Skip to content
Certification Guide

Free sample questions

A free set of SAA-C03 practice questions, grouped by domain. Try each one, then open the answer and the notes on every wrong option. For the full practice bank with instant feedback, use practice.

Every question is written for this guide. None are real exam questions.

Design Secure Architectures

  1. Question 1

    A company exposes a public REST API through Amazon API Gateway. Security reviewers found cross-site scripting payloads in request bodies sent to the API. The company wants protection against common web exploits without writing custom rules, and with the LEAST operational overhead. What should it do?

    • A.Write an AWS WAF byte match rule for each malicious string found in the logs, and update the rule every time a new payload appears, and keep a log of each change.
    • B.Enable Amazon Inspector so that it blocks cross-site scripting payloads that reach the API, using its network exposure findings.
    • C.Enable AWS Shield Standard, which inspects request bodies for cross-site scripting patterns and blocks any payload that matches.
    • D.Attach an AWS WAF web ACL to the REST API that uses the AWS managed Core rule set, which inspects request bodies for cross-site scripting patterns.
    Show answer and explanation

    Correct answer: D

    The AWS managed Core rule set is maintained by AWS and includes cross-site scripting rules for query arguments, the URI path, cookies and the request body. Attaching it to the API's web ACL needs no custom rules.

    Why the other options are wrong

    • A.A byte match rule is a custom rule, which the company wants to avoid, and each new payload would need another rule.
    • B.Inspector finds software vulnerabilities and network exposure in EC2, ECR and Lambda. It does not inspect or block web requests.
    • C.Shield Standard defends against network and transport layer DDoS attacks. It doesn't inspect request bodies for cross-site scripting.
  2. Question 2

    A public API is flooded by a few hundred IP addresses, each sending about 1,000 requests per minute. The team wants to limit each address to a fixed request rate, without blocking all traffic from the countries where the attacks originate. The team wants minimal configuration. What is the best choice?

    • A.An AWS WAF rate-based rule that aggregates requests by IP address, with a request limit per window and a block action for addresses above it.
    • B.An AWS WAF geo match rule that blocks every country where the attacks originate, which also stops legitimate customers from those countries.
    • C.A network ACL deny rule for each attacking address, updated every minute from the access logs by a scheduled job that the team must build and maintain.
    • D.AWS Shield Standard, which automatically limits the number of requests from each IP address during any flood, with no configuration needed on the team's side.
    Show answer and explanation

    Correct answer: A

    A rate-based rule counts requests per aggregation key, such as IP address, over an evaluation window, and applies its action to addresses above the limit. New attacking addresses are covered automatically, and the rest of the traffic is untouched.

    Why the other options are wrong

    • B.Country blocks also stop legitimate users in those countries, and they don't limit the request rate for each address.
    • C.This is manual work on every change, and the attackers keep changing addresses. It also uses up the limited rule count of the network ACL.
    • D.Shield Standard protects against network and transport layer DDoS attacks. It doesn't set per-address request limits.
  3. Question 3

    A company replicates objects from its source bucket in account A to a bucket in account B. The replicated objects must be owned by the account B bucket owner, so that account A cannot retain access through object ownership. Which configuration does this?

    • A.Enable bucket ACLs in account B and grant account A full control on each object.
    • B.Turn on S3 Object Lock in account B, which transfers ownership of each replica.
    • C.Set the replication rule's owner override to the destination account, and grant the replication role s3:ObjectOwnerOverrideToBucketOwner in the destination bucket policy.
    • D.Enable SSE-KMS on account B's bucket, which changes the owner of every replica to the key owner.
    Show answer and explanation

    Correct answer: C

    Replica ownership changes only when the replication rule sets an owner override to the destination account and the replication role has the s3:ObjectOwnerOverrideToBucketOwner permission in the destination bucket policy. Object Lock and encryption do not change ownership.

    Why the other options are wrong

    • A.ACL grants keep account A in control of each object, which is the outcome the question is trying to prevent.
    • B.Object Lock protects objects from deletion and does not change who owns them.
    • D.Encryption with a key does not change object ownership. The key owner is a separate question from the object owner.
  4. Question 4

    A company protects a public API with one AWS WAF web ACL. The web ACL must block SQL injection and stop any single client from sending a flood of requests. Which TWO rules should the web ACL include? (Choose TWO.)

    • A.The AWS managed SQL database rule group.
    • B.A geo match rule that blocks the countries where the attacks originate.
    • C.A byte match rule that blocks the word SELECT in query strings.
    • D.A rate-based rule that counts requests per client IP address and blocks addresses that exceed the limit.
    • E.An AWS Shield Advanced protection group for the API.
    Show answer and explanation

    Correct answer: A and D

    The managed SQL database rule group inspects requests for SQL injection patterns, and a rate-based rule limits each client's request rate. Together they cover both threats in one web ACL.

    Why the other options are wrong

    • B.Country blocks don't match SQL payloads, and they don't limit the rate for each client. They also block legitimate users in those countries.
    • C.A single keyword match blocks legitimate requests that contain the word, and it is a weak substitute for the managed SQL injection rules.
    • E.A protection group groups resources for Shield. It isn't a rule in the web ACL.
  5. Question 5

    Analysts sign in with an IAM user that has MFA and run reports from the AWS CLI by assuming a role named ReportRole in the same account. A report now takes up to 7 hours, and when the analysts call aws sts assume-role with --duration-seconds 28800, the call fails. ReportRole was created with default settings. Which change allows the 8-hour session?

