SOA-C03 domain - 22% of the exam

Reliability and Business Continuity

Reliability and Business Continuity is 22% of the SOA-C03 exam. These are the objectives it covers, each with practice questions, with every answer explained.

The domain in numbers

82
Practice questions
7
Objectives
70
Medium
12
Hard

Where people go wrong

  • Confusing the four disaster recovery strategies

    The wrong answers label a design with the wrong strategy. Pilot light keeps only the replicated database running while the application tier stays switched off until failover. Warm standby keeps the whole stack running at reduced capacity and only scales it up. Active-active serves live traffic from both Regions at full capacity, and backup and restore runs nothing until a disaster is declared. Each wrong option credits the design with the running components of a neighbouring strategy.

    See it in a sample question
  • Treating Multi-AZ as Regional protection or read scaling

    Multi-AZ is offered where the requirement is a second Region, or where reads must scale. The Multi-AZ standby sits in another Availability Zone of the same Region, so a full Region outage takes it down with the primary, and the standby adds no read throughput. A second Region needs its own copy of the data, such as a cross-Region read replica that can be promoted at failover. Read scaling needs read replicas, which Aurora Auto Scaling adds and removes against a target metric.

    See it in a sample question
  • Using a replica as a way to rewind a bad write

    Replicas, replication and standbys look like a way to roll back corruption. They copy the corruption as well: a read replica, a DynamoDB global table, S3 replication and a Multi-AZ standby all receive the overwrite or the bad rows. Rewinding needs retained history instead. Point-in-time restore covers RDS and DynamoDB, and S3 Versioning keeps the displaced object as a prior version.

    See it in a sample question
  • Picking a scaling policy by name instead of trigger

    The wrong answers choose a scaling policy that does not match what drives the demand. Target tracking holds one set point and manages its own alarms. Step scaling maps several breach bands to different adjustments. Scheduled scaling acts at fixed clock times, so it suits a known daily spike. Predictive scaling learns a recurring pattern from history and provisions ahead. Reactive policies cannot have capacity ready before the load arrives.

    See it in a sample question
  • Swapping grace period, cooldown and deregistration delay

    Several settings on the Auto Scaling and load balancer side get applied to each other's jobs. The health check grace period delays health evaluation after a launch, so a slow-starting application is not terminated while booting. Cooldown only spaces out scaling actions. Deregistration delay lets in-flight requests finish on a target being removed, and a lifecycle hook is what pauses termination to run clean-up.

    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 sampleReliability and Business Continuityhard

A regulated workload runs in one Region with an RTO of about an hour and an RPO of a few minutes, but the business will only fund a recovery Region that keeps costs minimal until a disaster is declared. The team replicates the database continuously to a small standby instance in the second Region but keeps the application tier switched off there. At failover they will start and scale up the application tier. Which disaster recovery strategy does this design represent?

  • ABackup and restore, because data is copied to the second Region and the application is rebuilt from those backups only after a disaster is declared.
  • BPilot light, because the core data is replicated live to an always-on database while the application tier stays off until it is started at failover. Correct
  • CWarm standby, because a scaled-down but fully running copy of the whole stack already serves a fraction of traffic in the second Region at all times.
  • DMulti-site active-active, because the second Region holds a replica of the data and can take over production traffic during a Regional event.
Recognise pilot light as the strategy that keeps only the core data live across Regions while the application tier stays dormant until failover. The four common DR strategies sit on a spectrum of cost versus recovery speed. Pilot light keeps the minimal critical core, usually a continuously replicated database, always running in the recovery Region while compute and other tiers stay switched off to save money, then starts and scales them at failover. The live replication gives a small RPO and the dormant application tier keeps standing costs low, which is precisely the trade-off described.

Why A is wrong: Backup and restore keeps no live infrastructure running in the recovery Region and rebuilds everything from snapshots, so it cannot deliver the few-minute RPO a continuously replicated standby database provides.

