Amazon AWS Certified Solutions Architect Associate SAA-C03 Elastic Compute Sizing and Scaling Signals Practice Test

 

Topic 12 covers elastic compute sizing and scaling signals for the AWS Certified Solutions Architect – Associate certification. These original practice questions apply the verified SAA-C03 objectives to practical decisions and troubleshooting. Select one answer unless a fixed number is requested. For broader preparation, visit the AWS Certified Solutions Architect Associate SAA-C03 Exam Dumps page. Each option includes an explanation of the relevant behavior and scenario constraints.

Question 1

A JVM service has low CPU but repeatedly runs out of memory under normal load. Which EC2 family characteristic should be prioritized?

1. Use a compute-optimized EC2 instance family.
2. Use a memory-optimized EC2 instance family.
3. Benchmark compatible instance architectures and choose the one that meets performance requirements with the desired price/performance.
4. Use a storage-optimized instance when local high-throughput/IOPS storage is the dominant requirement.
5. Use an accelerated-computing instance with the required GPU or accelerator.

Correct Answer: B

 

Correct Answer

Answer B is correct because Memory-optimized instances provide a higher memory-to-vCPU ratio for workloads constrained by RAM. This directly meets the decisive requirement: memory-optimized instance.

Incorrect Answers

Answer A is incorrect because Compute-optimized instances provide a higher compute profile for CPU-bound workloads. It may be useful elsewhere, but it does not meet the decisive requirement here: memory-optimized instance.

Answer C is incorrect because Architecture choice should follow software compatibility and measured workload behavior rather than an assumption that one instruction set is universally faster. It may be useful elsewhere, but it does not meet the decisive requirement here: memory-optimized instance.

Answer D is incorrect because Storage-optimized instances are designed for workloads that need high local storage performance. It may be useful elsewhere, but it does not meet the decisive requirement here: memory-optimized instance.

Answer E is incorrect because Accelerated instance families expose hardware accelerators for workloads that require GPU or specialized processing. It may be useful elsewhere, but it does not meet the decisive requirement here: memory-optimized instance.

 

Question 2

A scientific service is CPU-bound and keeps memory mostly idle. Which family characteristic is most appropriate?

1. Use an accelerated-computing instance with the required GPU or accelerator.
2. Use a memory-optimized EC2 instance family.
3. Use a storage-optimized instance when local high-throughput/IOPS storage is the dominant requirement.
4. Use a compute-optimized EC2 instance family.
5. Benchmark compatible instance architectures and choose the one that meets performance requirements with the desired price/performance.

Correct Answer: D

 

Correct Answer

Answer D is correct because Compute-optimized instances provide a higher compute profile for CPU-bound workloads. This directly meets the decisive requirement: compute-optimized instance.

Incorrect Answers

Answer A is incorrect because Accelerated instance families expose hardware accelerators for workloads that require GPU or specialized processing. It may be useful elsewhere, but it does not meet the decisive requirement here: compute-optimized instance.

Answer B is incorrect because Memory-optimized instances provide a higher memory-to-vCPU ratio for workloads constrained by RAM. It may be useful elsewhere, but it does not meet the decisive requirement here: compute-optimized instance.

Answer C is incorrect because Storage-optimized instances are designed for workloads that need high local storage performance. It may be useful elsewhere, but it does not meet the decisive requirement here: compute-optimized instance.

Answer E is incorrect because Architecture choice should follow software compatibility and measured workload behavior rather than an assumption that one instruction set is universally faster. It may be useful elsewhere, but it does not meet the decisive requirement here: compute-optimized instance.

 

Question 3

A machine-learning inference application requires GPU acceleration. Which instance characteristic is required?

1. Use a storage-optimized instance when local high-throughput/IOPS storage is the dominant requirement.
2. Use an accelerated-computing instance with the required GPU or accelerator.
3. Use a memory-optimized EC2 instance family.
4. Benchmark compatible instance architectures and choose the one that meets performance requirements with the desired price/performance.
5. Use a compute-optimized EC2 instance family.

Correct Answer: B

 

Correct Answer

Answer B is correct because Accelerated instance families expose hardware accelerators for workloads that require GPU or specialized processing. This directly meets the decisive requirement: accelerated computing instance.

Incorrect Answers

Answer A is incorrect because Storage-optimized instances are designed for workloads that need high local storage performance. It may be useful elsewhere, but it does not meet the decisive requirement here: accelerated computing instance.

Answer C is incorrect because Memory-optimized instances provide a higher memory-to-vCPU ratio for workloads constrained by RAM. It may be useful elsewhere, but it does not meet the decisive requirement here: accelerated computing instance.

Answer D is incorrect because Architecture choice should follow software compatibility and measured workload behavior rather than an assumption that one instruction set is universally faster. It may be useful elsewhere, but it does not meet the decisive requirement here: accelerated computing instance.

Answer E is incorrect because Compute-optimized instances provide a higher compute profile for CPU-bound workloads. It may be useful elsewhere, but it does not meet the decisive requirement here: accelerated computing instance.

 

Question 4

A data-processing node is limited by local NVMe throughput rather than CPU. Which family characteristic should be evaluated?

1. Use a compute-optimized EC2 instance family.
2. Benchmark compatible instance architectures and choose the one that meets performance requirements with the desired price/performance.
3. Use a memory-optimized EC2 instance family.
4. Use a storage-optimized instance when local high-throughput/IOPS storage is the dominant requirement.
5. Use an accelerated-computing instance with the required GPU or accelerator.

Correct Answer: D

 

Correct Answer

Answer D is correct because Storage-optimized instances are designed for workloads that need high local storage performance. This directly meets the decisive requirement: storage-optimized instance.

Incorrect Answers

Answer A is incorrect because Compute-optimized instances provide a higher compute profile for CPU-bound workloads. It may be useful elsewhere, but it does not meet the decisive requirement here: storage-optimized instance.

Answer B is incorrect because Architecture choice should follow software compatibility and measured workload behavior rather than an assumption that one instruction set is universally faster. It may be useful elsewhere, but it does not meet the decisive requirement here: storage-optimized instance.

Answer C is incorrect because Memory-optimized instances provide a higher memory-to-vCPU ratio for workloads constrained by RAM. It may be useful elsewhere, but it does not meet the decisive requirement here: storage-optimized instance.

Answer E is incorrect because Accelerated instance families expose hardware accelerators for workloads that require GPU or specialized processing. It may be useful elsewhere, but it does not meet the decisive requirement here: storage-optimized instance.

 

Question 5

An application can run on either x86 or Arm and benchmarks show equal functionality. Which architecture decision should be based on measured price/performance and software compatibility?