    • A.Attach a session policy to the AssumeRole call that sets the duration to 8 hours.
    • B.Change the maximum session duration of ReportRole to 8 hours, which the analysts can then request.
    • C.Have the analysts call GetSessionToken first and then assume ReportRole with those credentials.
    • D.Create access keys for the analysts so that the role is no longer needed.
    Show answer and explanation

    Correct answer: B

    A role's maximum session duration defaults to 1 hour and can be set from 1 to 12 hours. AssumeRole fails when the requested duration exceeds it, so raising it to 8 hours lets the 28,800-second request succeed. The analysts assume the role from IAM user credentials, so role chaining's one-hour cap doesn't apply.

    Why the other options are wrong

    • A.Session policies can only narrow permissions; they don't control how long a session lasts.
    • C.Starting from session-token credentials doesn't raise the role's limit. The request still exceeds ReportRole's 1-hour default maximum and fails.
    • D.Long-lived access keys give up temporary credentials, and the analysts need the role's permissions anyway.
  6. Question 6

    A security lead wants one place to view findings from GuardDuty, Inspector and Macie across many accounts, plus a score against the AWS Foundational Security Best Practices standard. Which service should the lead use?

    • A.Amazon Detective
    • B.AWS Config
    • C.AWS Security Hub CSPM
    • D.AWS Firewall Manager, which deploys WAF and Shield protections across many accounts from one central policy.
    Show answer and explanation

    Correct answer: C

    Security Hub CSPM collects findings from GuardDuty, Inspector, Macie and supported third-party products, runs security checks, and reports a score against standards such as the AWS Foundational Security Best Practices standard.

    Why the other options are wrong

    • A.Detective helps investigate a finding using linked log data. It doesn't collect findings from other services or score the account against a standard.
    • B.AWS Config records resource configuration and evaluates configuration rules. It doesn't aggregate findings from security services.
    • D.Firewall Manager deploys and manages protections across accounts, such as WAF rules. It doesn't aggregate findings or score security standards.
  7. Question 7

    A research institute runs one standalone AWS account and has decided never to adopt AWS Organizations. Staff already sign in to every internal system through the institute's SAML 2.0 identity provider. Staff in two directory groups need AWS Management Console access with different permissions, and nobody should have an IAM user. Which solution meets these requirements?

    • A.Create an IAM SAML identity provider and two roles that trust it, one per directory group, with the group chosen in the SAML assertion.
    • B.Enable an account instance of IAM Identity Center and assign two permission sets for the account to the groups.
    • C.Create an IAM user for each staff member, synchronized nightly from the directory, and place them in two IAM groups.
    • D.Create an Amazon Cognito user pool federated with the identity provider and grant the pool's groups console access.
    Show answer and explanation

    Correct answer: A

    IAM SAML 2.0 federation gives federated users console access through roles without IAM users. The identity provider sends the role in its assertion based on group membership, and each role's trust policy allows the SAML provider.

    Why the other options are wrong

    • B.Account instances of IAM Identity Center manage access only to supported applications in that account; permission sets for AWS account access need an organization instance, which requires AWS Organizations.
    • C.This creates IAM users, which the requirement forbids, and copies identities that can drift from the directory.
    • D.Cognito user pools sign in users to your own applications; they don't give workforce users access to the AWS Management Console.
  8. Question 8

    A company must encrypt every new object in an existing S3 bucket with a customer managed KMS key. The applications that upload the objects cannot be changed. Which setting meets this requirement?

    • A.Client-side encryption in each application, using the customer managed key.
    • B.Default bucket encryption set to SSE-KMS with the customer managed key.
    • C.A bucket policy that denies uploads without an encryption header.
    • D.S3 Versioning, so that each upload is kept as a version.
    Show answer and explanation

    Correct answer: B

    Default bucket encryption applies SSE-KMS to every object uploaded without its own encryption header, so the uploading applications need no change. A bucket policy that denies unencrypted uploads makes those uploads fail, and it does not encrypt anything itself.

    Why the other options are wrong

    • A.Each application would have to change, which the question rules out.
    • C.The uploads would fail rather than be encrypted, so the applications would break.
    • D.Versioning keeps old versions and does not encrypt objects.
  9. Question 9

    An administrator wants all members of the IAM group Analysts to read objects in a reporting bucket in the same account. They write a bucket policy whose Principal is the ARN of the Analysts group, but saving the policy fails with an invalid principal error. Which change grants the access with the LEAST effort?

    • A.Replace the group ARN with the ARN of each analyst's IAM user and edit the policy whenever the group changes.
    • B.Nest the Analysts group inside a new group that is allowed as a principal.
    • C.Attach an identity-based policy that allows s3:GetObject on the bucket to the Analysts group.
    • D.Grant the Analysts group read access with an access control list (ACL) on the bucket.
    Show answer and explanation

    Correct answer: C

    IAM groups can't be principals in resource-based policies because groups don't authenticate; they only carry permissions. Attaching the read permission to the group's identity policy gives every member access, and in the same account no bucket policy statement is needed.