Why B is correct: Pilot light keeps only the critical core, typically the replicated database, running continuously while the rest of the stack is dormant and brought up at failover, which matches the live standby database and switched-off application tier exactly.

Why C is wrong: Warm standby runs the application tier continuously at reduced capacity, but here the application tier is switched off, so the design stops short of a full always-on stack.

Why D is wrong: Active-active runs full production capacity serving live traffic in both Regions at once, whereas this design keeps the application tier off in the second Region until failover, so it is not active-active.

Free sampleReliability and Business Continuitymedium

A regulated workload already runs an Amazon RDS Multi-AZ database and stateless instances across two Availability Zones in one Region. Leadership now requires the workload to keep operating from a separate geographic location if the entire Region becomes unavailable. Which addition addresses this new requirement that the existing Multi-AZ design does not cover?

  • AAdd a third Availability Zone to the Auto Scaling group so the workload spreads across more zones and tolerates additional in-Region failures.
  • BIncrease the Multi-AZ standby instance size so failover completes faster and the database can absorb a larger Region-level traffic shift.
  • CBuild infrastructure and a cross-Region database copy in a second Region so the workload can recover there if the primary Region is unavailable. Correct
  • DEnable additional automated backups and copy the database snapshots to a second Availability Zone for faster restore during a Region outage.
Recognise that Multi-AZ addresses in-Region zone failures, while surviving a Region outage requires resources and data in a second Region. Multi-AZ deployments and multi-zone instances provide high availability against the loss of an Availability Zone, but all of that capacity lives within a single Region. Continuing to operate when an entire Region is unavailable is a disaster recovery problem that requires infrastructure and a replicated database in a second Region, which the in-Region Multi-AZ design does not provide.

Why A is wrong: Adding a third zone improves in-Region resilience but every zone still sits inside the same Region, so a full Region failure would take the whole workload offline.

Why B is wrong: A larger standby only affects in-Region failover behaviour and capacity; it provides no resource in another Region to fail over to when the whole Region is lost.

Why C is correct: Multi-AZ protects against zone failures within one Region only, so surviving a full Region outage requires resources and a database copy in a second Region to recover into.

Why D is wrong: Snapshots copied to another zone remain inside the same Region, so they are unavailable during a Region outage and do not provide a separate geographic recovery location.

Free sampleReliability and Business Continuitymedium

A team runs a critical DynamoDB table and must be able to restore it to any second within the last thirty-five days after an accidental bulk write corrupts records, without having to schedule and manage individual on-demand backups themselves. Which DynamoDB capability meets this recovery requirement most directly?

  • ASchedule a daily on-demand backup of the table and restore from the most recent daily backup when corruption is found.
  • BTurn on DynamoDB Streams and replay the captured change records to roll the table back to the moment before the corruption.
  • CConfigure a global table replica in a second Region and promote the replica to recover the data after the corruption occurs.
  • DEnable point-in-time recovery on the table so it can be restored to any second within the preceding thirty-five days. Correct
Enable DynamoDB point-in-time recovery to restore a table to any second within the last thirty-five days without managing manual backups. DynamoDB point-in-time recovery continuously backs up table data and lets you restore to any chosen second within a rolling thirty-five-day window, which is the right fit for recovering from accidental writes. On-demand backups only restore to discrete snapshot points, Streams retain changes for just twenty-four hours, and a global table replicates the corruption rather than preserving an earlier state, so none of those meet the to-the-second requirement.

Why A is wrong: Daily on-demand backups only allow recovery to each backup instant, so changes since the last daily snapshot are lost, which cannot meet a to-the-second recovery objective over thirty-five days.

Why B is wrong: Streams retain change records for only twenty-four hours and would require custom replay logic, so they are neither a thirty-five-day window nor a managed restore mechanism.

Why C is wrong: A global table replicates writes to other Regions almost immediately, so the bad bulk write propagates to every replica and there is no clean earlier state to promote.

