Skip to content

AWS Cloud Interview Questions: Multi-Region Disaster Recovery and Failover

Thirteen AWS interview questions on disaster recovery, covering RTO and RPO, regional failover, data replication, recovery testing, and operational readiness.

Ekanatha Reddy Urivakili14 min2,939 words

Multi-region disaster recovery is easy to describe and difficult to operate. A diagram with two regions does not prove that a system can recover: data may be stale, dependencies may remain unavailable, and a traffic switch may send users to an unready application.

These questions focus on the decisions and failure modes that distinguish a deployable recovery plan from a diagram. Answer in terms of business objectives, service behavior, and tested procedures—not a list of AWS products.

Fundamentals

1. What are RTO and RPO, and how do they shape an AWS recovery design?

What the interviewer is testing: Whether you can turn a business requirement into measurable engineering constraints.

Model answer:

  • Recovery time objective (RTO) is the maximum acceptable time to restore a service after disruption.
  • Recovery point objective (RPO) is the maximum acceptable amount of data loss, measured as time between the disruption and the latest recoverable data.
  • Define these per business capability, not just for “the application.” A read-only catalogue and a payment workflow may have very different tolerances.
  • Select architecture only after defining the objectives. Backups, replicated standby, and active-active systems have different costs, operational complexity, and recovery behavior.

An RPO of zero is not a casual design choice. It implies stronger consistency and availability constraints, and may make writes unavailable during some failures. Ask which failures and which data the objective covers.

Follow-up: How would you verify an RTO in production?
Hint: Measure elapsed time from declared incident to validated service, not merely infrastructure startup.

2. Is replication a substitute for backups?

What the interviewer is testing: Whether you distinguish availability from recoverability.

Model answer:

No. Replication can make data available in another location, but it can also copy unwanted changes: accidental deletion, corruption, or a malicious update may propagate. Backups provide a separate recovery path, ideally with retention controls and access protections distinct from the production write path.

A sound design may use both:

  • Replication to reduce recovery time after a regional failure.
  • Point-in-time recovery or versioned backups to recover from logical damage.
  • Restore exercises to prove that backups are usable and that dependencies, permissions, and application versions are available.

For Amazon S3, Cross-Region Replication is asynchronous and requires versioning on the relevant buckets. It should not be described as an immutable backup by itself. Review S3 replication documentation and define separate retention and deletion protections.

Follow-up: What would you do if corrupted data had already replicated?
Hint: Identify a known-good recovery point and restore into an isolated destination before replacing production data.

3. What does Multi-AZ protect you from, and what does it not protect you from?

What the interviewer is testing: Whether you understand failure domains.

Model answer:

Multi-AZ designs address failures within a Region, such as an Availability Zone disruption, depending on the service and configuration. They can improve availability without requiring the application to operate across Regions.

They do not, on their own, protect against a Region-wide outage, a bad deployment applied everywhere, compromised credentials, or logical data corruption. A multi-region recovery plan addresses a broader failure domain, but adds replication, routing, and operational complexity.

Start with the failure being addressed. Do not propose multi-region architecture just because “high availability” was mentioned; first establish the service’s impact tolerance and recovery objectives.

Follow-up: Would you make every workload multi-region?
Hint: Compare business impact and recovery targets with the cost and failure modes of the added architecture.

4. How do Route 53 failover and AWS Global Accelerator differ?

What the interviewer is testing: Whether you can choose a traffic-management mechanism based on client behavior and operational needs.

Model answer:

OptionHow it directs trafficUseful whenImportant consideration
Route 53 DNS failoverReturns DNS answers based on configured routing and health checksDNS-level regional routing is sufficientClient and resolver caching can delay the effect of a changed answer
AWS Global AcceleratorRoutes traffic through static anycast IP addresses to healthy regional endpointsYou want a stable network entry point and managed endpoint routingIt does not make an unhealthy application or its data recoverable

Neither option proves that the destination is ready to serve. Health checks should represent meaningful readiness, and routing changes should be coordinated with data promotion and application dependencies. See the Route 53 failover documentation and Global Accelerator documentation.

Follow-up: Why might a DNS change not move all users immediately?
Hint: Consider caching outside the authoritative DNS service.

Intermediate

5. When would you choose backup and restore, pilot light, warm standby, or active-active?

What the interviewer is testing: Whether you can explain the cost, complexity, and recovery trade-offs of common DR strategies.

Model answer:

Choose the least complex pattern that meets the agreed objectives. The names describe broad strategies; actual recovery behavior depends on implementation and testing.