    Why the other options are wrong

    • A.Listing every user works, but the bucket policy must be edited whenever someone joins or leaves, which is more effort than one group policy.
    • B.IAM groups can't be nested, and no group can be a principal.
    • D.S3 ACLs grant access to AWS accounts or predefined Amazon S3 groups, not to IAM groups.
  10. Question 10

    A company's public certificate from AWS Certificate Manager was issued with email validation. The certificate renews next year, and the company wants renewals to keep working without changes to its DNS process later. What should it do?

    • A.Keep email validation, because renewals always reuse the same validation method and that method never changes between renewal cycles.
    • B.Switch the certificate from email to DNS validation before the renewal, because email validation is being deprecated for renewals.
    • C.Import a certificate from a third-party CA and renew it by hand each year, so that the renewal date stays under the team's own control.
    • D.Turn off automatic renewal so that the certificate is not validated again, which keeps the domain's validation status unchanged.
    Show answer and explanation

    Correct answer: B

    ACM is deprecating email validation for new public certificates and renewals throughout 2027. Switching the certificate to DNS validation before renewal keeps renewal automatic, because the DNS validation record stays in place.

    Why the other options are wrong

    • A.Email validation is being deprecated for renewals, so keeping it is the risky option the question is testing.
    • C.An imported certificate needs manual renewal, which the question's automation goal does not allow.
    • D.Turning off renewal lets the certificate expire, which breaks the service.
  11. Question 11

    A data team's AWS account is in an OU whose only SCP allows actions for Amazon S3, AWS Glue and Amazon Athena. The FullAWSAccess policy has been replaced by this allow-list SCP at the OU. A data engineer's IAM role in the account has the AdministratorAccess managed policy. The engineer tries to create an Amazon RDS database and to start an AWS Glue job. What happens?

    • A.Both requests succeed, because AdministratorAccess allows all actions.
    • B.Both requests fail, because an SCP and an identity-based policy can't both apply to the same role.
    • C.The RDS request succeeds and the Glue request fails, because SCPs override only managed policies.
    • D.The Glue request succeeds and the RDS request fails, because the SCP and the identity policy intersect.
    Show answer and explanation

    Correct answer: D

    An action must be allowed by every applicable SCP and by the role's identity-based policy. Glue is allowed by both. RDS is allowed by AdministratorAccess but not by the SCP, so it is implicitly denied.

    Why the other options are wrong

    • A.AdministratorAccess can't exceed the SCP. SCPs cap every user and role in member accounts, even administrators.
    • B.SCPs and identity-based policies always apply together; the result is their intersection, not a conflict.
    • C.SCPs don't distinguish managed from inline policies, and the outcomes are reversed: RDS is outside the SCP.

Design Resilient Architectures

  1. Question 1

    A company uses an Aurora PostgreSQL global database with a secondary cluster in a recovery Region. The company wants an Aurora-enforced upper bound on how far the recovery Region can fall behind, even if that limits writes on the primary cluster. Which Aurora capability provides this?

    • A.Enable Multi-AZ on each secondary cluster, so that a failure in one zone of the recovery Region does not affect the RPO.
    • B.Set the rds.global_db_rpo parameter, so Aurora blocks commits on the primary when the secondary's lag passes the bound.
    • C.Copy a snapshot of the primary to the secondary every hour, so the secondary holds a recovery point that is never more than an hour old.
    • D.Disable replication while the primary Region is healthy to save cost, and turn it back on only after an outage starts.
    Show answer and explanation

    Correct answer: B

    On an Aurora PostgreSQL global database, the rds.global_db_rpo parameter sets an upper bound on RPO for the secondary clusters. Aurora tracks replication lag against that bound, and when the bound is exceeded it blocks transactions on the primary cluster. That trade-off is what makes the RPO enforceable.

    Why the other options are wrong

    • A.Multi-AZ protects within one Region and does not bound cross-Region replication lag.
    • C.Hourly snapshots give a recovery point at least an hour old and do not bound replication lag.
    • D.Disabling replication removes the secondary's data, so it cannot meet any RPO.
  2. Question 2

    A ticketing company's only AWS Lambda function in its account serves a public API. During a major on-sale event, requests fail with throttling errors because the function reaches the account's 1,000 concurrent execution quota. Some of the requests are purchase confirmations that can be processed a few seconds later. Which TWO actions will reduce the failures? (Choose TWO.)

    • A.Configure provisioned concurrency of 1,000 on the function.
    • B.Request a higher concurrent executions quota for the Region through Service Quotas.
    • C.Set reserved concurrency of 1,000 on the function.
    • D.Send the purchase confirmations to an Amazon SQS queue and process them with a separate consumer, so bursts wait in the queue instead of failing.
    • E.Increase the function's timeout from 30 seconds to 15 minutes.
    Show answer and explanation

    Correct answer: B and D

    The concurrency quota of 1,000 per account per Region is a default that can be increased. Moving the work that tolerates delay onto a queue turns rejected bursts into a backlog that a consumer drains at a controlled rate.

    Why the other options are wrong

    • A.Provisioned concurrency counts against the same account concurrency quota. It removes cold starts but adds no capacity beyond the limit.
    • C.Reserved concurrency can't exceed the account quota and only sets aside part of it. With one function in the account it changes nothing.
    • E.A longer timeout makes each invocation hold its concurrency longer, which makes throttling worse.
  3. Question 3

    A retailer publishes order events to a custom Amazon EventBridge event bus. A bug in a downstream consumer caused it to ignore events for the past 6 hours. After the fix is deployed, the team must reprocess those events without asking the producers to send them again. Which approach should a solutions architect have put in place to make this possible?