1. Use a memory-optimized EC2 instance family.
2. Use an accelerated-computing instance with the required GPU or accelerator.
3. Benchmark compatible instance architectures and choose the one that meets performance requirements with the desired price/performance.
4. Use a compute-optimized EC2 instance family.
5. Use a storage-optimized instance when local high-throughput/IOPS storage is the dominant requirement.

Correct Answer: C

 

Correct Answer

Answer C is correct because Architecture choice should follow software compatibility and measured workload behavior rather than an assumption that one instruction set is universally faster. This directly meets the decisive requirement: choose architecture using measured workload performance.

Incorrect Answers

Answer A is incorrect because Memory-optimized instances provide a higher memory-to-vCPU ratio for workloads constrained by RAM. It may be useful elsewhere, but it does not meet the decisive requirement here: choose architecture using measured workload performance.

Answer B is incorrect because Accelerated instance families expose hardware accelerators for workloads that require GPU or specialized processing. It may be useful elsewhere, but it does not meet the decisive requirement here: choose architecture using measured workload performance.

Answer D is incorrect because Compute-optimized instances provide a higher compute profile for CPU-bound workloads. It may be useful elsewhere, but it does not meet the decisive requirement here: choose architecture using measured workload performance.

Answer E is incorrect because Storage-optimized instances are designed for workloads that need high local storage performance. It may be useful elsewhere, but it does not meet the decisive requirement here: choose architecture using measured workload performance.

 

Question 6

A stateless CPU-bound web fleet slows whenever average CPU exceeds 65%. Which scaling metric is a reasonable target-tracking signal?

1. Use the load balancer requests-per-target metric.
2. Use average CPU utilization as the target-tracking metric.
3. Scale from queue backlog, such as messages per worker or oldest-message age.
4. Publish and scale on a custom CloudWatch metric that represents queued business work.
5. Use scheduled scaling to add capacity before the known event.

Correct Answer: B

 

Correct Answer

Answer B is correct because CPU is appropriate when measured application load is directly correlated with CPU saturation. This directly meets the decisive requirement: average CPU utilization.

Incorrect Answers

Answer A is incorrect because Requests per target can reflect service pressure more directly than CPU for some request-driven fleets. It may be useful elsewhere, but it does not meet the decisive requirement here: average CPU utilization.

Answer C is incorrect because Backlog metrics directly represent work waiting for the worker fleet and can scale consumers before user-visible delay becomes excessive. It may be useful elsewhere, but it does not meet the decisive requirement here: average CPU utilization.

Answer D is incorrect because A custom demand metric is preferable when infrastructure utilization does not accurately represent the workload bottleneck. It may be useful elsewhere, but it does not meet the decisive requirement here: average CPU utilization.

Answer E is incorrect because Scheduled scaling handles predictable recurring demand by changing desired capacity ahead of the expected burst. It may be useful elsewhere, but it does not meet the decisive requirement here: average CPU utilization.

 

Question 7

Workers consume messages from SQS and user delay is tied to the age of the oldest message. Which scaling signal best represents unmet work?

1. Use the load balancer requests-per-target metric.
2. Use scheduled scaling to add capacity before the known event.
3. Publish and scale on a custom CloudWatch metric that represents queued business work.
4. Use average CPU utilization as the target-tracking metric.
5. Scale from queue backlog, such as messages per worker or oldest-message age.

Correct Answer: E

 

Correct Answer

Answer E is correct because Backlog metrics directly represent work waiting for the worker fleet and can scale consumers before user-visible delay becomes excessive. This directly meets the decisive requirement: queue depth or oldest-message age.

Incorrect Answers

Answer A is incorrect because Requests per target can reflect service pressure more directly than CPU for some request-driven fleets. It may be useful elsewhere, but it does not meet the decisive requirement here: queue depth or oldest-message age.

Answer B is incorrect because Scheduled scaling handles predictable recurring demand by changing desired capacity ahead of the expected burst. It may be useful elsewhere, but it does not meet the decisive requirement here: queue depth or oldest-message age.

Answer C is incorrect because A custom demand metric is preferable when infrastructure utilization does not accurately represent the workload bottleneck. It may be useful elsewhere, but it does not meet the decisive requirement here: queue depth or oldest-message age.

Answer D is incorrect because CPU is appropriate when measured application load is directly correlated with CPU saturation. It may be useful elsewhere, but it does not meet the decisive requirement here: queue depth or oldest-message age.

 

Question 8

An ALB-backed service has variable request complexity, but requests per target strongly tracks saturation. Which metric is appropriate for target tracking?

1. Scale from queue backlog, such as messages per worker or oldest-message age.
2. Use scheduled scaling to add capacity before the known event.
3. Publish and scale on a custom CloudWatch metric that represents queued business work.
4. Use the load balancer requests-per-target metric.
5. Use average CPU utilization as the target-tracking metric.

Correct Answer: D

 

Correct Answer

Answer D is correct because Requests per target can reflect service pressure more directly than CPU for some request-driven fleets. This directly meets the decisive requirement: ALBRequestCountPerTarget.

Incorrect Answers

Answer A is incorrect because Backlog metrics directly represent work waiting for the worker fleet and can scale consumers before user-visible delay becomes excessive. It may be useful elsewhere, but it does not meet the decisive requirement here: ALBRequestCountPerTarget.

Answer B is incorrect because Scheduled scaling handles predictable recurring demand by changing desired capacity ahead of the expected burst. It may be useful elsewhere, but it does not meet the decisive requirement here: ALBRequestCountPerTarget.

Answer C is incorrect because A custom demand metric is preferable when infrastructure utilization does not accurately represent the workload bottleneck. It may be useful elsewhere, but it does not meet the decisive requirement here: ALBRequestCountPerTarget.

Answer E is incorrect because CPU is appropriate when measured application load is directly correlated with CPU saturation. It may be useful elsewhere, but it does not meet the decisive requirement here: ALBRequestCountPerTarget.

 

Question 9

A payroll service always receives a known burst at 09:00 on the first business day of the month. Which scaling method can add capacity before demand begins?

1. Use scheduled scaling to add capacity before the known event.
2. Publish and scale on a custom CloudWatch metric that represents queued business work.
3. Use the load balancer requests-per-target metric.
4. Scale from queue backlog, such as messages per worker or oldest-message age.
5. Use average CPU utilization as the target-tracking metric.

Correct Answer: A

 

Correct Answer

Answer A is correct because Scheduled scaling handles predictable recurring demand by changing desired capacity ahead of the expected burst. This directly meets the decisive requirement: scheduled scaling.