StrategyRecovery-region stateTypical trade-off
Backup and restoreBackups and deployment artifactsLower continuously running cost; recovery takes longer and depends on successful restoration
Pilot lightCore data or minimal infrastructure is availableFaster than rebuilding everything; scaling and dependency startup still take time
Warm standbyA smaller but functional service is runningFaster cutover; requires keeping the standby deployable and data sufficiently current
Active-activeMultiple Regions serve trafficPotentially fast traffic shift; highest complexity for consistency, conflict handling, and operations

Do not claim a fixed RTO for a pattern. A pilot light with untested infrastructure, missing secrets, or manual approvals may recover more slowly than a carefully automated backup-and-restore plan.

Follow-up: What is the hidden cost of active-active?
Hint: Include engineering and incident-response complexity, not only duplicated infrastructure.

6. How would you reason about data consistency with DynamoDB global tables?

What the interviewer is testing: Whether you understand that multi-region writes require explicit consistency and conflict decisions.

Model answer:

First establish which consistency mode and topology are actually configured and supported. DynamoDB global tables replicate data across Regions, but the application must not assume that replication makes every cross-region operation globally serialized.

For multi-writer designs, decide what happens when independent Regions update the same logical item. Avoid placing business invariants that require one globally ordered decision behind unconstrained concurrent writes. You might assign ownership of writes by entity, route a class of updates to one home Region, or design explicit conflict resolution.

Also account for failover behavior: a Region can be unreachable while other Regions are still accepting writes. The application needs a policy for whether to keep accepting writes, queue them, or reject them until ownership is clear. Review the DynamoDB global tables documentation.

Follow-up: How would you prevent two Regions from both acting as the write owner?
Hint: Discuss ownership, fencing, and what happens when control-plane state is unavailable.

7. How would you use Amazon Aurora Global Database for regional recovery?

What the interviewer is testing: Whether you can describe the database’s role without treating promotion as the whole recovery plan.

Model answer:

Aurora Global Database has a primary Region and secondary Regions. Secondary clusters can serve read traffic, while the primary handles writes. In a regional failure, recovery may require promoting a secondary, redirecting application connections, and verifying that the application is ready to use the new writer.

Explain the distinctions between planned switchover and unplanned recovery. Their procedures and data implications differ. Confirm how the selected configuration behaves, including replication lag, promotion steps, connection recovery, and the handling of in-flight operations. Do not promise a particular RTO or RPO without testing the chosen setup.

The database is one dependency in the recovery sequence. The service also needs compatible application configuration, credentials, network access, and a safe way to move traffic. See Aurora Global Database documentation.

Follow-up: What should happen to client connection pools during promotion?
Hint: Consider stale connections, retries, and whether a retry can safely repeat a write.

8. What are the important recovery caveats with S3 Cross-Region Replication?

What the interviewer is testing: Whether you know the difference between copying objects and replicating an application’s complete data semantics.

Model answer:

S3 replication is asynchronous, so an object may not yet exist in the destination when a failure occurs. Design around the possibility that the newest writes are missing, and monitor replication health rather than assuming that configuration implies successful replication.

Also inspect the full object and bucket behavior:

  • Replication configuration and object eligibility.
  • Versioning, delete-marker behavior, and retention requirements.
  • Encryption keys and destination permissions.
  • How the application discovers the recovery bucket and verifies object availability.

Test the actual recovery workflow. A destination bucket that contains objects but cannot be read by the recovery role is not a usable recovery copy.

Follow-up: How would you measure whether replication is keeping up with the business RPO?
Hint: Track replication state or lag for representative data, and test the recovery path under realistic write patterns.

9. How do you decide the order of operations during regional failover?

What the interviewer is testing: Whether you see recovery as a dependency and coordination problem rather than a traffic switch.

Model answer:

Write the sequence down and identify prerequisites. A typical plan might:

  1. Declare the incident and establish who can authorize recovery.
  2. Stop or fence writes in the failed or isolated Region where possible.
  3. Confirm the recovery data source and promote it if required.
  4. Start or scale the application and validate dependencies.
  5. Run readiness checks, then shift traffic.
  6. Confirm critical user journeys and monitor for regressions.

The exact ordering depends on the services and failure mode. In particular, do not send users to a Region that is healthy at the network layer but has stale data, missing secrets, or no writable database. Make each step observable and safe to retry.

Follow-up: What if the primary Region is partially reachable rather than clearly down?
Hint: Explain how you prevent two Regions from accepting conflicting writes.