    • A.A dead-letter queue configured on the consumer rule's target, then a redrive of the messages in that queue once the fix is deployed.
    • B.An Amazon SNS topic as a second target on every rule.
    • C.An EventBridge archive on the event bus, then a replay of the archived events for the affected time window.
    • D.Amazon CloudWatch Logs as a second target and a script that re-publishes the logged events.
    Show answer and explanation

    Correct answer: C

    An EventBridge archive stores events that match its pattern, for a retention period you choose. After a fix, you start a replay for a time window, and EventBridge sends those archived events back to the bus so the rules deliver them again.

    Why the other options are wrong

    • A.A target DLQ only catches events that EventBridge failed to deliver. These events were delivered and then ignored by buggy code, so they never reached the DLQ.
    • B.SNS pushes events to subscribers and keeps no history to replay later.
    • D.Logging events and writing a re-publish script works in principle, but it is custom tooling for a capability that EventBridge provides as a managed feature.
  4. Question 4

    A Java application connects to an Amazon RDS for PostgreSQL Multi-AZ DB instance through its DNS endpoint. After each failover, the application returns connection errors for several minutes until it is restarted, even though the database is available again within about two minutes. The team must fix this with the LEAST code change. Which change is most likely to resolve the problem?

    • A.Increase the connection pool size to 500 connections so that requests can wait for a free connection during the failover.
    • B.Set the JVM DNS cache time-to-live to 60 seconds or less.
    • C.Replace the endpoint name with the standby's IP address in the connection string, so the client connects directly.
    • D.Add a read replica and point all application queries at it, so that connections stop depending on the primary endpoint.
    Show answer and explanation

    Correct answer: B

    After a failover, the DNS record points to the standby. A JVM that caches the old address keeps connecting to the former primary until its cache expires. Setting the JVM DNS TTL to 60 seconds or less lets the application pick up the new address without code changes.

    Why the other options are wrong

    • A.A bigger pool keeps trying the old address, so it does not fix the stale DNS result.
    • C.The standby's IP address changes on failover, and a hard-coded address breaks the same way.
    • D.A read replica does not change the write path, and the application still caches the old primary address.
  5. Question 5

    A company runs 60 microservices on an on-premises Kubernetes cluster. All deployments are defined as Kubernetes manifests and Helm charts, and the platform team relies on Kubernetes tooling. The company wants to move to AWS while changing its deployment tooling as little as possible. Which solution meets these requirements?

    • A.Run the microservices as Amazon ECS services on AWS Fargate, converting each deployment to an ECS task definition.
    • B.Run the microservices as AWS Lambda functions.
    • C.Run the containers on Amazon EC2 instances with Docker and a deployment script.
    • D.Run the microservices on Amazon EKS and deploy them with the existing manifests and Helm charts.
    Show answer and explanation

    Correct answer: D

    Amazon EKS runs upstream Kubernetes with a managed control plane, so existing manifests, Helm charts and Kubernetes tooling keep working. ECS would be simpler only for a team without Kubernetes investment.

    Why the other options are wrong

    • A.ECS uses task definitions and services, not Kubernetes manifests, so every deployment would have to be rewritten.
    • B.Moving 60 containerized services to Lambda means rewriting them as functions and discarding the Kubernetes tooling.
    • C.This drops Kubernetes entirely and replaces it with custom scripts the team must build and maintain.
  6. Question 6

    A team runs an Amazon RDS for PostgreSQL database. The team wants the database to fail over automatically to a standby in another Availability Zone, and it does not want to change the application's connection string after a failover. Which configuration meets these requirements?

    • A.A read replica in the same Availability Zone as the primary.
    • B.A Multi-AZ DB instance deployment.
    • C.A manual snapshot taken every hour.
    • D.Two independent DB instances, with the application choosing one at random.
    Show answer and explanation

    Correct answer: B

    A Multi-AZ DB instance deployment keeps a synchronous standby in another Availability Zone and fails over automatically. RDS points the same DNS name at the standby, so the connection string does not change.

    Why the other options are wrong

    • A.A read replica is not a failover target by itself, and a replica in the same zone does not protect against a zone failure.
    • C.An hourly snapshot gives you a recovery point, but restoring it creates a new instance with a new endpoint, which takes much longer than an automatic failover.
    • D.Two independent instances with random selection split the data and do not fail over. The application would still break when its chosen instance is down.
  7. Question 7

    A checkout request passes through an API, four microservices and a database. Some checkouts take 9 seconds, but each service's own logs show normal response times. The team can add an SDK to each service in the next sprint and needs a view of where the time goes across the whole request path. Which AWS service should it use?

    • A.AWS X-Ray tracing service
    • B.AWS CloudTrail
    • C.AWS Config
    • D.Amazon Inspector
    Show answer and explanation

    Correct answer: A

    AWS X-Ray traces a single request as it moves through distributed services and builds a trace map that shows where the time goes. It is built for finding latency across service boundaries.

