SAA-C03 domain - 24% of the exam

Design High-Performing Architectures

Design High-Performing Architectures is 24% of the SAA-C03 exam. These are the objectives it covers, each with practice questions, with every answer explained.

The domain in numbers

86
Practice questions
9
Objectives
64
Medium
22
Hard

Where people go wrong

  • Paying for size or IOPS that gp3 already provides

    The wrong answers meet an IOPS or throughput target by growing a gp2 volume, moving to io2 or io1, or striping volumes, and each adds cost the requirement rules out. On gp2 the baseline is tied to capacity, so growing the volume buys storage nobody needs. gp3 provisions IOPS and throughput independently of size, such as 9,000 IOPS and 350 MiB/s on a 400 GiB volume. It also includes a free baseline of 3,000 IOPS and 125 MiB/s, so a workload inside that needs no change.

    See it in a sample question
  • Swapping cluster and spread placement groups

    The wrong answers attach each placement group to the other's goal. A spread placement group puts every instance on distinct hardware for fault isolation, so offering it to a tightly coupled HPC job only increases the network distance between nodes. A cluster placement group packs instances close together in one Availability Zone for low latency, so offering it to independent quorum members raises the chance that one hardware failure takes down several of them.

    See it in a sample question
  • Counting a standby as read capacity

    The wrong answers offer a Multi-AZ standby or an Aurora failover target as a way to offload reads. A standby exists for availability and does not serve application reads while the primary is healthy, so it adds no read capacity. RDS read replicas, or Aurora Replicas behind the cluster reader endpoint, add queryable copies. ElastiCache is the answer only when the same items are read repeatedly and can be served from memory.

    See it in a sample question
  • Swapping CloudFront, Global Accelerator and Transfer Acceleration

    The wrong answers give each edge service a job it does not do. CloudFront is offered for personalised, uncacheable API responses and for UDP game traffic, but it accelerates cacheable HTTP and HTTPS content. Global Accelerator suits those cases: static anycast IP addresses, TCP and UDP carried over the AWS backbone to the nearest healthy endpoint. CloudFront is also offered for slow long-distance uploads, which S3 Transfer Acceleration speeds into one bucket through an accelerate endpoint. Transfer Acceleration cannot front a load-balanced API.

    See it in a sample question
  • Expecting VPC peering to route transitively

    The wrong answers rely on VPC peering to link many networks. Peering is non-transitive, so a hub VPC cannot forward traffic between two spokes, and on-premises traffic arriving in one Region cannot cross a peering link to another. A full mesh of peering connections also grows quickly to maintain. A Transit Gateway acts as a central hub with transitive routing, shared across accounts with Resource Access Manager, and a Direct Connect gateway lets one private virtual interface reach VPCs in several Regions.

    See it in a sample question

Objectives in this domain

What to study

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

Sample questions from this domain

Free sampleDesign High-Performing Architecturesmedium

A company runs a transactional workload on an EC2 instance backed by a 400 GiB General Purpose SSD (gp2) volume. Monitoring shows the application needs a sustained 9,000 IOPS and 350 MiB/s of throughput, but the gp2 volume delivers only its size-derived baseline and the team does not want to grow the volume just to gain performance. A solutions architect must raise IOPS and throughput at the lowest cost without changing the volume size. Which action meets the requirement?

  • AIncrease the gp2 volume size to 3,000 GiB so the size-derived baseline reaches the required IOPS, then enable an instance burst balance to cover throughput peaks.
  • BMigrate the volume to Provisioned IOPS SSD (io2) and provision 9,000 IOPS, accepting the higher per-IOPS price for the guaranteed performance the workload needs.
  • CAttach a second gp2 volume of the same size and stripe the two in a RAID 0 array so the combined baseline meets the IOPS and throughput target.
  • DMigrate the volume to General Purpose SSD (gp3) and provision the 9,000 IOPS and 350 MiB/s throughput independently of the 400 GiB capacity. Correct
gp3 lets you provision IOPS and throughput independently of capacity, meeting performance targets without paying for extra volume size. General Purpose SSD gp3 separates provisioned IOPS and throughput from the volume capacity, unlike gp2 where the baseline scales with size. On a 400 GiB gp3 volume the team can provision up to the required 9,000 IOPS and 350 MiB/s directly, and gp3 carries a lower per-GiB price than gp2, so it raises performance at the lowest cost with no change to the volume size.

Why A is wrong: Growing a gp2 volume to lift baseline IOPS forces the team to pay for unused capacity, which directly contradicts the requirement to avoid increasing the volume size.

Why B is wrong: io2 can supply the IOPS but it costs more per provisioned IOPS than gp3, so it fails the lowest-cost requirement when gp3 already meets the target.

Why C is wrong: RAID 0 striping doubles cost and adds management effort while keeping the size-tied gp2 model, making it more expensive and complex than provisioning gp3 directly.

Why D is correct: gp3 decouples IOPS and throughput from capacity, so the team can provision exactly 9,000 IOPS and 350 MiB/s on the existing 400 GiB at lower cost than gp2.

Free sampleDesign High-Performing Architecturesmedium

A high-performance computing job runs across a fleet of EC2 instances that exchange large messages between nodes and are highly sensitive to inter-node network latency and throughput. The team wants to place these instances so they sit physically close together on the network to maximise packet-per-second performance within a single Availability Zone. Which placement strategy should they choose?