Senior and system-design level

10. Design a regional recovery plan for a stateless AWS API backed by a database and object storage.

What the interviewer is testing: Whether you can connect ingress, application readiness, data recovery, and operational procedures.

Model answer:

One possible design keeps a recovery stack deployed in a second Region, replicates or backs up data according to its recovery objectives, and routes traffic only after the recovery Region passes readiness checks. The application should be deployable from versioned infrastructure and release artifacts, not from undocumented manual steps.

Treat each data service separately. A database may require promotion; object storage may have asynchronous replication; caches may be rebuildable; queues may need replay or a controlled drain. For each, state the acceptable data loss and recovery action. Keep secrets, permissions, certificates, and required network paths available in the recovery Region.

A high-level flow looks like this:

flowchart LR
    client["Clients"] --> edge["Global entry"]
    edge --> primary["Primary region"]
    edge --> recovery["Recovery region"]
    primary --> primarydata["Primary data"]
    primarydata --> replication["Data replication"]
    replication --> recoverydata["Recovery data"]
    recovery --> recoverydata
    readiness["Readiness checks"] --> edge

A recovery Region is useful only when its application and data are ready to serve the required workload.

Follow-up: What would you deliberately keep out of the first version of this design?
Hint: Prioritize the workload and recovery objectives rather than cloning every production component.

11. How do you prevent split-brain writes during failover?

What the interviewer is testing: Whether you recognize that network isolation can make “the primary is down” ambiguous.

Model answer:

A partition can leave the old primary partially alive while the recovery Region is being promoted. If both sides accept writes, the result may be conflicting updates or violated business invariants. The plan needs an explicit authority model: who can promote, how write ownership changes, and how the old side is prevented from resuming as a writer.

Use fencing where the system supports it: revoke or invalidate the old writer’s authority, isolate its write path, or use a controlled ownership mechanism. A human approval step can be appropriate, but it must be clear who can act and what evidence they need. Automation should be idempotent so rerunning a recovery step does not create a second writer.

Do not confuse a health check with a fencing mechanism. A failed health check can trigger routing, but it does not necessarily stop the old Region from processing writes.

Follow-up: What would you do if you cannot prove the old writer is fenced?
Hint: Explain how your consistency and availability policy handles uncertainty.

12. How would you test disaster recovery without turning the exercise into an outage?

What the interviewer is testing: Whether you can validate the recovery process while controlling risk.

Model answer:

Test progressively. Start with a restore into an isolated environment, then exercise application startup and dependency checks, then conduct a planned regional recovery exercise with a scoped workload. Establish the exercise’s owner, stop conditions, communication plan, and rollback or return-to-primary procedure before it begins.

Measure what matters: time to a verified user journey, data freshness at recovery, failed runbook steps, and unexpected manual dependencies. Test both the documented steps and the people/process assumptions—such as access to credentials and authority to promote.

A tabletop exercise checks decisions and coordination; it does not prove that infrastructure works. A technical failover test proves more, but should still be scoped and controlled. Use findings to update runbooks and repeat the test after material architecture changes.

Follow-up: What evidence would make you trust the recovery plan?
Hint: Look for repeatable, timestamped results against the defined objectives.

13. What should you observe during a DR event, and what deserves an alert?

What the interviewer is testing: Whether you can make recovery measurable and distinguish actionable failures from dashboard noise.

Model answer:

Instrument the recovery path as a first-class service. Use RED metrics for the API—rate, errors, and duration—and USE metrics for constrained resources—utilization, saturation, and errors. Add recovery-specific signals: active Region, data replication health or lag, promotion state, readiness status, recovery-step duration, and the result of a synthetic critical journey.

Log decisions and transitions with structured fields. Avoid secrets and sensitive payloads. Trace a recovery request across the control plane and application, so the timeline shows whether delay came from promotion, deployment, readiness, or traffic movement.

For OpenTelemetry in a Python control API, the following initializes a tracer and creates a span for a readiness check. Install the OpenTelemetry SDK and OTLP exporter, and configure an OTLP collector endpoint before running it:

from opentelemetry import trace
from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.resources import Resource
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor

provider = TracerProvider(
    resource=Resource.create({"service.name": "dr-control-api"})
)
provider.add_span_processor(BatchSpanProcessor(OTLPSpanExporter()))
trace.set_tracer_provider(provider)

tracer = trace.get_tracer(__name__)

with tracer.start_as_current_span("recovery.readiness_check") as span:
    span.set_attribute("aws.region", "us-west-2")
    span.set_attribute("dr.role", "recovery")

Notice: Export spans to a collector and propagate trace context across service boundaries to connect the recovery timeline.

