SAA-C03 domain - 26% of the exam

Design Resilient Architectures

Design Resilient Architectures is 26% of the SAA-C03 exam. These are the objectives it covers, each with practice questions, with every answer explained.

The domain in numbers

95
Practice questions
7
Objectives
67
Medium
28
Hard

Where people go wrong

  • Choosing a scaling policy that cannot see the demand signal

    The wrong answers pick a scaling policy that does not match how demand arrives. Scheduled scaling is offered for load that drifts off any fixed timetable, and step or simple scaling for a requirement that rules out hand-set thresholds. In the other direction, reactive policies are offered for a surge whose timing is known, so capacity arrives after the peak starts. Target tracking holds average CPU at a chosen value, scheduled scaling suits a known time such as 09:00, and predictive scaling forecasts a cyclical pattern ahead of each peak.

    See it in a sample question
  • Expecting one queue to give every consumer a copy

    The wrong answers use a single SQS queue where several consumers each need every event. A queue hands each message to one consumer, so the tasks compete for messages instead of each receiving a copy. Direct calls and direct SNS subscriptions fail differently: they couple the publisher to every task or leave no buffer for a slow consumer. Publishing to an SNS topic with one SQS queue per consumer gives each its own durable copy, and new consumers subscribe without publisher changes.

    See it in a sample question
  • Trusting EC2 status checks to catch application failure

    A running instance that returns HTTP 503 still passes EC2 status checks, which cover only host and hypervisor reachability. The wrong answers lean on those default checks, or on a Route 53 health check or a CloudWatch alarm, none of which deregister a target. An ALB target group health check on the application path stops routing to the failing target. The Auto Scaling group replaces it only when its health check type is set to ELB.

    See it in a sample question
  • Treating a single-zone resource as surviving an AZ failure

    The wrong answers pick a resource that sits in one Availability Zone and expect it to survive that zone's loss. An EBS volume is replicated only within one zone, EFS One Zone storage and S3 One Zone-IA hold data in a single zone, and one NAT gateway leaves every private subnet without outbound access when its zone fails. EFS Regional storage and S3 Standard store data across multiple zones, and a NAT gateway per zone with zone-local routes gives each zone its own path.

    See it in a sample question
  • Matching the disaster recovery pattern to the wrong targets

    The wrong answers choose a pattern that either misses a recovery target or costs more than the targets justify. Daily backups breach a fifteen-minute recovery point, warm standby and active-active run compute or a full footprint that a four-hour recovery time does not need, and backup and restore cannot approach near-zero objectives. Backup and restore is the cheapest fit for a twenty-four hour objective, pilot light replicates data while compute stays off, and active-active suits near-zero targets.

    See it in a sample question

Objectives in this domain

What to study

The official documents this domain's practice questions cite most.

Sample questions from this domain

Free sampleDesign Resilient Architecturesmedium

A payroll application sees a sharp, repeatable load increase every weekday at 09:00 as staff log in, and demand drops again by 18:00. Reactive scaling reacts only after CPU has already climbed, causing slow responses for the first few minutes. Which Auto Scaling approach best ensures capacity is already in place before the morning surge?

  • AIncrease the cooldown period on the existing simple scaling policy so the group reacts more decisively when the morning CPU spike begins.
  • BLower the target value on the target tracking policy so the group keeps far more spare instances running throughout the day.
  • CSwitch to a step scaling policy with larger step sizes so the group adds many instances at once when CPU first rises.
  • DAdd a scheduled scaling action that raises the desired and minimum capacity shortly before 09:00 and lowers it after 18:00. Correct
Scheduled scaling pre-provisions Auto Scaling capacity for load changes that occur at known, repeatable times before demand actually arrives. When the timing of a load change is predictable, scheduled scaling sets desired and minimum capacity at a defined clock time, so instances finish launching and pass health checks before the surge hits; reactive policies act only after a metric has already breached, leaving an initial gap.

Why A is wrong: A longer cooldown only delays further scaling actions; it does nothing to add capacity ahead of 09:00 and would actually slow the group's reaction to the surge.