    Why the other options are wrong

    • B.CloudTrail records API activity in your account. It does not show the timing of a request as it moves between services.
    • C.AWS Config records resource configuration and changes. It does not measure request latency.
    • D.Amazon Inspector scans workloads for vulnerabilities. It does not trace requests.
  8. Question 8

    A software company runs a cleanup AWS Lambda function for each of its 300 customers once a day, at 01:00 in each customer's own time zone. Today a cron daemon on a single Amazon EC2 instance invokes the functions, and missed runs go unnoticed when the instance fails. Which solution meets these requirements with the LEAST operational overhead?

    • A.Run the cron daemon on two EC2 instances in different Availability Zones, and add a shared database lock so each daily job runs only once.
    • B.Have the Lambda function re-invoke itself every minute and check whether it is 01:00 in any customer's time zone before running that customer's cleanup.
    • C.Create an Amazon SQS delay queue and send each customer's message with a delay of 24 hours so it reappears for processing at 01:00 each day.
    • D.Create one EventBridge Scheduler schedule per customer, with a cron expression, the customer's time zone and the Lambda function as target.
    Show answer and explanation

    Correct answer: D

    EventBridge Scheduler is a serverless scheduler that supports cron and rate expressions with a time zone per schedule and invokes targets such as Lambda directly, with retries and an optional dead-letter queue. There is no scheduling server to run or monitor.

    Why the other options are wrong

    • A.Two instances remove one failure point but double the servers to patch, and both instances would run every job unless you add locking.
    • B.A self-invoking loop is fragile, runs about 1,440 times a day for no work, and hides failures the same way the cron server does.
    • C.SQS delays are limited to 15 minutes, so a 24-hour delay isn't possible.
  9. Question 9

    A research team submits thousands of independent containerized simulation jobs each day. Each job runs for 20 to 40 minutes. The team wants the jobs queued, retried on failure and run on compute that scales to zero between submissions, with no servers or custom scheduler to manage. Which solution meets these requirements?

    • A.Run the jobs with AWS Batch using a Fargate compute environment.
    • B.Invoke an AWS Lambda function for each job.
    • C.Run an Amazon ECS service on AWS Fargate with 100 tasks always running that pull jobs from a database table.
    • D.Run a fleet of Amazon EC2 instances with a custom scheduler that assigns jobs.
    Show answer and explanation

    Correct answer: A

    AWS Batch is a purpose-built, managed batch scheduler. Jobs wait in a job queue, Batch retries failed jobs according to the retry strategy, and a Fargate compute environment runs each container with no instances and no capacity left running when the queue is empty.

    Why the other options are wrong

    • B.A Lambda invocation can run for at most 15 minutes, so jobs of 20 to 40 minutes would be stopped.
    • C.Always-running tasks don't scale to zero, and the team would build its own queueing and retry logic on the table.
    • D.The team would manage instances and write the scheduler, which the requirements rule out.

Design High-Performing Architectures

  1. Question 1

    A SaaS company serves a cacheable marketing site and a partner API. The partner firewall must allowlist fixed addresses for the API, which is HTTPS only. The marketing pages should be cached at the edge. The company wants the LEAST operational overhead. Which design meets these requirements?

    • A.Global Accelerator alone in front of both the marketing site and the partner API, with the same two static addresses.
    • B.CloudFront alone for both hostnames, with the partner firewall allowlisting the CloudFront edge ranges it publishes.
    • C.CloudFront for the cacheable marketing pages, and Global Accelerator on its own hostname for partner API addresses.
    • D.Route 53 weighted records with a zero TTL pointing at each Region's ALB, for both hostnames and both protocols.
    Show answer and explanation

    Correct answer: C

    Use each service for what it does. CloudFront caches the marketing pages, and Global Accelerator gives the API two static addresses the partner can allowlist.

    Why the other options are wrong

    • A.Global Accelerator doesn't cache, so the marketing pages would reach the origin on every request.
    • B.CloudFront's edge addresses are shared and broad, so the partner can't allowlist a fixed pair for the API.
    • D.Route 53 records return changing endpoint addresses, and a zero TTL increases DNS queries without giving fixed IPs.
  2. Question 2

    A reporting service reads from an RDS for MySQL read replica. After an order is placed, the order confirmation page sometimes shows the order as missing, because the replica is a few seconds behind. Browsing reads should continue to use the replica. Which approach meets both requirements?

    • A.After checkout, send the confirmation read to the writer, and keep browsing reads on the replica.
    • B.Send every read to the replica, so the writer handles only writes and never serves a read.
    • C.Send every read to the writer, so the replica serves no traffic and the writer carries all read load.
    • D.Wait a fixed 60 seconds after each write before reading from the replica, to let it catch up each time.
    Show answer and explanation

    Correct answer: A

    Replication is asynchronous, so the replica can lag, and a read that must see the newest write goes to the source. Browsing reads can tolerate some lag, so they stay on the replica. A fixed wait is a guess, because lag isn't guaranteed to stay under any fixed time.

    Why the other options are wrong

    • B.This returns the stale results the confirmation page must avoid.
    • C.This removes the read capacity the browsing traffic needs from the replica.
    • D.Replica lag varies, so a fixed wait can still return stale data, and the delay hurts users.
  3. Question 3

    A DynamoDB table has 50 million order items. About 0.1 percent carry FlagForReview, a boolean that is true on flagged items. Flagged items also carry FlagDay, the flag date as a string such as 2026-11-16, and FlagDate, a timestamp. The support team needs the flagged orders listed by flag date, and the list must not scan the whole table. Which index design meets this requirement?