Follow-up: Which recovery signal would you page on first?
Hint: Choose a signal that indicates user impact or an imminent failure to meet the recovery objective.

Two operational checks

A minimal check can confirm that a DynamoDB table reports the expected replica Regions and statuses. Set TABLE_NAME and use credentials with permission to describe the table:

: "${TABLE_NAME:?Set TABLE_NAME}"
aws dynamodb describe-table \
  --table-name "$TABLE_NAME" \
  --query 'Table.Replicas[].{Region:RegionName,Status:ReplicaStatus}' \
  --output table

Notice: This inspects configured replica status; it does not prove that the application can read or write successfully in the recovery Region.

A more realistic smoke test checks whether a newly written source object becomes visible in a destination bucket. It assumes replication is configured, the test identity can write and read both buckets, and the object is eligible for replication. Set SOURCE_BUCKET and DESTINATION_BUCKET first.

#!/usr/bin/env bash
set -euo pipefail

: "${SOURCE_BUCKET:?Set SOURCE_BUCKET}"
: "${DESTINATION_BUCKET:?Set DESTINATION_BUCKET}"

file="$(mktemp)"
key="dr-smoke/$(date -u +%Y%m%dT%H%M%SZ)-$$"
trap 'rm -f "$file"' EXIT
printf 'replication smoke test\n' > "$file"

aws s3api put-object \
  --bucket "$SOURCE_BUCKET" \
  --key "$key" \
  --body "$file" >/dev/null

deadline=$((SECONDS + ${TIMEOUT_SECONDS:-120}))
while (( SECONDS < deadline )); do
  if aws s3api head-object \
      --bucket "$DESTINATION_BUCKET" \
      --key "$key" >/dev/null 2>&1; then
    printf 'Replicated object is readable: %s\n' "$key"
    exit 0
  fi
  sleep 3
done

printf 'Timed out waiting for replicated object: %s\n' "$key" >&2
exit 1

Notice: This tests a real write-to-read path; choose the timeout from your own recovery objectives and clean up test objects according to your retention policy.

Observability and recovery objectives

Treat the targets below as illustrative. Set production objectives with service owners, then verify that instrumentation and tests can measure them.

SLIExample objectiveWhat to measure
Recovery timeComplete recovery within the agreed RTOTime from incident declaration to validated critical journey
Data recovery pointRecovered data is within the agreed RPOAge or position of the newest recoverable business data
Recovery readinessRecovery Region passes its readiness checksApplication, data, identity, and dependency checks
Critical journey availabilityKey user workflows succeed after cutoverSynthetic or real transaction success by Region

Useful structured logs can make each transition auditable:

{
  "event": "dr.step.completed",
  "incident_id": "inc-example",
  "step": "recovery_readiness",
  "region": "us-west-2",
  "result": "success",
  "duration_ms": 1840,
  "runbook_version": "2026-11"
}

Page on conditions that need immediate human action, such as a recovery objective being at risk, failed promotion, an unexpected active writer, or a failed critical-journey check after cutover. Alert on sustained replication degradation according to the RPO, not on every transient metric fluctuation. Send lower-severity configuration drift and exercise findings to a non-paging workflow.

Common mistakes candidates make

  • Quoting an RTO without a measured recovery procedure. Infrastructure launch time is not the same as restored service.
  • Treating replication as backup. Replication may reproduce deletion or corruption.
  • Failing over traffic before data and dependencies are ready. Reachability is not readiness.
  • Assuming a DNS change is immediate. Resolver and client caching affect observed behavior.
  • Ignoring write ownership. A partial partition can leave two Regions accepting writes.
  • Using one recovery objective for every workload. Different business capabilities often need different recovery strategies.
  • Describing a service instead of its failure behavior. Explain what happens when replication lags, promotion fails, or credentials are unavailable.
  • Calling an untested runbook “automated.” Recovery steps need repeatable evidence and clear ownership.

Quick-revision checklist

  • Define RTO and RPO per business capability.
  • Name the failure domain each availability feature addresses.
  • Distinguish replication, backup, and point-in-time recovery.
  • Explain the selected recovery strategy and its trade-offs.
  • State how data promotion, write ownership, and split-brain prevention work.
  • Include identity, networking, secrets, and dependencies in the recovery plan.
  • Define readiness checks before shifting traffic.
  • Test restore, failover, and return-to-primary procedures.
  • Measure recovery with logs, metrics, traces, and synthetic journeys.
  • Keep recovery objectives tied to tested evidence, not architecture diagrams alone.

This article was generated with AI. #AIGenerated