Amazon AWS Certified Solutions Architect - Associate SAA-C03 Bundle

The option Increase task memory sizing after measuring working-set demand addresses a separate condition: a Fargate task's memory limit repeatedly causes OOM kills. Container memory limits determine possible OOM termination. The actual requirement is a rollout sends traffic to tasks before warmup completes. Recommended corrective action: Configure readiness checks, grace periods and health-based deployment. Healthy target admission must match app readiness.

Adjust service desired count and target-tracking scaling metrics would help if an ECS service's tasks are saturated even while the ASG EC2 hosts have spare room. Task scaling governs application replica count. That is not the symptom described here: a rollout sends traffic to tasks before warmup completes. The matching response is Configure readiness checks, grace periods and health-based deployment, because healthy target admission must match app readiness.

Increase task CPU resources and benchmark throughput solves a different operational problem where a container processes CPU-bound requests but task CPU is set very low. Task CPU allocations affect processing capacity. It does not directly address a rollout sends traffic to tasks before warmup completes. The case calls for Configure readiness checks, grace periods and health-based deployment, since healthy target admission must match app readiness.

Use a continuously ready service or Lambda with appropriate provisioned capacity would help if a low-latency API should not wait for a Batch queue with minutes of startup. Queue-oriented batch startup does not meet strict API response SLAs. That is not the symptom described here: a nightly analytic task needs many CPU workers for 45 minutes. The matching response is Use AWS Batch array jobs with appropriately scaled compute environments, because batch can scale parallel tasks for bounded jobs.

For the alternative failure a data-transformation service runs Spark processing on a large dataset, the control Use AWS Glue or EMR Spark where its service pattern fits is relevant: managed distributed Spark suits parallel data transforms. Here the evidence instead is a nightly analytic task needs many CPU workers for 45 minutes. The selected technical mechanism is Use AWS Batch array jobs with appropriately scaled compute environments: batch can scale parallel tasks for bounded jobs.

The competing option Use EC2 accelerated instances through Batch or suitable ML training services would be useful when the job has a large GPU training requirement and must schedule accelerated hosts. Accelerated hardware meets GPU workloads unavailable on generic compute. The present issue is a nightly analytic task needs many CPU workers for 45 minutes; it is more directly addressed by Use AWS Batch array jobs with appropriately scaled compute environments. Batch can scale parallel tasks for bounded jobs.

For the alternative failure a workload has long-lived private network dependencies but should not manage EC2 instances, the control Use ECS/Fargate tasks with VPC networking is relevant: fargate manages container runtime while preserving private VPC paths. Here the evidence instead is a low-latency API should not wait for a Batch queue with minutes of startup. The selected technical mechanism is Use a continuously ready service or Lambda with appropriate provisioned capacity: queue-oriented batch startup does not meet strict API response SLAs.

Use AWS Glue or EMR Spark where its service pattern fits solves a different operational problem where a data-transformation service runs Spark processing on a large dataset. Managed distributed Spark suits parallel data transforms. It does not directly address a low-latency API should not wait for a Batch queue with minutes of startup. The case calls for Use a continuously ready service or Lambda with appropriate provisioned capacity, since queue-oriented batch startup does not meet strict API response SLAs.

The competing option Use AWS Batch array jobs with appropriately scaled compute environments would be useful when a nightly analytic task needs many CPU workers for 45 minutes. Batch can scale parallel tasks for bounded jobs. The present issue is a low-latency API should not wait for a Batch queue with minutes of startup; it is more directly addressed by Use a continuously ready service or Lambda with appropriate provisioned capacity. Queue-oriented batch startup does not meet strict API response SLAs.

Use AWS Batch array jobs with appropriately scaled compute environments solves a different operational problem where a nightly analytic task needs many CPU workers for 45 minutes. Batch can scale parallel tasks for bounded jobs. It does not directly address a workload has long-lived private network dependencies but should not manage EC2 instances. The case calls for Use ECS/Fargate tasks with VPC networking, since fargate manages container runtime while preserving private VPC paths.

The competing option Use EC2 accelerated instances through Batch or suitable ML training services would be useful when the job has a large GPU training requirement and must schedule accelerated hosts. Accelerated hardware meets GPU workloads unavailable on generic compute. The present issue is a workload has long-lived private network dependencies but should not manage EC2 instances; it is more directly addressed by Use ECS/Fargate tasks with VPC networking. Fargate manages container runtime while preserving private VPC paths.