    • A.A global secondary index with FlagDay as the partition key and FlagDate as the sort key, which holds only the flagged items.
    • B.A global secondary index with FlagForReview as the partition key and FlagDate as the sort key, which lists flagged items in date order.
    • C.A local secondary index with FlagForReview as the sort key, which can be added to the table at any time.
    • D.A scan of the whole table with a filter on FlagForReview, run once per day by a scheduled job.
    Show answer and explanation

    Correct answer: A

    A sparse GSI keyed on FlagDay (a string) with FlagDate as the sort key holds only flagged items, so one Query returns a day's flags in time order without scanning. FlagForReview is a boolean, and DynamoDB key attributes must be string, number or binary, so no index can use it as a key (option b). An LSI must keep the table's partition key (OrderId) and can only be defined when the table is created (option c). A scan reads all 50 million items on each run (option d).

    Why the other options are wrong

    • B.FlagForReview is a boolean, and DynamoDB key attributes must be string, number or binary, so no index can use it as a key.
    • C.An LSI must keep the table's partition key (OrderId) and can only be created with the table, so it cannot group flags by date.
    • D.A scan reads all 50 million items on every run, which the requirement rules out.
  4. Question 4

    Analysts run Athena queries from several tools, and each tool sets its own query result location. The security team requires every result file to land in one approved S3 bucket encrypted with SSE-KMS, whatever a client sets. Which Athena setting enforces this?

    • A.Set the query result location in each analyst's tool, and document the approved bucket in the team wiki.
    • B.Turn on S3 Object Lambda on the results bucket, so each result file is re-encrypted with KMS as it is read.
    • C.Create a workgroup for each analyst tool, each with its own result location and no settings enforced.
    • D.Create a workgroup with the approved result location and SSE-KMS encryption, with enforcement turned on.
    Show answer and explanation

    Correct answer: D

    An Athena workgroup can define the query result location and encryption. When the workgroup's enforce-configuration setting is on, Athena overrides client-side values, so every tool writes results to the approved bucket with SSE-KMS. The other options rely on each client behaving correctly, or on Lambda, which does not control where Athena writes results.

    Why the other options are wrong

    • A.Client settings can be changed or skipped by each tool, so a written standard does not enforce anything.
    • B.Object Lambda does not control where Athena writes results or which key encrypts them.
    • C.Separate workgroups without enforcement still let each client choose its own result location.
  5. Question 5

    A team added a predictive scaling policy to an EC2 Auto Scaling group a week ago and removed the group's other scaling policies. The forecast in the console matches the real daily traffic pattern closely, yet the group never launches instances ahead of the morning peak. The team also wants the group to shrink each night. Which TWO actions should a solutions architect take? (Choose TWO.)

    • A.Wait until the policy has collected 14 full days of history before it scales.
    • B.Switch the predictive scaling policy from forecast only mode to forecast and scale mode.
    • C.Add a warm pool to the group.
    • D.Increase the group's default cooldown period.
    • E.Add a target tracking policy so the group can scale in when load falls.
    Show answer and explanation

    Correct answer: B and E

    A new predictive scaling policy starts in forecast only mode: it produces forecasts but doesn't change capacity until you switch it to forecast and scale. Predictive scaling also never scales in, so a dynamic policy such as target tracking is needed to remove capacity at night.

    Why the other options are wrong

    • A.Predictive scaling starts forecasting after 24 hours of data; it doesn't need 14 days before it can scale. The forecast already exists, so data isn't the problem.
    • C.A warm pool shortens how long new instances take to become ready, but nothing is launching them in the first place, and it doesn't scale in.
    • D.The default cooldown applies to simple scaling policies and has no effect on whether a predictive policy acts or whether the group scales in.
  6. Question 6

    A log analytics job on one EC2 instance reads 8 TB of log files sequentially in large blocks every night. The data must persist when the instance is stopped. Throughput matters and random IOPS do not. Which EBS volume type is the MOST cost-effective fit?

    • A.Throughput Optimized HDD (st1)
    • B.Provisioned IOPS SSD (io2 Block Express)
    • C.Instance store volumes
    • D.General Purpose SSD (gp3) at maximum IOPS
    Show answer and explanation

    Correct answer: A

    st1 is an HDD volume built for large, sequential, throughput-heavy workloads such as big data and log processing, delivering up to 500 MiB/s per volume. It cannot be a boot volume, so the instance boots from a separate volume, which is the normal layout for a data volume like this one.

    Why the other options are wrong

    • B.io2 Block Express is priced for high random IOPS and sub-millisecond latency, which this workload does not need.
    • C.Instance store data is lost when the instance stops, which breaks the persistence requirement.
    • D.Paying for maximum gp3 IOPS buys random I/O performance the sequential job does not use.
  7. Question 7

    A trading platform runs a self-managed database on a Nitro-based EC2 instance. The database needs 150,000 IOPS on one volume with consistent sub-millisecond latency. Which storage meets these performance requirements?

    • A.A gp3 volume with the maximum provisioned IOPS
    • B.An Amazon EFS file system with Elastic throughput
    • C.An io1 volume with Multi-Attach enabled
    • D.An io2 Block Express volume provisioned with 150,000 IOPS
    Show answer and explanation

    Correct answer: D

    io2 Block Express supports up to 256,000 IOPS and 4,000 MiB/s per volume on Nitro-based instances, with average latency under 500 microseconds. It is the only EBS volume type that meets 150,000 IOPS on a single volume.