Incorrect Answers

Answer B is incorrect because A custom demand metric is preferable when infrastructure utilization does not accurately represent the workload bottleneck. It may be useful elsewhere, but it does not meet the decisive requirement here: scheduled scaling.

Answer C is incorrect because Requests per target can reflect service pressure more directly than CPU for some request-driven fleets. It may be useful elsewhere, but it does not meet the decisive requirement here: scheduled scaling.

Answer D is incorrect because Backlog metrics directly represent work waiting for the worker fleet and can scale consumers before user-visible delay becomes excessive. It may be useful elsewhere, but it does not meet the decisive requirement here: scheduled scaling.

Answer E is incorrect because CPU is appropriate when measured application load is directly correlated with CPU saturation. It may be useful elsewhere, but it does not meet the decisive requirement here: scheduled scaling.

 

Question 10

A transcoding service is constrained by minutes of video waiting rather than CPU utilization. Which scaling signal should be used?

1. Use average CPU utilization as the target-tracking metric.
2. Publish and scale on a custom CloudWatch metric that represents queued business work.
3. Scale from queue backlog, such as messages per worker or oldest-message age.
4. Use scheduled scaling to add capacity before the known event.
5. Use the load balancer requests-per-target metric.

Correct Answer: B

 

Correct Answer

Answer B is correct because A custom demand metric is preferable when infrastructure utilization does not accurately represent the workload bottleneck. This directly meets the decisive requirement: custom backlog-work metric.

Incorrect Answers

Answer A is incorrect because CPU is appropriate when measured application load is directly correlated with CPU saturation. It may be useful elsewhere, but it does not meet the decisive requirement here: custom backlog-work metric.

Answer C is incorrect because Backlog metrics directly represent work waiting for the worker fleet and can scale consumers before user-visible delay becomes excessive. It may be useful elsewhere, but it does not meet the decisive requirement here: custom backlog-work metric.

Answer D is incorrect because Scheduled scaling handles predictable recurring demand by changing desired capacity ahead of the expected burst. It may be useful elsewhere, but it does not meet the decisive requirement here: custom backlog-work metric.

Answer E is incorrect because Requests per target can reflect service pressure more directly than CPU for some request-driven fleets. It may be useful elsewhere, but it does not meet the decisive requirement here: custom backlog-work metric.

 

Question 11

New EC2 instances take eight minutes before they can serve traffic. Scaling decisions treat them as fully ready immediately and trigger repeated scale-out. Which setting should reflect startup time?

1. Use Auto Scaling health checks and replacement of unhealthy instances.
2. Set the Auto Scaling group minimum capacity to four.
3. Use connection draining/job checkpointing or scale-in protection so in-flight non-interruptible work is not terminated abruptly.
4. Configure appropriate instance warmup for the scaling policy.
5. Tune stabilization/cooldown and choose a metric/target that avoids rapid oscillation.

Correct Answer: D

 

Correct Answer

Answer D is correct because Warmup prevents new instances from distorting scaling metrics before they have finished initialization and can carry normal load. This directly meets the decisive requirement: instance warmup.

Incorrect Answers

Answer A is incorrect because A self-healing group terminates and replaces capacity that no longer passes configured health evaluation. It may be useful elsewhere, but it does not meet the decisive requirement here: instance warmup.

Answer B is incorrect because Minimum capacity defines the floor below which the group will not scale in. It may be useful elsewhere, but it does not meet the decisive requirement here: instance warmup.

Answer C is incorrect because Safe scale-in must account for long-running work rather than treating every instance as disposable at any moment. It may be useful elsewhere, but it does not meet the decisive requirement here: instance warmup.

Answer E is incorrect because Scaling should not react to every short fluctuation when instance replacement itself is expensive or slow. It may be useful elsewhere, but it does not meet the decisive requirement here: instance warmup.

 

Question 12

A target-tracking group scales in and out rapidly around a threshold, causing frequent instance replacement. Which tuning goal is appropriate?

1. Use connection draining/job checkpointing or scale-in protection so in-flight non-interruptible work is not terminated abruptly.
2. Tune stabilization/cooldown and choose a metric/target that avoids rapid oscillation.
3. Use Auto Scaling health checks and replacement of unhealthy instances.
4. Configure appropriate instance warmup for the scaling policy.
5. Set the Auto Scaling group minimum capacity to four.

Correct Answer: B

 

Correct Answer

Answer B is correct because Scaling should not react to every short fluctuation when instance replacement itself is expensive or slow. This directly meets the decisive requirement: stabilize scale-in behavior/cooldown.

Incorrect Answers

Answer A is incorrect because Safe scale-in must account for long-running work rather than treating every instance as disposable at any moment. It may be useful elsewhere, but it does not meet the decisive requirement here: stabilize scale-in behavior/cooldown.

Answer C is incorrect because A self-healing group terminates and replaces capacity that no longer passes configured health evaluation. It may be useful elsewhere, but it does not meet the decisive requirement here: stabilize scale-in behavior/cooldown.

Answer D is incorrect because Warmup prevents new instances from distorting scaling metrics before they have finished initialization and can carry normal load. It may be useful elsewhere, but it does not meet the decisive requirement here: stabilize scale-in behavior/cooldown.

Answer E is incorrect because Minimum capacity defines the floor below which the group will not scale in. It may be useful elsewhere, but it does not meet the decisive requirement here: stabilize scale-in behavior/cooldown.

 

Question 13

A production service must always keep at least four instances available even at night. Which Auto Scaling parameter provides this floor?

1. Configure appropriate instance warmup for the scaling policy.
2. Tune stabilization/cooldown and choose a metric/target that avoids rapid oscillation.
3. Set the Auto Scaling group minimum capacity to four.
4. Use connection draining/job checkpointing or scale-in protection so in-flight non-interruptible work is not terminated abruptly.
5. Use Auto Scaling health checks and replacement of unhealthy instances.

Correct Answer: C

 

Correct Answer

Answer C is correct because Minimum capacity defines the floor below which the group will not scale in. This directly meets the decisive requirement: minimum capacity.

Incorrect Answers

Answer A is incorrect because Warmup prevents new instances from distorting scaling metrics before they have finished initialization and can carry normal load. It may be useful elsewhere, but it does not meet the decisive requirement here: minimum capacity.

Answer B is incorrect because Scaling should not react to every short fluctuation when instance replacement itself is expensive or slow. It may be useful elsewhere, but it does not meet the decisive requirement here: minimum capacity.

Answer D is incorrect because Safe scale-in must account for long-running work rather than treating every instance as disposable at any moment. It may be useful elsewhere, but it does not meet the decisive requirement here: minimum capacity.

