AWS Certified Solutions Architect - Associate SAA-C03 Practice Test 91
Question 900
A Fargate task requests 4 vCPUs and a large memory allocation, but metrics never exceed 0.5 vCPU or 2 GiB. The app has no special reserve-memory needs. What change targets resource waste?
- Lower the number of running ECS tasks with service scaling while retaining the oversized CPU and memory allocation of each remaining task.
- Set each task to 0.5 vCPU while keeping a 16-GiB memory allocation, without checking the Fargate sizing matrix.
- Purchase Compute Savings Plans only and keep the original oversized task definition unchanged.
- Benchmark supported smaller Fargate CPU-memory combinations and revise task sizing with safe peak headroom.
Explanation
Correct Answer
D:Fargate charges for the task’s configured vCPU and memory allocations, not just measured use. Load-test the smallest supported combination that accommodates real memory peaks, CPU burst demand and safe headroom before adjusting the task definition.
Incorrect Answers
A:Service scaling can reduce overall task count at low demand but does not address waste in the specified per-task 4-vCPU, high-memory resource shape.
B:Fargate enforces supported CPU-memory combinations; 0.5-vCPU Linux Fargate tasks do not support a 16-GiB memory setting.
C:Commitment discounts can lower eligible compute rates but do not correct overprovisioned per-task resources, nor remove resource waste at peak.
Question 901
A Fargate service supports a modest overnight request rate but keeps 70 tasks running because that count was needed during Black Friday. No service autoscaling is configured. What should be done?
- Use ECS Service Auto Scaling and minimum/maximum counts based on actual requests or appropriate utilization signals.
- Keep the Black Friday desired count all year because Fargate task billing stops automatically when tasks are idle.
- Reduce only log retention without reviewing the 70 continuously running containers.
- Increase the service desired count at midnight to minimize costs.
Explanation
Correct Answer
A:Scaling down unnecessary desired tasks avoids paying for idle task allocations while retaining needed baseline availability.
Incorrect Answers
B:Running Fargate tasks incur charges for their resource allocations.
C:Log savings do not reclaim the largest idle task CPU/memory expense.
D:Adding tasks to low load generally raises allocated resource spending.
Question 902
A bioinformatics company already uses an AWS Batch managed EC2 Spot compute environment with checkpointable jobs. The job queue regularly experiences interruptions because a few cheap instance pools are oversubscribed. The organization wants an AWS-managed instance-pool selection strategy that considers BOTH Spot price and available capacity, instead of simply selecting the cheapest pool. Which setting best meets that requirement?
- Set allocationStrategy=BEST_FIT and assume it has the same Spot pool diversification and interruption behavior as price-capacity optimization.
- Use the AWS Batch SPOT_PRICE_CAPACITY_OPTIMIZED instance-type allocation strategy, with a diverse set of eligible instance types.
- Select a fixed single EC2 instance size and one Availability Zone; this guarantees the lowest interruption rate on any Spot pool.
- Set the compute environment to On-Demand but call it Spot pricing; allocation strategies can convert normal running-hour prices into Spot prices.
Explanation
Correct Answer
B:AWS Batch SPOT_PRICE_CAPACITY_OPTIMIZED considers both price and available Spot capacity to choose eligible instance pools that are comparatively less likely to be interrupted while preserving low prices. Compatible type/AZ flexibility improves pool selection.
Incorrect Answers
A:BEST_FIT is not the documented Spot price-and-capacity allocation strategy and does not give the requested cost/interruption balance.
C:Restricting flexibility to a single size and AZ shrinks available Spot capacity pools and cannot guarantee fewer interruptions.
D:On-Demand and Spot are distinct capacity purchasing models; an allocation label alone does not turn On-Demand into Spot discounted usage.
Question 903
A Batch transform requests 32 vCPUs and 128 GiB per job, but observed peaks are only 4 vCPUs and 12 GiB. Its excessive requests block scheduling and waste allocated worker resources. What should change?
- Request still more vCPUs to reduce the bill despite low actual CPU usage.
- Reduce job memory below observed peak because scheduling speed is more important than correctness.
- Set realistic job-level vCPU/memory requests based on measured peaks plus reasonable headroom.
