Load balancer rules do not control permitted task resource configuration.
Question 693
An ECS Fargate API service has predictable daily traffic peaks, but its desired count is static. Scale-out takes a minute for tasks to become ready. What operating policy should it use?
- Configure ECS Service Auto Scaling with meaningful demand metrics and startup-aware target health, optionally scheduling predictable pre-scaling.
- Increase the database backup retention as the exclusive mechanism to scale Fargate tasks.
- Set desired count permanently to one and assume the platform will transparently duplicate it under load.
- Increase only the task's DNS TTL to schedule more task copies.
Explanation
Correct Answer
A:Task count should track real demand while allowing for startup time and readiness before routing traffic.
Incorrect Answers
B:Backup retention does not alter ECS desired count.
C:Fargate runs requested tasks; it does not automatically infer every application's necessary replicas.
D:DNS cache duration cannot launch new Fargate tasks.
Question 694
Three ECS Fargate microservices each need access to different S3 prefixes. Engineers plan to give every task the same broad credentials through environment variables. Which identity model limits risk?
- Store one account root access key in all three task environments for convenience.
- Assign per-service ECS task IAM roles with only required S3 actions and resources.
- Use only the ECS task execution role as a blanket application data-access grant for every container.
- Enable VPC DNS resolution and remove IAM policies from all S3 calls.
Explanation
Correct Answer
B:A task role supplies application credentials to that workload and can be scoped independently, unlike shared static keys.
Incorrect Answers
A:Root keys create severe shared privilege and leak risk across services.
C:The execution role primarily supports agent actions such as image pulls and logging, not as a replacement for tightly scoped application task roles.
D:DNS resolution cannot authorize access to private S3 objects.
Question 695
An EKS deployment uses HPA and increases desired pods from 4 to 20 during a traffic surge. Eight new pods remain Pending with FailedScheduling events because all worker nodes lack allocatable CPU. Adding more pod replicas does not create more worker-node capacity. Which mechanism addresses this bottleneck?
- Increase the HPA minimum replica count from 20 to 40 while leaving node capacity fixed.
- Increase the cluster's HPA minimum replica count while leaving all worker-node capacity unchanged, assuming the PodDisruptionBudget will schedule the extra pods.
- Configure EKS-supported cluster-compute autoscaling such as Karpenter with valid NodePools to provision nodes for unschedulable pods.
- Increase the CPU limits for every Pending pod but leave its CPU requests and the node fleet unchanged.
Explanation
Correct Answer
C:HPA adjusts the number of pods, while Karpenter or supported cluster autoscalers provision node capacity when the scheduler cannot place those pods. NodePools must allow the requested instance attributes and capacity.
Incorrect Answers
A:More desired pods on exhausted nodes will remain Pending and do not solve the node-resource shortage.
B:Horizontal Pod Autoscaler changes desired pod count; a PodDisruptionBudget protects availability during disruption. Neither allocates EC2 node CPU or schedules Pending pods when no node has capacity.
D:Scheduling depends on each pod's requests and available node allocatable CPU. Increasing limits alone cannot supply missing worker capacity or resolve FailedScheduling events.
Question 696
An EKS Horizontal Pod Autoscaler requests 25 replicas, but twelve pods remain Pending with Insufficient cpu messages; all available nodes are saturated. What capability is missing?
- Only raise HPA maximum replicas while leaving the node pool full.
- Reduce Kubernetes Service DNS TTL so Pending pods are scheduled faster.
- Increase a CloudWatch logs retention setting and expect it to create more node cores.
- Provide node capacity through a compatible node autoscaler, managed provisioning or node group scaling.
Explanation
Correct Answer
D:HPA increases desired pods, but a scheduler cannot place them without adequate worker-node compute resources.
Incorrect Answers
A:Additional desired pods increase the Pending count without provisioning CPUs.
B:DNS is not the scheduler's missing allocatable CPU capacity.
C:CloudWatch Logs retention controls how long stored log events are kept; it does not modify EC2 vCPU capacity, EMR worker-node allocations or Spark executor slots. The workload must scale compute resources rather than log retention.
Question 697
A new EKS application pod opens its port immediately but needs 40 seconds to load configuration. The load balancer sends traffic before it is usable, causing 503s; the pod also lacks accurate resource requests. What should be configured?