Answer E is incorrect because A self-healing group terminates and replaces capacity that no longer passes configured health evaluation. It may be useful elsewhere, but it does not meet the decisive requirement here: minimum capacity.

 

Question 14

A worker is processing a non-interruptible 30-minute job when scale-in selects it for termination. Which design should protect work?

1. Set the Auto Scaling group minimum capacity to four.
2. Configure appropriate instance warmup for the scaling policy.
3. Use connection draining/job checkpointing or scale-in protection so in-flight non-interruptible work is not terminated abruptly.
4. Tune stabilization/cooldown and choose a metric/target that avoids rapid oscillation.
5. Use Auto Scaling health checks and replacement of unhealthy instances.

Correct Answer: C

 

Correct Answer

Answer C is correct because Safe scale-in must account for long-running work rather than treating every instance as disposable at any moment. This directly meets the decisive requirement: drain/protect instance or externalize job state.

Incorrect Answers

Answer A is incorrect because Minimum capacity defines the floor below which the group will not scale in. It may be useful elsewhere, but it does not meet the decisive requirement here: drain/protect instance or externalize job state.

Answer B is incorrect because Warmup prevents new instances from distorting scaling metrics before they have finished initialization and can carry normal load. It may be useful elsewhere, but it does not meet the decisive requirement here: drain/protect instance or externalize job state.

Answer D is incorrect because Scaling should not react to every short fluctuation when instance replacement itself is expensive or slow. It may be useful elsewhere, but it does not meet the decisive requirement here: drain/protect instance or externalize job state.

Answer E is incorrect because A self-healing group terminates and replaces capacity that no longer passes configured health evaluation. It may be useful elsewhere, but it does not meet the decisive requirement here: drain/protect instance or externalize job state.

 

Question 15

An instance repeatedly fails the application health check but remains in the group. Which mechanism should replace unhealthy capacity?

1. Use connection draining/job checkpointing or scale-in protection so in-flight non-interruptible work is not terminated abruptly.
2. Tune stabilization/cooldown and choose a metric/target that avoids rapid oscillation.
3. Configure appropriate instance warmup for the scaling policy.
4. Set the Auto Scaling group minimum capacity to four.
5. Use Auto Scaling health checks and replacement of unhealthy instances.

Correct Answer: E

 

Correct Answer

Answer E is correct because A self-healing group terminates and replaces capacity that no longer passes configured health evaluation. This directly meets the decisive requirement: Auto Scaling health replacement.

Incorrect Answers

Answer A is incorrect because Safe scale-in must account for long-running work rather than treating every instance as disposable at any moment. It may be useful elsewhere, but it does not meet the decisive requirement here: Auto Scaling health replacement.

Answer B is incorrect because Scaling should not react to every short fluctuation when instance replacement itself is expensive or slow. It may be useful elsewhere, but it does not meet the decisive requirement here: Auto Scaling health replacement.

Answer C is incorrect because Warmup prevents new instances from distorting scaling metrics before they have finished initialization and can carry normal load. It may be useful elsewhere, but it does not meet the decisive requirement here: Auto Scaling health replacement.

Answer D is incorrect because Minimum capacity defines the floor below which the group will not scale in. It may be useful elsewhere, but it does not meet the decisive requirement here: Auto Scaling health replacement.

 

Question 16

A Lambda function is CPU-bound and completes much faster when more memory is configured. Which tuning fact should guide testing?

1. Limit concurrency or add a queue/rate limiter so Lambda cannot exceed the downstream service limit.
2. Benchmark Lambda memory settings because CPU and other resources scale with the configured memory.
3. Set reserved concurrency to cap the function’s maximum concurrent executions.
4. Use provisioned concurrency for the latency-sensitive function.
5. Move the long task to a compute service suited to longer execution or decompose it into shorter workflow steps.

Correct Answer: B

 

Correct Answer

Answer B is correct because Lambda performance tuning should be measured because increasing memory can also increase available CPU and reduce runtime. This directly meets the decisive requirement: benchmark memory allocation.

Incorrect Answers

Answer A is incorrect because Serverless scaling must be constrained by the slowest protected dependency rather than by Lambda capacity alone. It may be useful elsewhere, but it does not meet the decisive requirement here: benchmark memory allocation.

Answer C is incorrect because Reserved concurrency can place a hard ceiling on one function and protect a downstream dependency from excessive parallel calls. It may be useful elsewhere, but it does not meet the decisive requirement here: benchmark memory allocation.

Answer D is incorrect because Provisioned concurrency keeps initialized execution environments ready to reduce cold-start latency for configured capacity. It may be useful elsewhere, but it does not meet the decisive requirement here: benchmark memory allocation.

Answer E is incorrect because A workload that exceeds Lambda duration limits needs a different execution model rather than repeated timeouts. It may be useful elsewhere, but it does not meet the decisive requirement here: benchmark memory allocation.

 

Question 17

A public function must never allow more than 100 concurrent database connections. Which Lambda control can cap function concurrency?

1. Set reserved concurrency to cap the function’s maximum concurrent executions.
2. Benchmark Lambda memory settings because CPU and other resources scale with the configured memory.
3. Move the long task to a compute service suited to longer execution or decompose it into shorter workflow steps.
4. Use provisioned concurrency for the latency-sensitive function.
5. Limit concurrency or add a queue/rate limiter so Lambda cannot exceed the downstream service limit.

Correct Answer: A

 

Correct Answer

Answer A is correct because Reserved concurrency can place a hard ceiling on one function and protect a downstream dependency from excessive parallel calls. This directly meets the decisive requirement: reserved concurrency.

Incorrect Answers

Answer B is incorrect because Lambda performance tuning should be measured because increasing memory can also increase available CPU and reduce runtime. It may be useful elsewhere, but it does not meet the decisive requirement here: reserved concurrency.

Answer C is incorrect because A workload that exceeds Lambda duration limits needs a different execution model rather than repeated timeouts. It may be useful elsewhere, but it does not meet the decisive requirement here: reserved concurrency.

Answer D is incorrect because Provisioned concurrency keeps initialized execution environments ready to reduce cold-start latency for configured capacity. It may be useful elsewhere, but it does not meet the decisive requirement here: reserved concurrency.

Answer E is incorrect because Serverless scaling must be constrained by the slowest protected dependency rather than by Lambda capacity alone. It may be useful elsewhere, but it does not meet the decisive requirement here: reserved concurrency.

 

Question 18

A latency-sensitive Lambda API experiences cold-start delay during sudden traffic. Which feature can keep initialized execution environments ready?

