SAA-C03
Free SAA-C03 practice questions
10 questions from our SAA-C03 practice tests, each with the correct answer and a written explanation. Read them with the answers shown, or switch to quiz mode and answer first.
Question 1 of 10, Design Secure Architectures
A multinational company uses AWS Organizations. Legal requires that workloads in the Production OU run only in eu-west-1 and eu-central-1, but administrators must still be able to use global services such as IAM and Route 53. Which approach enforces this with the LEAST ongoing effort?
- ACreate a customer managed IAM policy in each production account that allows actions only in the two Regions and attach it to all existing users and roles
- BAttach an SCP to the Production OU that denies actions when aws:RequestedRegion is outside the two Regions, using NotAction to exempt global services Correct answer
- CEnable the AWS Control Tower Region deny control on the landing zone, then rely on drift detection emails to catch accounts enrolled later outside Control Tower
- DDeploy AWS Config rules in every Region with automatic remediation that deletes any resource created outside eu-west-1 and eu-central-1
Answer and explanation: An SCP that denies on aws:RequestedRegion (with NotAction for global services) is a preventive guardrail for every principal in the OU, including future accounts and roles, with nothing to maintain per account. Per-account IAM policies must be attached to every new principal and are bypassed by anyone who can create a role. Config remediation is detective: resources are created first and then deleted. Relying on drift emails for accounts outside Control Tower leaves gaps and needs manual follow-up, so it is not the least-effort enforcement.
Question 2 of 10, Design Secure Architectures
A company must share encrypted EBS snapshots of a database server with a partner's AWS account for forensic analysis. The snapshots are currently encrypted with the AWS managed key aws/ebs. What must the solutions architect do so the partner can create volumes from them?
- AShare the existing snapshots with the partner account and add the partner's account ID as a principal in the aws/ebs key policy
- BUse the EBS direct APIs to read the snapshot blocks and write them to an S3 bucket encrypted with SSE-S3 that grants the partner read access
- CCreate a launch-ready AMI from each snapshot, share the AMIs with the partner account, and let the partner's account re-encrypt them on copy
- DCopy the snapshots re-encrypted with a customer managed KMS key, allow the partner account in that key's policy, and share the copies Correct answer
Answer and explanation: Snapshots encrypted with the AWS managed aws/ebs key cannot be shared because that key policy cannot be edited; copying them under a customer managed key and granting the partner account use of the key makes sharing possible. The aws/ebs key policy cannot be modified. AMIs backed by aws/ebs-encrypted snapshots have the same restriction, so they cannot be shared either. Dumping blocks to S3 with the direct APIs is a custom export the partner cannot turn back into volumes without building tooling.
Question 3 of 10, Design Secure Architectures
Developers in a sandbox account need to create IAM roles for their Lambda functions. Security is concerned that a developer could create a role with more permissions than the developer has and escalate privileges. Which control addresses this concern?
- AAttach the ReadOnlyAccess managed policy to developers and have the platform team create each Lambda execution role on request through a ticket
- BAllow role creation only when an approved permissions boundary is attached, enforced with an iam:PermissionsBoundary condition Correct answer
- CRestrict developers to the AWSLambdaBasicExecutionRole managed policy and use an SCP to block lambda:CreateFunction unless the role name starts with dev-
- DGrant developers iam:* in the sandbox and enable CloudTrail with an EventBridge rule that notifies security whenever CreateRole or AttachRolePolicy is called
Answer and explanation: Requiring an approved permissions boundary (enforced with the iam:PermissionsBoundary condition key) caps the maximum permissions of any role developers create, so they can build Lambda roles without escalating. CloudTrail notifications are detective and only alert after an over-privileged role exists. Read-only access with ticketed role creation blocks the self-service requirement. A role-name prefix check does not limit what policies are attached to the role, so escalation is still possible.
Question 4 of 10, Design Resilient Architectures
A team runs Amazon RDS for PostgreSQL as a single-standby Multi-AZ DB instance. They want faster failovers, typically under 35 seconds, and the ability to send read traffic to standby capacity without managing separate read replicas. Which deployment option should they choose?
- AMigrate to Aurora PostgreSQL Serverless v2 with a single writer and no reader instances
- BKeep the Multi-AZ DB instance, move to a larger instance class, and add an in-Region read replica for reads
- CA Multi-AZ DB cluster with one writer and two readable standby instances in different AZs Correct answer
- DAdd a cross-Region read replica and promote it if the primary Availability Zone fails
Answer and explanation: RDS Multi-AZ DB clusters have two readable standbys in separate AZs using semi-synchronous replication, giving typical failovers under 35 seconds and a reader endpoint. A cross-Region replica is for DR and needs manual promotion. A larger Multi-AZ instance keeps a non-readable standby, and the added read replica is exactly the separate replica the team wants to avoid. Aurora without reader instances has no standby capacity to read from or fail over to quickly.
Question 5 of 10, Design Resilient Architectures
A clickstream pipeline must allow three independent consumer applications to read the same records, and a new consumer added next month must be able to reprocess the last 30 days of data. Which service should the solutions architect choose?