    Why the other options are wrong

    • A.gp3 tops out at 80,000 IOPS per volume and offers single-digit millisecond, not sub-millisecond, latency.
    • B.EFS is network file storage with about 1 ms read and 2.7 ms write latency; it is not the right choice for a single database volume that needs sub-millisecond latency.
    • C.io1 tops out at 64,000 IOPS per volume, and Multi-Attach adds sharing, not performance.
  8. Question 8

    A company is deploying a 30-node Apache Cassandra cluster on EC2 in one Availability Zone. The database keeps three replicas of each row and must place them so that a single rack failure can't take out more than one replica. Cassandra must be able to see which hardware group each node runs on. Which solution meets these requirements?

    • A.Launch the nodes into a partition placement group with three partitions.
    • B.Launch the nodes into a spread placement group.
    • C.Launch the nodes into a cluster placement group.
    • D.Launch the nodes on Dedicated Hosts so each node has its own physical server.
    Show answer and explanation

    Correct answer: A

    Partitions in a partition placement group never share racks, and EC2 exposes each instance's partition number. Mapping each partition to a Cassandra rack lets the database place the three replicas in different partitions. The number of instances is limited only by account quotas.

    Why the other options are wrong

    • B.A spread placement group allows at most seven running instances per AZ, so it can't hold 30 nodes in one AZ, and it exposes no partition layout to the application.
    • C.A cluster placement group packs instances close together for low latency, which raises the chance that one hardware failure hits several replicas.
    • D.Dedicated Hosts give physical server isolation for licensing or compliance, but they don't guarantee separate racks or tell Cassandra how nodes are grouped.

Design Cost-Optimized Architectures

  1. Question 1

    An Aurora cluster processes a high volume of small reads and writes. Its monthly Aurora I/O charges are larger than its instance charges. The team wants the LOWEST cost with the same availability. What should it do?

    • A.Compare the bills, then switch the cluster to Aurora I/O-Optimized
    • B.Keep Aurora Standard and add read replicas to spread the I/O across more instances, which keeps the same engine and the same data path
    • C.Move to DynamoDB, because it has no I/O charges
    • D.Keep Aurora Standard and reduce the backup retention period
    Show answer and explanation

    Correct answer: A

    When I/O charges are larger than instance charges, I/O-Optimized removes the largest part of the bill, and the higher instance rate costs less than the I/O it removes. Aurora's guidance is that I/O-Optimized pays off when I/O is 25 percent or more of Aurora spend. Read replicas add instances, and they add their own I/O. DynamoDB is a migration with application changes that the stem does not allow. Backup retention doesn't touch I/O.

    Why the other options are wrong

    • B.Replicas add instances and read I/O, which raises both parts of the bill.
    • C.A move to DynamoDB is a re-architecture, and the stem asks for the same availability without that change.
    • D.Backup retention affects backup storage, not the I/O charges that drive this bill.
  2. Question 2

    A startup's Amazon S3 spend grows every month as customers upload more data. The founders want an email warning before the month's S3 spend is expected to exceed USD 5,000, not after the money is spent. Which solution meets this requirement with the LEAST operational overhead?

    • A.Download the Cost and Usage Report daily and check totals in a spreadsheet.
    • B.Create an Amazon CloudWatch billing alarm on the estimated charges for Amazon S3 at USD 5,000.
    • C.Turn on AWS Cost Anomaly Detection for Amazon S3 with email alerts.
    • D.Create an AWS Budgets cost budget filtered to Amazon S3 with an alert on forecasted spend.
    Show answer and explanation

    Correct answer: D

    AWS Budgets can alert on forecasted spend, before costs accrue, as well as on actual spend. A monthly cost budget filtered to Amazon S3 with a forecast-based email alert warns the founders automatically.

    Why the other options are wrong

    • A.The Cost and Usage Report is detailed data for analysis; checking it by hand every day is high operational overhead.
    • B.A CloudWatch billing alarm triggers only when current charges exceed the threshold; it doesn't use projections, so the warning comes after USD 5,000 is already spent.
    • C.Cost Anomaly Detection flags unusual spend patterns and accounts for natural growth; it doesn't warn against a fixed USD 5,000 forecast threshold.
  3. Question 3

    A ticketing company expects a surge for a three-week concert presale. It needs to be certain it can launch 30 c7i.2xlarge instances in one specific Availability Zone, next to its existing database, for those three weeks only. It already has a Compute Savings Plan that covers part of its usage. Which approach meets the requirement MOST cost-effectively?

    • A.Buy 30 zonal Standard Reserved Instances on a 1-year term, which reserves the capacity in that AZ.
    • B.Request 30 Spot Instances with a persistent request, so they relaunch whenever capacity returns.
    • C.Create an On-Demand Capacity Reservation for 30 instances in that AZ and cancel it after the event.
    • D.Buy an additional Compute Savings Plan sized to 30 instances, so the capacity is covered at a discount.
    Show answer and explanation

    Correct answer: C

    On-Demand Capacity Reservations reserve capacity in one Availability Zone for any duration with no long-term commitment, and you can cancel them when the event ends. They are billed at the On-Demand rate, and the existing Savings Plan can discount them.