1. Limit concurrency or add a queue/rate limiter so Lambda cannot exceed the downstream service limit.
2. Use provisioned concurrency for the latency-sensitive function.
3. Benchmark Lambda memory settings because CPU and other resources scale with the configured memory.
4. Set reserved concurrency to cap the function’s maximum concurrent executions.
5. Move the long task to a compute service suited to longer execution or decompose it into shorter workflow steps.

Correct Answer: B

 

Correct Answer

Answer B is correct because Provisioned concurrency keeps initialized execution environments ready to reduce cold-start latency for configured capacity. This directly meets the decisive requirement: provisioned concurrency.

Incorrect Answers

Answer A is incorrect because Serverless scaling must be constrained by the slowest protected dependency rather than by Lambda capacity alone. It may be useful elsewhere, but it does not meet the decisive requirement here: provisioned concurrency.

Answer C is incorrect because Lambda performance tuning should be measured because increasing memory can also increase available CPU and reduce runtime. It may be useful elsewhere, but it does not meet the decisive requirement here: provisioned concurrency.

Answer D is incorrect because Reserved concurrency can place a hard ceiling on one function and protect a downstream dependency from excessive parallel calls. It may be useful elsewhere, but it does not meet the decisive requirement here: provisioned concurrency.

Answer E is incorrect because A workload that exceeds Lambda duration limits needs a different execution model rather than repeated timeouts. It may be useful elsewhere, but it does not meet the decisive requirement here: provisioned concurrency.

 

Question 19

A Lambda consumer can process thousands of messages per second, but a downstream API allows only 50 requests per second. Which design must coordinate scaling?

1. Set reserved concurrency to cap the function’s maximum concurrent executions.
2. Move the long task to a compute service suited to longer execution or decompose it into shorter workflow steps.
3. Use provisioned concurrency for the latency-sensitive function.
4. Benchmark Lambda memory settings because CPU and other resources scale with the configured memory.
5. Limit concurrency or add a queue/rate limiter so Lambda cannot exceed the downstream service limit.

Correct Answer: E

 

Correct Answer

Answer E is correct because Serverless scaling must be constrained by the slowest protected dependency rather than by Lambda capacity alone. This directly meets the decisive requirement: bound concurrency/request rate.

Incorrect Answers

Answer A is incorrect because Reserved concurrency can place a hard ceiling on one function and protect a downstream dependency from excessive parallel calls. It may be useful elsewhere, but it does not meet the decisive requirement here: bound concurrency/request rate.

Answer B is incorrect because A workload that exceeds Lambda duration limits needs a different execution model rather than repeated timeouts. It may be useful elsewhere, but it does not meet the decisive requirement here: bound concurrency/request rate.

Answer C is incorrect because Provisioned concurrency keeps initialized execution environments ready to reduce cold-start latency for configured capacity. It may be useful elsewhere, but it does not meet the decisive requirement here: bound concurrency/request rate.

Answer D is incorrect because Lambda performance tuning should be measured because increasing memory can also increase available CPU and reduce runtime. It may be useful elsewhere, but it does not meet the decisive requirement here: bound concurrency/request rate.

 

Question 20

A task routinely needs longer than the maximum Lambda execution duration. Which architecture change is required?

1. Use provisioned concurrency for the latency-sensitive function.
2. Move the long task to a compute service suited to longer execution or decompose it into shorter workflow steps.
3. Limit concurrency or add a queue/rate limiter so Lambda cannot exceed the downstream service limit.
4. Set reserved concurrency to cap the function’s maximum concurrent executions.
5. Benchmark Lambda memory settings because CPU and other resources scale with the configured memory.

Correct Answer: B

 

Correct Answer

Answer B is correct because A workload that exceeds Lambda duration limits needs a different execution model rather than repeated timeouts. This directly meets the decisive requirement: move to container/compute or split workflow.

Incorrect Answers

Answer A is incorrect because Provisioned concurrency keeps initialized execution environments ready to reduce cold-start latency for configured capacity. It may be useful elsewhere, but it does not meet the decisive requirement here: move to container/compute or split workflow.

Answer C is incorrect because Serverless scaling must be constrained by the slowest protected dependency rather than by Lambda capacity alone. It may be useful elsewhere, but it does not meet the decisive requirement here: move to container/compute or split workflow.

Answer D is incorrect because Reserved concurrency can place a hard ceiling on one function and protect a downstream dependency from excessive parallel calls. It may be useful elsewhere, but it does not meet the decisive requirement here: move to container/compute or split workflow.

Answer E is incorrect because Lambda performance tuning should be measured because increasing memory can also increase available CPU and reduce runtime. It may be useful elsewhere, but it does not meet the decisive requirement here: move to container/compute or split workflow.

 

Question 21

An ECS task is killed for exceeding its memory limit even though the cluster has free memory. Which setting should be corrected?

1. Use ECS Service Auto Scaling on an appropriate service metric.
2. Increase or correct the task CPU/memory definition to match measured requirements.
3. Use the ECS daemon scheduling strategy.
4. Use an ECS capacity provider with managed scaling.
5. Provide compatible GPU-capable container instances and matching placement/resource requirements.

Correct Answer: B

 

Correct Answer

Answer B is correct because Task-level resource limits can terminate a container even when unused cluster capacity exists outside that task allocation. This directly meets the decisive requirement: task resource definition.

Incorrect Answers

Answer A is incorrect because Service Auto Scaling changes desired task count independently of the underlying cluster capacity mechanism. It may be useful elsewhere, but it does not meet the decisive requirement here: task resource definition.

Answer C is incorrect because Daemon scheduling places one copy of a task on each eligible container instance and is suited to host-level agents. It may be useful elsewhere, but it does not meet the decisive requirement here: task resource definition.

Answer D is incorrect because Capacity providers can coordinate EC2 Auto Scaling group capacity with ECS task placement demand. It may be useful elsewhere, but it does not meet the decisive requirement here: task resource definition.

Answer E is incorrect because Tasks cannot be scheduled successfully when the cluster capacity does not satisfy their declared resource needs. It may be useful elsewhere, but it does not meet the decisive requirement here: task resource definition.

 

Question 22

An ECS service wants the cluster to add EC2 instances automatically when tasks cannot be placed. Which feature coordinates cluster capacity?

1. Provide compatible GPU-capable container instances and matching placement/resource requirements.
2. Increase or correct the task CPU/memory definition to match measured requirements.
3. Use the ECS daemon scheduling strategy.
4. Use ECS Service Auto Scaling on an appropriate service metric.
5. Use an ECS capacity provider with managed scaling.

Correct Answer: E

 

Correct Answer

Answer E is correct because Capacity providers can coordinate EC2 Auto Scaling group capacity with ECS task placement demand. This directly meets the decisive requirement: capacity provider managed scaling.