- AAmazon MQ for RabbitMQ with a fanout exchange and durable queues per consumer
- BAmazon Kinesis Data Streams with retention extended to 30 days or more Correct answer
- CAn SNS topic fanning out to three SQS queues, with a new queue subscribed when the next consumer arrives
- DAn SQS standard queue with a 14-day retention period and three separate consumer groups
Answer and explanation: Kinesis Data Streams lets multiple consumers read the same records independently and supports retention up to 365 days, so a new consumer can replay 30 days. SQS has no consumer groups; a message is processed by one consumer and retention maxes out at 14 days. An SNS fan-out queue subscribed next month receives only new messages, not the last 30 days. RabbitMQ queues likewise deliver only messages published after the queue exists.
Question 6 of 10, Design Resilient Architectures
An Aurora PostgreSQL cluster has a writer and three Aurora Replicas: one db.r6g.large reporting replica and two db.r6g.4xlarge replicas sized like the writer. During a failover test, Aurora promoted the small reporting replica, causing performance problems. How should the architect prevent this?
- AGive the large replicas failover tier 0 or 1 and the reporting replica tier 15 Correct answer
- BMove the reporting workload to an Aurora cross-Region replica so it is not a failover candidate in this Region
- CEnable Multi-AZ on each large replica individually so Aurora prefers Multi-AZ instances during failover
- DDisable automatic failover on the cluster and promote a large replica manually during outages
Answer and explanation: Aurora promotes the replica with the highest priority (lowest tier number), preferring the largest instance among ties, so setting tiers fixes the selection. Aurora failover cannot be disabled, and manual promotion would lengthen outages. A cross-Region replica would add latency for reporting and change the architecture, which is unnecessary when tiers solve it. Multi-AZ is a cluster-level property in Aurora, not a per-replica setting.
Question 7 of 10, Design High-Performing Architectures
Five separate applications read from the same Kinesis data stream using shared-throughput polling, and read latency has climbed to several seconds because they compete for each shard's 2 MB/s read limit. Each application needs low-latency, independent reads. What should be done?
- AIncrease the shard count fivefold so each application gets more read capacity
- BHave producers aggregate records with the KPL to reduce the number of reads
- CRegister each application as an enhanced fan-out consumer Correct answer
- DHave all applications share a single KCL application name
Answer and explanation: Enhanced fan-out gives each registered consumer its own 2 MB/s per shard via push (SubscribeToShard) with latency around 70 ms. Adding shards raises cost and the apps still share each shard's 2 MB/s and five-reads-per-second limits with polling. Record aggregation does not reduce contention between applications. Sharing a KCL application name splits shards among the apps rather than letting each read all data.
Question 8 of 10, Design High-Performing Architectures
A global news application uses Aurora MySQL in us-east-1. Readers in Asia and Europe experience high latency on read-heavy pages. The company wants low-latency local reads in those Regions with minimal replication lag, typically under a second. What should be used?
- AElastiCache for Redis OSS in us-east-1 with Global Datastore
- BAurora MySQL cross-Region binlog read replicas in each Region
- CCross-Region snapshot copies restored into new clusters each night
- DAurora Global Database with secondary clusters in Asia and Europe Correct answer
Answer and explanation: Aurora Global Database replicates at the storage layer with typical lag under one second, and secondary clusters serve local reads. Nightly snapshots are stale. Binlog-based cross-Region replicas work but typically have higher, variable lag and add load on the writer. A cache in us-east-1 still incurs cross-continent latency, and Global Datastore adds a cache layer without replicating the database.
Question 9 of 10, Design Cost-Optimized Architectures
An analytics team runs heavy Amazon Redshift queries only during month-end close, about 4 days a month. The provisioned cluster sits idle the rest of the time. Which option reduces cost with the least operational effort?
- AUnload the data to S3 and use Athena only during month-end close
- BMove to Amazon Redshift Serverless, billed only for compute used while queries run Correct answer
- CPause the provisioned cluster after each close with a scheduled action and resize it with elastic resize before the next one
- DBuy 1-year Reserved Nodes for the existing cluster
Answer and explanation: Redshift Serverless provisions and scales compute automatically and bills per RPU-hour only while workloads run, fitting infrequent heavy use with no scheduling to manage. Reserved Nodes pay for an always-on cluster. Pause/resume plus resize schedules can cut cost but require ongoing operational management. Rewriting the warehouse onto Athena is a migration with more effort and may not match Redshift performance for heavy close queries.
Question 10 of 10, Design Cost-Optimized Architectures
A company runs several hundred Python Lambda functions with steady traffic. The code has no architecture-specific native dependencies. What simple change can lower the Lambda compute bill?
- AMove the functions into a VPC so they avoid data transfer charges
- BRaise every function to 10,240 MB so invocations finish faster and cost less
- CSwitch the functions to the arm64 (Graviton) architecture Correct answer
- DEnable provisioned concurrency on every function to get the lower per-GB-second rate
Answer and explanation: Lambda on arm64 is billed at a lower price per GB-second and often performs better, and pure Python code typically runs unchanged. Maximum memory raises the price per millisecond far more than it shortens most functions. Provisioned concurrency has a lower duration rate but adds a charge for keeping environments ready, which does not lower the bill across hundreds of functions. VPC placement does not reduce Lambda pricing.