Why D is correct: Point-in-time recovery continuously captures changes and lets the table be restored to any second within the last thirty-five days, which directly meets the fine-grained, self-managed recovery requirement.

Free sampleReliability and Business Continuitymedium

A web tier runs in an EC2 Auto Scaling group, and operators want the group to keep average CPU utilisation across the fleet close to 50 percent, adding or removing instances automatically as traffic rises and falls throughout the day. They want the simplest policy that maintains this set point without them defining individual thresholds for each capacity step. Which scaling policy meets this requirement with the least ongoing tuning?

  • AA simple scaling policy that adds two instances whenever a CPU alarm breaches and then waits for a cooldown before evaluating again.
  • BA step scaling policy with several CPU alarm bands that each add a different number of instances as utilisation climbs higher.
  • CA scheduled scaling action that sets desired capacity higher during the day and lower at night based on the usual traffic curve.
  • DA target tracking scaling policy on the average CPU utilisation metric with the target value set to 50 percent. Correct
Use a target tracking scaling policy to hold a metric at a chosen set point with the least manual threshold tuning. Target tracking works like a thermostat: you name a metric and a target value, and Auto Scaling provisions and manages the underlying CloudWatch alarms, computing the capacity changes needed to keep the metric near the target. This removes the per-band alarm and step design that simple and step scaling require, which is why it is the lowest-maintenance fit for a stable CPU set point.

Why A is wrong: Simple scaling reacts to one alarm with a fixed change and a blocking cooldown, so it cannot hold a continuous set point and needs the team to hand-tune the threshold and step.

Why B is wrong: Step scaling reacts faster than simple scaling but still forces operators to design and maintain every alarm band and step, which is exactly the per-threshold tuning they want to avoid.

Why C is wrong: Scheduled scaling changes capacity on a clock and is blind to the live CPU metric, so it cannot track an actual utilisation set point as traffic varies unpredictably.

Why D is correct: Target tracking creates and manages the CloudWatch alarms for you and adjusts capacity to hold the metric at the chosen set point, which is the lowest-effort way to keep CPU near 50 percent.

Free sampleReliability and Business Continuitymedium

Instances in an Auto Scaling group take roughly three minutes to start their application after launch. The group uses Elastic Load Balancing health checks, and operators find that freshly launched instances are repeatedly marked unhealthy and terminated before the application finishes starting, putting the group into a loop of launching and replacing instances. Which setting should they adjust so the new instances are given time to start before health evaluation can terminate them?

  • AReduce the load balancer health check interval so the target group probes each new instance more frequently and detects the healthy state sooner after start-up.
  • BIncrease the Auto Scaling health check grace period so the group ignores health check results for each instance until its application has had time to start. Correct
  • CIncrease the scaling cooldown so the group pauses between scaling activities and does not launch replacement instances while the previous ones are starting up.
  • DSwitch the group from Elastic Load Balancing health checks back to EC2 status checks so that application start-up time no longer affects the reported health.
Set an adequate health check grace period so Auto Scaling does not terminate instances before their application finishes starting. The health check grace period tells Auto Scaling to disregard health check results for a set number of seconds after an instance launches, which prevents the group from acting on the expected unhealthy reports during application start-up while still detecting genuine failures once the period elapses.

Why A is wrong: A shorter interval probes more often but still evaluates instances during the three-minute start-up window, so the group keeps marking them unhealthy and terminating them.

Why B is correct: The grace period delays health evaluation after launch, giving the application time to start before a failing check can trigger termination, which breaks the replacement loop.

Why C is wrong: Cooldown governs the gap between scaling actions, not how soon health checks evaluate a new instance, so the start-up termination loop continues.

Why D is wrong: EC2 status checks ignore the application, so a hung process would never be detected, weakening real failure detection rather than fixing the start-up loop.

Other domains in this exam

See also the SOA-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.