- Add more S3 object tags as a substitute for task resource declarations.
Explanation
Correct Answer
C:Declared resource requirements affect how Batch schedules jobs and the capacity paid for or kept idle.
Incorrect Answers
A:Larger allocations can worsen placement and resource waste.
B:Underrequested memory can cause OOM failures and costly retries.
D:Tags do not determine Batch job CPU or memory allocation.
Question 904
A company uses a temporary AWS Batch compute environment only for a two-hour monthly reconciliation. After completion, several EC2 worker hosts remain provisioned for the next 30 days. Which lifecycle control helps?
- Leave the maximum worker count running because Batch never incurs charges when there are no queued jobs.
- Increase completed job log retention as a way to stop EC2 costs.
- Move the worker AMI to another AWS Region while retaining the original running hosts.
- Scale the compute environment down or decommission idle temporary capacity after confirmed job completion.
Explanation
Correct Answer
D:Releasing idle worker infrastructure prevents needless instance-hours between separated processing windows.
Incorrect Answers
A:Provisioned EC2 compute still incurs charges while active.
B:Increasing the number of days that batch logs are retained changes log-storage usage, not whether EC2 worker instances continue running. Avoiding compute charges requires actual worker scale-in, shutdown or fleet lifecycle controls.
C:Image copying does not stop active compute billing.
Question 905
A distributed nightly ETL job reads the same 6-TB raw dataset from another Region on every run, producing large cross-Region transfer charges. Compute costs are modest. Which placement decision may help?
- Stage or retain appropriate data near the processing Region, balancing copy/storage costs against repeated transfer and freshness requirements.
- Keep all data remote and scale worker vCPUs solely to reduce billed network bytes.
- Always make a fresh full copy for every step even when a suitable local staged dataset exists.
- Move compute to the cheapest hourly Region without checking data locality or output destinations.
Explanation
Correct Answer
A:Repeated remote input movement can dominate processing economics; local reuse can reduce recurring data-transfer expense.
Incorrect Answers
B:Faster CPU does not inherently eliminate the same repeated cross-Region payload.
C:Creating another complete copy of an already staged dataset at each pipeline step adds requests, storage and data-transfer work. Reusing a suitable intermediate dataset or designing a direct data flow can reduce redundant movement.
D:Lower instance rates may be outweighed by much higher network traffic.
Question 906
A cost-reduction script is scheduled to stop every EC2 instance each night, including the company's 24/7 payment database and web tier. What constraint should change the policy?
- Put the 24/7 payment and web instances on a nightly EventBridge Scheduler stop policy with no exception for customer traffic.
- Apply schedules only to verified nonproduction or safely stoppable resources while maintaining production availability requirements.
- Create an Auto Scaling scheduled action that reduces the production group desired capacity to zero nightly even though the service has global users.
- Buy a lower-cost one-year compute commitment but continue to assume production hosts may be shut down regardless of service hours.
Explanation
Correct Answer
B:Restrict scheduled stop/start automation to development, test or demonstrably idle workloads that have an accepted outage window. For customer-facing services retain minimum healthy capacity and use safe scaling and rate optimization instead.
Incorrect Answers
A:A schedule can save idle nonproduction compute but deliberately stopping a required always-on payment tier violates the stated availability requirement.
C:Auto Scaling scheduling is valid, but the proposed zero-capacity policy would intentionally remove the production service during its required operating window.
D:A Savings Plan or reservation changes billing economics, not the business requirement for uninterrupted production application capacity.
Question 907
A developer uses a supported EBS-backed EC2 instance with a large in-memory workspace. It must be off overnight, but restoring application memory state from scratch takes 20 minutes. Which shutdown mode should be evaluated?
- Use terminate and assume in-memory process state survives in the next new instance automatically.
- Use EC2 stop exclusively because all stopped instances automatically retain RAM like hibernation.
- Evaluate EC2 hibernation for supported instance/AMI/EBS configurations, comparing preserved memory state and storage costs.
- Store the instance's private IP in Route 53 and assume this recreates application memory after a stop.
Explanation
Correct Answer
C:Hibernation saves RAM contents to EBS for faster stateful resume where service prerequisites are met.