Incorrect Answers

Answer A is incorrect because Tasks cannot be scheduled successfully when the cluster capacity does not satisfy their declared resource needs. It may be useful elsewhere, but it does not meet the decisive requirement here: capacity provider managed scaling.

Answer B is incorrect because Task-level resource limits can terminate a container even when unused cluster capacity exists outside that task allocation. It may be useful elsewhere, but it does not meet the decisive requirement here: capacity provider managed scaling.

Answer C is incorrect because Daemon scheduling places one copy of a task on each eligible container instance and is suited to host-level agents. It may be useful elsewhere, but it does not meet the decisive requirement here: capacity provider managed scaling.

Answer D is incorrect because Service Auto Scaling changes desired task count independently of the underlying cluster capacity mechanism. It may be useful elsewhere, but it does not meet the decisive requirement here: capacity provider managed scaling.

 

Question 23

A containerized API needs more tasks when average service CPU rises. Which feature scales the service task count?

1. Use an ECS capacity provider with managed scaling.
2. Use the ECS daemon scheduling strategy.
3. Increase or correct the task CPU/memory definition to match measured requirements.
4. Provide compatible GPU-capable container instances and matching placement/resource requirements.
5. Use ECS Service Auto Scaling on an appropriate service metric.

Correct Answer: E

 

Correct Answer

Answer E is correct because Service Auto Scaling changes desired task count independently of the underlying cluster capacity mechanism. This directly meets the decisive requirement: ECS Service Auto Scaling.

Incorrect Answers

Answer A is incorrect because Capacity providers can coordinate EC2 Auto Scaling group capacity with ECS task placement demand. It may be useful elsewhere, but it does not meet the decisive requirement here: ECS Service Auto Scaling.

Answer B is incorrect because Daemon scheduling places one copy of a task on each eligible container instance and is suited to host-level agents. It may be useful elsewhere, but it does not meet the decisive requirement here: ECS Service Auto Scaling.

Answer C is incorrect because Task-level resource limits can terminate a container even when unused cluster capacity exists outside that task allocation. It may be useful elsewhere, but it does not meet the decisive requirement here: ECS Service Auto Scaling.

Answer D is incorrect because Tasks cannot be scheduled successfully when the cluster capacity does not satisfy their declared resource needs. It may be useful elsewhere, but it does not meet the decisive requirement here: ECS Service Auto Scaling.

 

Question 24

Tasks require GPUs but the current capacity provider uses instance types with no GPU. Which scheduling issue must be corrected?

1. Use ECS Service Auto Scaling on an appropriate service metric.
2. Provide compatible GPU-capable container instances and matching placement/resource requirements.
3. Use the ECS daemon scheduling strategy.
4. Increase or correct the task CPU/memory definition to match measured requirements.
5. Use an ECS capacity provider with managed scaling.

Correct Answer: B

 

Correct Answer

Answer B is correct because Tasks cannot be scheduled successfully when the cluster capacity does not satisfy their declared resource needs. This directly meets the decisive requirement: capacity/placement compatibility.

Incorrect Answers

Answer A is incorrect because Service Auto Scaling changes desired task count independently of the underlying cluster capacity mechanism. It may be useful elsewhere, but it does not meet the decisive requirement here: capacity/placement compatibility.

Answer C is incorrect because Daemon scheduling places one copy of a task on each eligible container instance and is suited to host-level agents. It may be useful elsewhere, but it does not meet the decisive requirement here: capacity/placement compatibility.

Answer D is incorrect because Task-level resource limits can terminate a container even when unused cluster capacity exists outside that task allocation. It may be useful elsewhere, but it does not meet the decisive requirement here: capacity/placement compatibility.

Answer E is incorrect because Capacity providers can coordinate EC2 Auto Scaling group capacity with ECS task placement demand. It may be useful elsewhere, but it does not meet the decisive requirement here: capacity/placement compatibility.

 

Question 25

A service must place one task on every EC2 container instance for a local monitoring agent. Which ECS scheduling strategy fits?

1. Use the ECS daemon scheduling strategy.
2. Use ECS Service Auto Scaling on an appropriate service metric.
3. Use an ECS capacity provider with managed scaling.
4. Provide compatible GPU-capable container instances and matching placement/resource requirements.
5. Increase or correct the task CPU/memory definition to match measured requirements.

Correct Answer: A

 

Correct Answer

Answer A is correct because Daemon scheduling places one copy of a task on each eligible container instance and is suited to host-level agents. This directly meets the decisive requirement: daemon scheduling.

Incorrect Answers

Answer B is incorrect because Service Auto Scaling changes desired task count independently of the underlying cluster capacity mechanism. It may be useful elsewhere, but it does not meet the decisive requirement here: daemon scheduling.

Answer C is incorrect because Capacity providers can coordinate EC2 Auto Scaling group capacity with ECS task placement demand. It may be useful elsewhere, but it does not meet the decisive requirement here: daemon scheduling.

Answer D is incorrect because Tasks cannot be scheduled successfully when the cluster capacity does not satisfy their declared resource needs. It may be useful elsewhere, but it does not meet the decisive requirement here: daemon scheduling.

Answer E is incorrect because Task-level resource limits can terminate a container even when unused cluster capacity exists outside that task allocation. It may be useful elsewhere, but it does not meet the decisive requirement here: daemon scheduling.

 

Question 26

A simulation consists of 50,000 independent jobs with no ordering dependency. Which AWS service is designed to queue and schedule this batch workload?

1. Move eligible preprocessing closer to the data source using an appropriate edge or hybrid compute location.
2. Use Amazon EMR for the managed Spark/big-data processing cluster.
3. Express explicit job or workflow dependencies between processing stages.
4. Use AWS Batch to queue and schedule the independent jobs.
5. Use checkpointing together with interruption-tolerant capacity such as Spot where the workload can safely resume.

Correct Answer: D

 

Correct Answer

Answer D is correct because AWS Batch manages job queues and compute environments for large volumes of batch work. This directly meets the decisive requirement: AWS Batch.

Incorrect Answers

Answer A is incorrect because Processing near the source reduces WAN bandwidth and latency when shipping all raw data to a Region is inefficient. It may be useful elsewhere, but it does not meet the decisive requirement here: AWS Batch.

Answer B is incorrect because EMR is designed to run distributed data-processing frameworks such as Apache Spark on AWS. It may be useful elsewhere, but it does not meet the decisive requirement here: AWS Batch.

Answer C is incorrect because A dependency graph prevents downstream stages from starting before prerequisite jobs have completed. It may be useful elsewhere, but it does not meet the decisive requirement here: AWS Batch.