Use a continuously ready service or Lambda with appropriate provisioned capacity would help if a low-latency API should not wait for a Batch queue with minutes of startup. Queue-oriented batch startup does not meet strict API response SLAs. That is not the symptom described here: a workload has long-lived private network dependencies but should not manage EC2 instances. The matching response is Use ECS/Fargate tasks with VPC networking, because fargate manages container runtime while preserving private VPC paths.

The competing option Use ECS/Fargate tasks with VPC networking would be useful when a workload has long-lived private network dependencies but should not manage EC2 instances. Fargate manages container runtime while preserving private VPC paths. The present issue is a data-transformation service runs Spark processing on a large dataset; it is more directly addressed by Use AWS Glue or EMR Spark where its service pattern fits. Managed distributed Spark suits parallel data transforms.

Use AWS Batch array jobs with appropriately scaled compute environments would help if a nightly analytic task needs many CPU workers for 45 minutes. Batch can scale parallel tasks for bounded jobs. That is not the symptom described here: a data-transformation service runs Spark processing on a large dataset. The matching response is Use AWS Glue or EMR Spark where its service pattern fits, because managed distributed Spark suits parallel data transforms.

For the alternative failure the job has a large GPU training requirement and must schedule accelerated hosts, the control Use EC2 accelerated instances through Batch or suitable ML training services is relevant: accelerated hardware meets GPU workloads unavailable on generic compute. Here the evidence instead is a data-transformation service runs Spark processing on a large dataset. The selected technical mechanism is Use AWS Glue or EMR Spark where its service pattern fits: managed distributed Spark suits parallel data transforms.

The option Use AWS Batch array jobs with appropriately scaled compute environments addresses a separate condition: a nightly analytic task needs many CPU workers for 45 minutes. Batch can scale parallel tasks for bounded jobs. The actual requirement is the job has a large GPU training requirement and must schedule accelerated hosts. Recommended corrective action: Use EC2 accelerated instances through Batch or suitable ML training services. Accelerated hardware meets GPU workloads unavailable on generic compute.

Use AWS Glue or EMR Spark where its service pattern fits would help if a data-transformation service runs Spark processing on a large dataset. Managed distributed Spark suits parallel data transforms. That is not the symptom described here: the job has a large GPU training requirement and must schedule accelerated hosts. The matching response is Use EC2 accelerated instances through Batch or suitable ML training services, because accelerated hardware meets GPU workloads unavailable on generic compute.

Use a continuously ready service or Lambda with appropriate provisioned capacity solves a different operational problem where a low-latency API should not wait for a Batch queue with minutes of startup. Queue-oriented batch startup does not meet strict API response SLAs. It does not directly address the job has a large GPU training requirement and must schedule accelerated hosts. The case calls for Use EC2 accelerated instances through Batch or suitable ML training services, since accelerated hardware meets GPU workloads unavailable on generic compute.

Use backlog-per-consumer or oldest-message age scaling signals would help if queue lag grows despite average CPU below 20 percent. Queue lag measures processing shortage directly. That is not the symptom described here: a service CPU metric has periodic spikes that trigger needless rapid scaling. The matching response is Use appropriate metric windows and cooldowns/warmup settings, because scaling stabilization prevents thrashing.

Diagnose downstream I/O and service traces before changing scale target solves a different operational problem where an ALB reports elevated target response time with normal request count. Latency may reflect dependency contention not capacity count. It does not directly address a service CPU metric has periodic spikes that trigger needless rapid scaling. The case calls for Use appropriate metric windows and cooldowns/warmup settings, since scaling stabilization prevents thrashing.

The competing option Expand subnet IP capacity or placement design before further scale-out would be useful when the autoscaling group cannot launch more instances because all subnet IPs are exhausted. Scale-out requires free addresses in selected subnets. The present issue is a service CPU metric has periodic spikes that trigger needless rapid scaling; it is more directly addressed by Use appropriate metric windows and cooldowns/warmup settings. Scaling stabilization prevents thrashing.

For the alternative failure a service CPU metric has periodic spikes that trigger needless rapid scaling, the control Use appropriate metric windows and cooldowns/warmup settings is relevant: scaling stabilization prevents thrashing. Here the evidence instead is queue lag grows despite average CPU below 20 percent. The selected technical mechanism is Use backlog-per-consumer or oldest-message age scaling signals: queue lag measures processing shortage directly.