Why B is wrong: A lower target keeps extra capacity around the clock, which can cover the surge but wastes money during the long off-peak hours and still reacts rather than pre-provisioning for a known time.

Why C is wrong: Bigger step sizes speed the catch-up after CPU rises, but the group still launches only once the spike has begun, so the first minutes of the 09:00 surge stay underserved.

Why D is correct: Scheduled scaling changes capacity on a known clock pattern, so raising desired capacity just before 09:00 means instances are warm and registered before the predictable surge, then scaling back after 18:00.

Free sampleDesign Resilient Architecturesmedium

A photo-sharing platform must react to every new image upload by triggering three independent tasks at the same time: generate thumbnails, update a search index, and run content moderation. Each task is owned by a separate team and may be added or removed over time without changing the uploader code. Which messaging design best satisfies these requirements?

  • AWrite each upload event to a single Amazon SQS standard queue that all three tasks poll, so every task competes to read the same message and processes its share.
  • BHave the uploader invoke each task's API endpoint in sequence so the thumbnail, index, and moderation services run one after another for every uploaded image.
  • CPublish an upload event to an Amazon SNS topic and subscribe one Amazon SQS queue per task, letting each team consume its own queue independently of the others. Correct
  • DPublish the upload event to an Amazon SNS topic and subscribe the three task functions directly, accepting that a failed delivery to one subscriber drops that event for that task.
Apply the SNS-to-SQS fan-out pattern so a single event durably triggers multiple independent consumers that can change without touching the producer. SNS broadcasts each published message to all subscribers. Subscribing one SQS queue per task gives every task its own durable copy that it processes in parallel and at its own pace, and teams are added or removed by changing subscriptions rather than the uploader.

Why A is wrong: A single SQS queue delivers each message to only one consumer, so the three tasks would compete and most events would reach just one task rather than all three.

Why B is wrong: Direct synchronous calls couple the uploader to every task, force sequential rather than parallel processing, and require uploader code changes whenever a task is added or removed.

Why C is correct: The SNS fan-out pattern delivers each event to multiple SQS queues, so each task gets its own durable copy and teams can be added or removed by managing subscriptions, not uploader code.

Why D is wrong: Subscribing compute directly to SNS works for parallel fan-out, but without an intermediate queue a delivery failure can lose the event, unlike a durable per-task queue.

Free sampleDesign Resilient Architecturesmedium

A stateless web fleet runs on an Amazon EC2 Auto Scaling group behind an Application Load Balancer across three Availability Zones. Occasionally an instance keeps running but its application process hangs and returns HTTP 503 to every request, and the team wants the load balancer to stop sending user traffic to such an instance and the Auto Scaling group to replace it automatically with a fresh instance. Which two configurations together meet these requirements? Select TWO.

  • ALeave the Auto Scaling group health check type set to EC2 so the group relies only on the underlying instance system and hardware status checks to decide whether to replace an instance.
  • BAttach an Amazon CloudWatch alarm on group average CPU utilisation and let a target tracking scaling policy add instances when CPU is high.
  • CConfigure an Application Load Balancer target group health check that probes an application HTTP path and marks a target unhealthy on non-200 responses, so traffic stops routing to it. Correct
  • DEnable cross-zone load balancing on the Application Load Balancer so requests are spread evenly and the hung instance recovers under lighter load.
  • ESet the Auto Scaling group health check type to ELB so the group treats the load balancer's unhealthy verdict as a reason to terminate and replace that instance. Correct
Application-level ELB health checks combined with an ELB-type Auto Scaling health check both remove a hung instance from rotation and trigger its automatic replacement. An EC2 status check cannot see an application that has hung while the instance still runs, so detection must come from an Application Load Balancer health check that probes the app and fails on 503 responses. Setting the Auto Scaling group health check type to ELB makes the group honour that verdict and replace the instance, giving both traffic removal and self-healing.

Why A is wrong: EC2 status checks see the instance as running and healthy when only the application has hung, so the group never replaces it; this is the default trap that misses application failures.