Incorrect Answers
A:Termination does not preserve arbitrary RAM as a restorable running process.
B:Normal stop and hibernation have different state-preservation behavior.
D:Route 53 stores DNS records mapping names to addresses; it does not snapshot EC2 process memory. After an instance stop or replacement, application state must be recovered from durable backing storage or a supported checkpoint.
Question 908
A project stops unused EC2 machines but continues paying for unattached EBS volumes and idle public IPv4 addresses. Finance says stopped compute should mean zero bill. What should be audited?
- Assume every associated EBS block and public IPv4 address becomes permanently free when EC2 enters stopped state.
- Delete all network accounts without verifying whether any address or volume is still needed.
- Add a Savings Plan commitment for future compute use and assume its discount automatically eliminates unattached EBS and public IPv4 charges.
- Inventory ancillary billable resources such as EBS volumes, snapshots and public IPv4 allocations and remove confirmed unnecessary items.
Explanation
Correct Answer
D:Stopping an instance can stop compute billing but does not automatically eliminate related storage or address charges.
Incorrect Answers
A:Storage/address services may bill independently of instance CPU state.
B:Mass-deleting accounts, addresses or volumes without evaluating live dependencies can break production traffic and remove recovery assets. A safe cost cleanup requires resource attribution and an approved dependency-aware retirement plan.
C:Compute commitments discount covered compute usage, not separate EBS storage or public IPv4 inventory. The unused ancillary resources still accrue charges until addressed.
Question 909
A JVM analytics server spends little CPU time but its 150-GiB heap frequently spills to disk on a general-purpose instance. Extra vCPUs remain idle. Which right-sizing direction matches the bottleneck?
- Evaluate a memory-optimized family with enough RAM rather than paying for unused CPU capacity.
- Choose an even higher-core compute-optimized instance with less RAM because CPU count always reduces swapping.
- Use a GPU-only instance solely because JVM garbage collection allocates objects.
- Reduce DB query timeout to zero instead of addressing the host's memory shortage.
Explanation
Correct Answer
A:High resident memory demand favors a larger memory-per-vCPU ratio when CPU is not saturated.
Incorrect Answers
B:An application that swaps because it lacks RAM will not necessarily improve by switching to more CPU cores with even less memory. Memory-optimized capacity or an application memory reduction addresses the measured bottleneck more directly.
C:General JVM heap pressure does not establish GPU acceleration needs.
D:A timeout setting cannot create resident memory capacity.
Question 910
A media company runs CPU-intensive transcoding jobs around the clock, averaging 90% CPU over each 24-hour period. Finance proposes replacing the stable-performance EC2 fleet with lower-base-price T3 Unlimited instances because it assumes “Unlimited” means all sustained CPU is included in that hourly price. The jobs cannot tolerate CPU-credit throttling. How should the architect evaluate this proposal?
- Adopt T3 Unlimited immediately; all CPU use above the T3 baseline is free once Unlimited mode is enabled.
- Model the proposed T3 baseline and billable surplus CPU credits over the sustained utilization profile, then benchmark total price and throughput against an appropriate nonburstable instance.
- Compare only the base EC2 per-hour price, treating the initial Unlimited CPU-credit surplus as permanently available for the sustained 90% workload.
- Choose T3 Standard mode instead because it guarantees sustained 90% CPU even after all accrued credits are depleted.
Explanation
Correct Answer
B:The baseline and rolling usage determine whether a burstable Unlimited instance incurs billed surplus CPU credit consumption. Comparing all-in charges and performance to steady-performance compute is the defensible way to assess the supposed savings.
Incorrect Answers
A:Unlimited mode can sustain CPU above the baseline, but surplus CPU credits can become billable when the rolling 24-hour average exceeds baseline. For around-the-clock 90% CPU, evaluate those charges and observed price/performance against a nonburstable instance before declaring savings.
C:The base hourly T3 price and initial credit availability do not capture billable excess CPU on a persistently busy Unlimited instance. The rolling 24-hour baseline and surplus-credit charging metrics must be included in the total-cost comparison.
D:Standard mode uses earned and accrued credits for bursts and returns to its baseline performance when credits are depleted. It cannot guarantee sustained high CPU beyond the baseline for this always-busy workload.