Answer E is incorrect because Checkpointing limits lost work when discounted interruptible capacity is reclaimed. It may be useful elsewhere, but it does not meet the decisive requirement here: AWS Batch.

 

Question 27

A workflow has three stages and stage 2 cannot start until every stage-1 job completes. Which design is required?

1. Express explicit job or workflow dependencies between processing stages.
2. Use checkpointing together with interruption-tolerant capacity such as Spot where the workload can safely resume.
3. Use AWS Batch to queue and schedule the independent jobs.
4. Move eligible preprocessing closer to the data source using an appropriate edge or hybrid compute location.
5. Use Amazon EMR for the managed Spark/big-data processing cluster.

Correct Answer: A

 

Correct Answer

Answer A is correct because A dependency graph prevents downstream stages from starting before prerequisite jobs have completed. This directly meets the decisive requirement: job/workflow dependencies.

Incorrect Answers

Answer B is incorrect because Checkpointing limits lost work when discounted interruptible capacity is reclaimed. It may be useful elsewhere, but it does not meet the decisive requirement here: job/workflow dependencies.

Answer C is incorrect because AWS Batch manages job queues and compute environments for large volumes of batch work. It may be useful elsewhere, but it does not meet the decisive requirement here: job/workflow dependencies.

Answer D is incorrect because Processing near the source reduces WAN bandwidth and latency when shipping all raw data to a Region is inefficient. It may be useful elsewhere, but it does not meet the decisive requirement here: job/workflow dependencies.

Answer E is incorrect because EMR is designed to run distributed data-processing frameworks such as Apache Spark on AWS. It may be useful elsewhere, but it does not meet the decisive requirement here: job/workflow dependencies.

 

Question 28

A Spark analytics job processes terabytes of data and the team wants a managed big-data cluster service. Which service should be used?

1. Move eligible preprocessing closer to the data source using an appropriate edge or hybrid compute location.
2. Use checkpointing together with interruption-tolerant capacity such as Spot where the workload can safely resume.
3. Use Amazon EMR for the managed Spark/big-data processing cluster.
4. Use AWS Batch to queue and schedule the independent jobs.
5. Express explicit job or workflow dependencies between processing stages.

Correct Answer: C

 

Correct Answer

Answer C is correct because EMR is designed to run distributed data-processing frameworks such as Apache Spark on AWS. This directly meets the decisive requirement: Amazon EMR.

Incorrect Answers

Answer A is incorrect because Processing near the source reduces WAN bandwidth and latency when shipping all raw data to a Region is inefficient. It may be useful elsewhere, but it does not meet the decisive requirement here: Amazon EMR.

Answer B is incorrect because Checkpointing limits lost work when discounted interruptible capacity is reclaimed. It may be useful elsewhere, but it does not meet the decisive requirement here: Amazon EMR.

Answer D is incorrect because AWS Batch manages job queues and compute environments for large volumes of batch work. It may be useful elsewhere, but it does not meet the decisive requirement here: Amazon EMR.

Answer E is incorrect because A dependency graph prevents downstream stages from starting before prerequisite jobs have completed. It may be useful elsewhere, but it does not meet the decisive requirement here: Amazon EMR.

 

Question 29

A long batch job can checkpoint intermediate results to S3. Spot interruption is acceptable if work can resume. Which design improves efficiency?

1. Express explicit job or workflow dependencies between processing stages.
2. Use AWS Batch to queue and schedule the independent jobs.
3. Use Amazon EMR for the managed Spark/big-data processing cluster.
4. Use checkpointing together with interruption-tolerant capacity such as Spot where the workload can safely resume.
5. Move eligible preprocessing closer to the data source using an appropriate edge or hybrid compute location.

Correct Answer: D

 

Correct Answer

Answer D is correct because Checkpointing limits lost work when discounted interruptible capacity is reclaimed. This directly meets the decisive requirement: use interruption-tolerant capacity and resume from checkpoint.

Incorrect Answers

Answer A is incorrect because A dependency graph prevents downstream stages from starting before prerequisite jobs have completed. It may be useful elsewhere, but it does not meet the decisive requirement here: use interruption-tolerant capacity and resume from checkpoint.

Answer B is incorrect because AWS Batch manages job queues and compute environments for large volumes of batch work. It may be useful elsewhere, but it does not meet the decisive requirement here: use interruption-tolerant capacity and resume from checkpoint.

Answer C is incorrect because EMR is designed to run distributed data-processing frameworks such as Apache Spark on AWS. It may be useful elsewhere, but it does not meet the decisive requirement here: use interruption-tolerant capacity and resume from checkpoint.

Answer E is incorrect because Processing near the source reduces WAN bandwidth and latency when shipping all raw data to a Region is inefficient. It may be useful elsewhere, but it does not meet the decisive requirement here: use interruption-tolerant capacity and resume from checkpoint.

 

Question 30

A factory must preprocess sensor data near the source before sending summarized data to the Region because the WAN is constrained. Which placement principle applies?

1. Use Amazon EMR for the managed Spark/big-data processing cluster.
2. Express explicit job or workflow dependencies between processing stages.
3. Move eligible preprocessing closer to the data source using an appropriate edge or hybrid compute location.
4. Use AWS Batch to queue and schedule the independent jobs.
5. Use checkpointing together with interruption-tolerant capacity such as Spot where the workload can safely resume.

Correct Answer: C

 

Correct Answer

Answer C is correct because Processing near the source reduces WAN bandwidth and latency when shipping all raw data to a Region is inefficient. This directly meets the decisive requirement: edge/local compute.

Incorrect Answers

Answer A is incorrect because EMR is designed to run distributed data-processing frameworks such as Apache Spark on AWS. It may be useful elsewhere, but it does not meet the decisive requirement here: edge/local compute.

Answer B is incorrect because A dependency graph prevents downstream stages from starting before prerequisite jobs have completed. It may be useful elsewhere, but it does not meet the decisive requirement here: edge/local compute.

Answer D is incorrect because AWS Batch manages job queues and compute environments for large volumes of batch work. It may be useful elsewhere, but it does not meet the decisive requirement here: edge/local compute.

Answer E is incorrect because Checkpointing limits lost work when discounted interruptible capacity is reclaimed. It may be useful elsewhere, but it does not meet the decisive requirement here: edge/local compute.

 

Question 31

A producer writes 10,000 messages per second while workers drain 2,000 per second. Web CPU is low. Which component must scale or be throttled?