Why B is wrong: Target tracking on CPU adjusts capacity for load but never identifies or replaces an individual hung instance, so the unhealthy target keeps lingering in the fleet.

Why C is correct: An application-level HTTP health check detects the hung process returning 503 and removes the failing target from rotation, which is exactly what stops user traffic reaching it.

Why D is wrong: Cross-zone load balancing only distributes requests more evenly across targets; it does neither detection nor replacement, and a hung process will not recover from lighter traffic.

Why E is correct: With the ELB health check type the group acts on the load balancer view, terminating the instance the target group reports unhealthy and launching a replacement automatically.

Free sampleDesign Resilient Architecturesmedium

A data engineering team must store large volumes of processed analytics output that is rarely changed but must survive concurrent failure of multiple storage devices and even an entire Availability Zone without manual replication work. They want the highest durability with no operational overhead for managing redundancy. Which storage choice meets this requirement?

  • AA General Purpose SSD Amazon EBS volume attached to a single EC2 instance, backed up to daily snapshots for additional safety.
  • BAn Amazon EFS file system mounted from multiple instances, relying on the file system to replicate data across zones automatically.
  • CAn instance store volume on a memory-optimised EC2 instance, with a nightly copy job pushing the data to a second instance for redundancy.
  • DAn Amazon S3 bucket using a standard storage class, which stores objects redundantly across multiple devices in at least three Availability Zones. Correct
Amazon S3 stores objects redundantly across at least three Availability Zones for very high durability without team-managed replication. S3 standard storage classes are designed for 99.999999999 percent durability by automatically replicating each object across multiple devices in a minimum of three Availability Zones. This redundancy survives device and AZ failures with no replication effort, which EBS, instance store, and even EFS do not match for rarely changed object data.

Why A is wrong: An EBS volume is replicated only within one Availability Zone and is tied to one instance, so it cannot survive an AZ failure on its own.

Why B is wrong: EFS does store across multiple zones, but it is a file system priced for shared POSIX access rather than the most durable, lowest-overhead object store for rarely changed output.

Why C is wrong: Instance store is ephemeral and tied to the host lifecycle, so data is lost when the instance stops, making it unsuitable for durable analytics output.

Why D is correct: S3 standard storage classes store each object redundantly across at least three Availability Zones, delivering eleven nines of durability with no replication work by the team.

Free sampleDesign Resilient Architectureshard

A payments company must survive the loss of an entire AWS Region with a recovery time objective and recovery point objective both close to zero, and it can absorb the cost of running full capacity in two Regions. Which architecture best meets these near-zero recovery targets?

  • AA pilot light in the second Region where only the replicated database runs and compute is started after a Region failure is confirmed.
  • BA multi-site active-active deployment serving live traffic from both Regions at full capacity with continuous data replication between them. Correct
  • CA warm standby running a scaled-down fleet that is enlarged to production capacity once the primary Region is declared lost.
  • DNightly cross-Region backups restored into a freshly built environment as soon as a Region outage is detected by monitoring.
Select multi-site active-active when both RTO and RPO must be near zero and full dual-Region cost is acceptable. Because both Regions already run full production capacity and replicate data continuously, a Region failure requires only redirecting traffic to the survivor, which is why active-active is the only pattern that drives both RTO and RPO close to zero.

Why A is wrong: Pilot light is cheaper but compute must be launched and scaled after failover, so recovery takes far longer than the near-zero RTO the business demands.

Why B is correct: Active-active already runs full production in both Regions with live replication, so losing one Region needs only traffic shifting, giving recovery objectives close to zero.

Why C is wrong: Warm standby recovers in minutes because of the scale-up step, which is good but still slower than the near-zero RTO that active-active provides.

Why D is wrong: Backup and restore is the slowest pattern with hours of rebuild time and nightly data loss, so it cannot approach a near-zero RTO or RPO.

Other domains in this exam

See also the SAA-C03 cert hub, the study guide, and the cheat sheet.

Examworthy is not affiliated with or endorsed by Amazon Web Services. Original, blueprint-aligned practice material only.