- Add truthful Kubernetes readiness probes and CPU/memory requests matching scheduling needs before routing service traffic.
- Replace readiness checks with a liveness probe that is always successful.
- Increase the deployment's metadata label count without specifying resource requests or health checks.
- Force traffic to every Running pod regardless of initialization state.
Explanation
Correct Answer
A:Readiness gates protect clients from unready pods, while resource requests allow the scheduler to place tasks on adequate nodes.
Incorrect Answers
B:Liveness cannot establish whether the application is ready to accept traffic.
C:Labels alone do not reserve resources or report application readiness.
D:Running phase does not imply the service endpoint can process real requests.
Question 698
An EKS cluster has ample managed Kubernetes control-plane availability, but its only worker node is CPU-bound and pods are Pending. The team requests a larger managed control plane as the fix. Which part must scale?
- Increase EKS API-server or managed control-plane throughput without adding schedulable worker CPU.
- Increase or autoscale worker node or compatible serverless compute resources needed by pods.
- Install Cluster Autoscaler but leave the backing node group maximum size fixed at its exhausted single-node capacity.
- Raise resource requests for Pending pods to make them schedule earlier on the same saturated worker.
Explanation
Correct Answer
B:Unschedulable application pods require adequate worker-node allocatable resources or supported Fargate pod compute. A managed EKS control plane does not replace nodes; scaling node groups or compatible autoscaled pod capacity addresses the bottleneck.
Incorrect Answers
A:A managed Kubernetes control plane schedules and administers workloads; its API capacity is not the CPU resource pool for application pods.
C:Autoscaling cannot provision additional schedulable CPU if its underlying capacity-provider/node-group limits prevent any additional nodes.
D:Increased pod CPU requests make scheduling constraints stricter, not easier, when available node allocatable CPU is already exhausted.
Question 699
A Glue-style CSV normalization job reads from S3 and performs lightweight string parsing. Engineers propose dedicated GPU instances for the whole month even though profiling shows I/O waiting and minimal vector computation. What is better?
- Choose P-family GPU instances solely because the dataset contains millions of rows.
- Buy a larger GPU memory tier without changing the slow source-read path.
- Use appropriately sized CPU or managed data-processing compute matched to I/O needs instead of idle GPUs.
- Use GPU-intensive training instances for all periodic text parsing because GPU instances are automatically always cheaper.
Explanation
Correct Answer
C:GPU hardware adds cost without accelerating the bottleneck when the workload is basic I/O-bound record processing.
Incorrect Answers
A:Row count does not prove GPU-compatible parallel kernels or GPU-bound processing.
B:More accelerator RAM cannot remove the measured I/O wait.
D:Specialized hardware price/performance depends on actual accelerator utilization.
Question 700
A machine-learning inference container requires 45 GiB of GPU-resident weights and activations per model replica, and latency rules out paging model layers to host RAM. A proposed instance has four 12-GiB GPUs, while the inference framework runs each model replica on one GPU and does not support tensor or pipeline parallelism. Which design decision resolves the memory constraint?
- Place four full model replicas on the four 12-GiB GPUs and enable EC2 enhanced networking to address the per-replica GPU allocation failure.
- Select a larger host-memory EC2 instance with the same 12-GiB GPUs and increase CPU-to-GPU PCIe transfer bandwidth to provide 45 GiB of device-resident VRAM.
- Increase the number of 12-GiB GPUs and treat aggregate VRAM as a contiguous address space, leaving the one-GPU-per-replica model unchanged.
- Select an accelerator with sufficient usable VRAM per model replica, or redesign and validate inference software for supported model parallelism.
Explanation
Correct Answer
D:The resident weights plus working set must fit in accessible accelerator memory. A larger single device meets the current software model; validated model sharding is the alternative if supported.
Incorrect Answers
A:Enhanced networking affects packet processing and network performance. It does not let a 45-GiB resident model fit in a 12-GiB device allocation when each replica is assigned one GPU.
B:Host RAM capacity and PCIe throughput do not increase the physical device VRAM available to a GPU-resident replica, and the scenario explicitly excludes paging as a latency workaround.
C:Multiple GPUs provide separate device memories. Without supported distribution across devices, an individual replica remains constrained by one 12-GiB card.