1. Increase the relevant AWS service quota after validating the required scale.
2. Scale the worker/consumer tier or apply producer backpressure.
3. Scale or redesign the database tier because it is now the saturated dependency.
4. Choose a partition key or sharding strategy that distributes request load more evenly.
5. Choose instance/network configuration with sufficient network bandwidth or reduce transfer demand.

Correct Answer: B

 

Correct Answer

Answer B is correct because The backlog shows that consumer throughput, not the web tier, is below the arrival rate. This directly meets the decisive requirement: worker/consumer capacity.

Incorrect Answers

Answer A is incorrect because A hard account or service quota stops capacity growth even when scaling policies request more resources. It may be useful elsewhere, but it does not meet the decisive requirement here: worker/consumer capacity.

Answer C is incorrect because Adding application servers cannot improve throughput once a shared database writer is the limiting resource. It may be useful elsewhere, but it does not meet the decisive requirement here: worker/consumer capacity.

Answer D is incorrect because A hot partition limits throughput while leaving capacity on other partitions unused. It may be useful elsewhere, but it does not meet the decisive requirement here: worker/consumer capacity.

Answer E is incorrect because A saturated network interface can cap throughput even when compute utilization is low. It may be useful elsewhere, but it does not meet the decisive requirement here: worker/consumer capacity.

 

Question 32

An Auto Scaling group needs 200 instances but stops at 100 because the account limit is reached. What is the bottleneck?

1. Increase the relevant AWS service quota after validating the required scale.
2. Scale the worker/consumer tier or apply producer backpressure.
3. Choose instance/network configuration with sufficient network bandwidth or reduce transfer demand.
4. Choose a partition key or sharding strategy that distributes request load more evenly.
5. Scale or redesign the database tier because it is now the saturated dependency.

Correct Answer: A

 

Correct Answer

Answer A is correct because A hard account or service quota stops capacity growth even when scaling policies request more resources. This directly meets the decisive requirement: service quota.

Incorrect Answers

Answer B is incorrect because The backlog shows that consumer throughput, not the web tier, is below the arrival rate. It may be useful elsewhere, but it does not meet the decisive requirement here: service quota.

Answer C is incorrect because A saturated network interface can cap throughput even when compute utilization is low. It may be useful elsewhere, but it does not meet the decisive requirement here: service quota.

Answer D is incorrect because A hot partition limits throughput while leaving capacity on other partitions unused. It may be useful elsewhere, but it does not meet the decisive requirement here: service quota.

Answer E is incorrect because Adding application servers cannot improve throughput once a shared database writer is the limiting resource. It may be useful elsewhere, but it does not meet the decisive requirement here: service quota.

 

Question 33

Application servers scale out successfully but a single database writer is saturated and response time keeps rising. Which tier is now limiting throughput?

1. Scale the worker/consumer tier or apply producer backpressure.
2. Scale or redesign the database tier because it is now the saturated dependency.
3. Choose a partition key or sharding strategy that distributes request load more evenly.
4. Increase the relevant AWS service quota after validating the required scale.
5. Choose instance/network configuration with sufficient network bandwidth or reduce transfer demand.

Correct Answer: B

 

Correct Answer

Answer B is correct because Adding application servers cannot improve throughput once a shared database writer is the limiting resource. This directly meets the decisive requirement: database tier.

Incorrect Answers

Answer A is incorrect because The backlog shows that consumer throughput, not the web tier, is below the arrival rate. It may be useful elsewhere, but it does not meet the decisive requirement here: database tier.

Answer C is incorrect because A hot partition limits throughput while leaving capacity on other partitions unused. It may be useful elsewhere, but it does not meet the decisive requirement here: database tier.

Answer D is incorrect because A hard account or service quota stops capacity growth even when scaling policies request more resources. It may be useful elsewhere, but it does not meet the decisive requirement here: database tier.

Answer E is incorrect because A saturated network interface can cap throughput even when compute utilization is low. It may be useful elsewhere, but it does not meet the decisive requirement here: database tier.

 

Question 34

A fleet has spare CPU but all instances saturate their network bandwidth during data transfer. Which instance resource is the bottleneck?

1. Choose a partition key or sharding strategy that distributes request load more evenly.
2. Increase the relevant AWS service quota after validating the required scale.
3. Scale the worker/consumer tier or apply producer backpressure.
4. Scale or redesign the database tier because it is now the saturated dependency.
5. Choose instance/network configuration with sufficient network bandwidth or reduce transfer demand.

Correct Answer: E

 

Correct Answer

Answer E is correct because A saturated network interface can cap throughput even when compute utilization is low. This directly meets the decisive requirement: network bandwidth.

Incorrect Answers

Answer A is incorrect because A hot partition limits throughput while leaving capacity on other partitions unused. It may be useful elsewhere, but it does not meet the decisive requirement here: network bandwidth.

Answer B is incorrect because A hard account or service quota stops capacity growth even when scaling policies request more resources. It may be useful elsewhere, but it does not meet the decisive requirement here: network bandwidth.

Answer C is incorrect because The backlog shows that consumer throughput, not the web tier, is below the arrival rate. It may be useful elsewhere, but it does not meet the decisive requirement here: network bandwidth.

Answer D is incorrect because Adding application servers cannot improve throughput once a shared database writer is the limiting resource. It may be useful elsewhere, but it does not meet the decisive requirement here: network bandwidth.

 

Question 35

One shard receives 80% of requests because the partition key is skewed while other shards are idle. What must be corrected?

1. Choose instance/network configuration with sufficient network bandwidth or reduce transfer demand.
2. Increase the relevant AWS service quota after validating the required scale.
3. Scale the worker/consumer tier or apply producer backpressure.
4. Scale or redesign the database tier because it is now the saturated dependency.
5. Choose a partition key or sharding strategy that distributes request load more evenly.

Correct Answer: E

 

Correct Answer

Answer E is correct because A hot partition limits throughput while leaving capacity on other partitions unused. This directly meets the decisive requirement: partition-key distribution.

Incorrect Answers

Answer A is incorrect because A saturated network interface can cap throughput even when compute utilization is low. It may be useful elsewhere, but it does not meet the decisive requirement here: partition-key distribution.

Answer B is incorrect because A hard account or service quota stops capacity growth even when scaling policies request more resources. It may be useful elsewhere, but it does not meet the decisive requirement here: partition-key distribution.

Answer C is incorrect because The backlog shows that consumer throughput, not the web tier, is below the arrival rate. It may be useful elsewhere, but it does not meet the decisive requirement here: partition-key distribution.

Answer D is incorrect because Adding application servers cannot improve throughput once a shared database writer is the limiting resource. It may be useful elsewhere, but it does not meet the decisive requirement here: partition-key distribution.