The competing option Expand subnet IP capacity or placement design before further scale-out would be useful when the autoscaling group cannot launch more instances because all subnet IPs are exhausted. Scale-out requires free addresses in selected subnets. The present issue is queue lag grows despite average CPU below 20 percent; it is more directly addressed by Use backlog-per-consumer or oldest-message age scaling signals. Queue lag measures processing shortage directly.

The option Pre-scale capacity with scheduled actions around the ETL window addresses a separate condition: a scheduled ETL job creates a predictable burst of API calls each night. Scheduled scale reduces predictable burst delays. The actual requirement is queue lag grows despite average CPU below 20 percent. Recommended corrective action: Use backlog-per-consumer or oldest-message age scaling signals. Queue lag measures processing shortage directly.

The competing option Use appropriate metric windows and cooldowns/warmup settings would be useful when a service CPU metric has periodic spikes that trigger needless rapid scaling. Scaling stabilization prevents thrashing. The present issue is an ALB reports elevated target response time with normal request count; it is more directly addressed by Diagnose downstream I/O and service traces before changing scale target. Latency may reflect dependency contention not capacity count.

The option Pre-scale capacity with scheduled actions around the ETL window addresses a separate condition: a scheduled ETL job creates a predictable burst of API calls each night. Scheduled scale reduces predictable burst delays. The actual requirement is an ALB reports elevated target response time with normal request count. Recommended corrective action: Diagnose downstream I/O and service traces before changing scale target. Latency may reflect dependency contention not capacity count.

Use backlog-per-consumer or oldest-message age scaling signals would help if queue lag grows despite average CPU below 20 percent. Queue lag measures processing shortage directly. That is not the symptom described here: an ALB reports elevated target response time with normal request count. The matching response is Diagnose downstream I/O and service traces before changing scale target, because latency may reflect dependency contention not capacity count.

The option Diagnose downstream I/O and service traces before changing scale target addresses a separate condition: an ALB reports elevated target response time with normal request count. Latency may reflect dependency contention not capacity count. The actual requirement is a scheduled ETL job creates a predictable burst of API calls each night. Recommended corrective action: Pre-scale capacity with scheduled actions around the ETL window. Scheduled scale reduces predictable burst delays.

Expand subnet IP capacity or placement design before further scale-out would help if the autoscaling group cannot launch more instances because all subnet IPs are exhausted. Scale-out requires free addresses in selected subnets. That is not the symptom described here: a scheduled ETL job creates a predictable burst of API calls each night. The matching response is Pre-scale capacity with scheduled actions around the ETL window, because scheduled scale reduces predictable burst delays.

For the alternative failure a service CPU metric has periodic spikes that trigger needless rapid scaling, the control Use appropriate metric windows and cooldowns/warmup settings is relevant: scaling stabilization prevents thrashing. Here the evidence instead is a scheduled ETL job creates a predictable burst of API calls each night. The selected technical mechanism is Pre-scale capacity with scheduled actions around the ETL window: scheduled scale reduces predictable burst delays.

The option Use appropriate metric windows and cooldowns/warmup settings addresses a separate condition: a service CPU metric has periodic spikes that trigger needless rapid scaling. Scaling stabilization prevents thrashing. The actual requirement is the autoscaling group cannot launch more instances because all subnet IPs are exhausted. Recommended corrective action: Expand subnet IP capacity or placement design before further scale-out. Scale-out requires free addresses in selected subnets.

For the alternative failure a scheduled ETL job creates a predictable burst of API calls each night, the control Pre-scale capacity with scheduled actions around the ETL window is relevant: scheduled scale reduces predictable burst delays. Here the evidence instead is the autoscaling group cannot launch more instances because all subnet IPs are exhausted. The selected technical mechanism is Expand subnet IP capacity or placement design before further scale-out: scale-out requires free addresses in selected subnets.

Use backlog-per-consumer or oldest-message age scaling signals solves a different operational problem where queue lag grows despite average CPU below 20 percent. Queue lag measures processing shortage directly. It does not directly address the autoscaling group cannot launch more instances because all subnet IPs are exhausted. The case calls for Expand subnet IP capacity or placement design before further scale-out, since scale-out requires free addresses in selected subnets.