    Why the other options are wrong

    • A.Zonal RIs do reserve capacity, but a 1-year term pays for about 49 weeks the company doesn't need.
    • B.Spot gives no capacity guarantee and can be interrupted, which fails 'be certain'.
    • D.Savings Plans are only a billing discount. They never reserve capacity, and the new commitment would run for at least a year.
  4. Question 4

    A company runs a fleet of x86 Amazon EC2 instances for a Java microservice that has no native x86 dependencies. It also runs dozens of Python AWS Lambda functions. The company wants better price performance for both with minimal code changes. Which approach should it take?

    • A.Move the microservice to larger x86 instance sizes and raise the memory setting of every Lambda function.
    • B.Move the microservice to Dedicated Hosts and turn on provisioned concurrency for the Lambda functions.
    • C.Buy EC2 Instance Savings Plans for the current x86 family, and leave the Lambda functions unchanged.
    • D.Test, then move the microservice to Graviton (arm64) instance types and switch the functions to arm64.
    Show answer and explanation

    Correct answer: D

    Java and interpreted Python run on Arm without code changes when there are no native x86 dependencies. Graviton-based EC2 types and the arm64 Lambda architecture typically give better price performance than their x86 equivalents, so a tested switch is the direct lever.

    Why the other options are wrong

    • A.Bigger instances and more memory raise cost; they only improve performance per dollar if the workload is starved, which the stem doesn't say.
    • B.Dedicated Hosts are for licensing and compliance, and provisioned concurrency adds cost to remove cold starts. Neither improves price performance.
    • C.A commitment discount on x86 helps cost but locks the fleet to that family, and it does nothing for the Lambda functions.
  5. Question 5

    An IoT platform writes about 50 million sensor readings per month to Amazon S3 as separate JSON objects of roughly 4 KB each. Readings older than 90 days are kept for 5 years only for rare investigations, where a delay of 48 hours is acceptable. The team's lifecycle rule to move objects to S3 Glacier Deep Archive after 90 days is not moving the objects at all. What should a solutions architect do to reduce storage cost MOST effectively?

    • A.Combine each day's readings into larger compressed files, archive those to S3 Glacier Deep Archive, and expire the originals.
    • B.Add an object size filter so the lifecycle rule also transitions objects smaller than 128 KB to S3 Glacier Deep Archive.
    • C.Change the lifecycle rule to transition the readings to S3 Standard-IA after 30 days and keep them there for the full 5 years.
    • D.Move the readings to S3 Intelligent-Tiering and activate the Deep Archive Access tier so that idle objects archive by themselves.
    Show answer and explanation

    Correct answer: A

    By default, S3 Lifecycle does not transition objects smaller than 128 KB, because a transition request is billed per object and small objects save little. Archived objects also carry about 40 KB of extra billed metadata each, which is ten times the size of a 4 KB reading. Bundling readings into larger files cuts both the request count and the metadata overhead, so archiving becomes worthwhile.

    Why the other options are wrong

    • B.A size filter makes the transition happen, but each 4 KB object then pays a transition request plus about 40 KB of archive metadata, which can cost more than it saves.
    • C.S3 Standard-IA bills each object as at least 128 KB, so 4 KB objects would be billed 32 times their size.
    • D.S3 Intelligent-Tiering does not monitor or tier objects under 128 KB; they always stay in the Frequent Access tier.
  6. Question 6

    A steady ElastiCache cluster serves a web application. The team wants to lower the cache's cost without reducing its hit rate or the node count it needs. Which TWO actions do this? (Choose TWO.)

    • A.Add nodes so each one carries less load
    • B.Move from reserved nodes to On-Demand nodes to keep the cluster flexible
    • C.Commit to ElastiCache reserved nodes for the steady node count
    • D.Remove the cache and add read replicas to the database instead
    • E.Choose Valkey over Redis OSS for the same node type
    Show answer and explanation

    Correct answer: C and E

    Reserved nodes discount the steady node count. Valkey nodes are priced below the equivalent Redis OSS nodes, so the engine choice lowers the rate without changing the node count. More nodes add cost. On-Demand removes the discount. Replicas add database instances, and they usually cost more than the cache they replace.

    Why the other options are wrong

    • A.More nodes add cost, and each one is a running charge.
    • B.On-Demand nodes give up the reserved discount on a steady count.
    • D.Read replicas are billed database instances, and they typically cost more than the cache they would replace.
  7. Question 7

    A branch office sends about 200 GB a month to AWS and needs an encrypted link set up within a day. No steady high-volume link is planned. Which option is the MOST cost-effective?

    • A.AWS Direct Connect with a 10 Gbps port.
    • B.A transit gateway with a Direct Connect gateway.
    • C.Interface endpoints for the application.
    • D.A Site-to-Site VPN.
    Show answer and explanation

    Correct answer: D

    A Site-to-Site VPN is encrypted by design, can be set up quickly, and charges per connection-hour plus internet data transfer. At 200 GB a month, that is far below the cost of a dedicated port and its commitment. Direct Connect is not encrypted by itself and needs provisioning time.

    Why the other options are wrong

    • A.A 10 Gbps port has a large port-hour charge and needs provisioning time, far more than 200 GB a month justifies.
    • B.A transit gateway and Direct Connect gateway add hourly charges on top of a link this small.
    • C.Interface endpoints reach AWS services, not a branch office.