  • ALaunch the instances in a spread placement group so each one runs on distinct underlying hardware to reduce the chance of correlated failures.
  • BLaunch the instances in a cluster placement group so they are packed close together in one Availability Zone for low-latency, high-throughput networking. Correct
  • CLaunch the instances in a partition placement group so they are divided into separate racks for large distributed and replicated workloads.
  • DLaunch the instances across multiple Availability Zones behind a Network Load Balancer so the cluster gains both low latency and zone resilience.
A cluster placement group packs instances close together in one Availability Zone for the lowest latency and highest throughput between nodes. Cluster placement groups co-locate instances on hardware that is physically near each other within a single Availability Zone, giving the high packet-per-second performance and low latency that tightly coupled HPC traffic needs; spread and partition groups instead separate instances for fault isolation and so increase inter-node distance.

Why A is wrong: A spread placement group deliberately separates instances onto different hardware for fault isolation, which is the opposite of packing nodes close together for the lowest possible inter-node latency.

Why B is correct: A cluster placement group packs instances onto hardware close together within a single Availability Zone, delivering the lowest network latency and highest packet-per-second throughput between nodes, which is exactly what tightly coupled HPC traffic requires.

Why C is wrong: A partition placement group spreads instances across distinct racks to limit the blast radius of a hardware failure for big data systems; it does not pack nodes tightly for the lowest latency this HPC job needs.

Why D is wrong: Spreading across Availability Zones adds inter-zone network distance and latency between nodes, which harms the tightly coupled message passing this workload depends on, and a load balancer does not reduce node-to-node latency.

Free sampleDesign High-Performing Architecturesmedium

A reporting and analytics team runs heavy SELECT queries against a production Amazon Aurora MySQL cluster, and those long-running reads are slowing down the transactional traffic on the writer. The team wants to move the reporting reads off the writer onto separate capacity within the same Region, while keeping a single managed cluster. Which approach best meets this requirement?

  • AAdd Aurora Replicas to the cluster and point the reporting queries at the cluster reader endpoint so analytics reads run on replica capacity separate from the writer. Correct
  • BInsert an Amazon ElastiCache cluster in front of Aurora so the reporting queries are answered from memory instead of touching the writer instance directly.
  • CCreate an Amazon Aurora global database with a secondary Region and run the reporting queries against the remote secondary cluster to relieve the writer.
  • DEnable an Aurora Multi-AZ standby instance and route the reporting queries to that standby so they no longer compete with traffic on the writer.
Aurora Replicas behind the reader endpoint offload read-heavy queries from the writer within the same Region using one managed cluster. Aurora Replicas attach to the same shared distributed storage volume as the writer and are reached through the cluster reader endpoint, so steering analytical SELECTs to that endpoint runs them on dedicated replica instances; the writer is then left to process transactional writes and short reads without competing for CPU and I/O.

Why A is correct: Aurora Replicas share the cluster storage volume and serve reads through the reader endpoint, so directing reporting queries there moves the heavy SELECT load onto separate replica instances and frees the writer for transactional traffic within one managed cluster.

Why B is wrong: ElastiCache suits repeated identical lookups, but ad hoc analytical SELECTs with varied filters cache poorly and would mostly miss, so it does not reliably offload these long-running, low-repeat reporting queries from the writer.

Why C is wrong: A global database does add read capacity, but it spans Regions for disaster recovery and adds cross-Region replication and cost, which exceeds the requirement to keep the workload in one Region and one cluster.

Why D is wrong: An Aurora failover target exists for availability and does not serve application reads while the writer is healthy, so reporting queries cannot be routed to it to offload the writer.

Free sampleDesign High-Performing Architecturesmedium

A media company lets users in Asia, Europe, and South America upload large video files of several gigabytes each directly into a single Amazon S3 bucket hosted in us-east-1. Users far from that Region report slow, inconsistent upload speeds over the public internet. The team wants to speed up these long-distance uploads to the existing bucket with the least change to the client application. Which approach best improves upload performance?

  • APut an Amazon CloudFront distribution in front of the bucket and let clients download cached objects from the nearest edge location to cut latency.
  • BCreate read replica buckets in eu-west-1 and ap-southeast-1 using S3 Cross-Region Replication and have nearby users upload into the closest replica bucket.
  • CIncrease the multipart upload part size to a few gigabytes so each large file is sent as a single part and fewer network round trips are required.
  • DEnable S3 Transfer Acceleration on the bucket and have clients upload through the bucket's accelerate endpoint, which routes data over the AWS edge network. Correct
Use S3 Transfer Acceleration to speed long-distance uploads to a single bucket over the AWS edge network with minimal client change. S3 Transfer Acceleration routes client uploads to the nearest CloudFront edge location and carries them to the bucket over the optimised AWS backbone, improving throughput for geographically distant clients while requiring only a switch to the bucket accelerate endpoint.

Why A is wrong: CloudFront caching accelerates downloads of repeatedly read objects, not the large inbound uploads that are the stated bottleneck here.

Why B is wrong: Cross-Region Replication copies objects after they are written and is one-directional, so it cannot accelerate the inbound upload path users complain about.

Why C is wrong: Oversized single parts remove the parallelism that speeds multipart uploads and do nothing about the distance latency causing the slow transfers.

Why D is correct: Transfer Acceleration uses CloudFront edge locations and the AWS backbone to speed distant uploads to the same bucket with only an endpoint change.

Free sampleDesign High-Performing Architectureshard

A company operates 18 VPCs spread across three AWS accounts and plans to add roughly 10 more over the next year. Every VPC must be able to route to every other VPC, and the networking team wants to avoid managing a growing number of point-to-point connections and overlapping route table entries as VPCs are added. Which approach scales connectivity with the least ongoing operational effort?

  • ACreate a full mesh of VPC peering connections between every pair of VPCs and add explicit routes in each VPC route table for all other VPC CIDR ranges.
  • BEstablish AWS Direct Connect links between the accounts so that the dedicated circuits carry inter-VPC traffic and propagate routes automatically across all VPCs.
  • CConfigure AWS PrivateLink endpoints in each VPC so that interface endpoints expose every other VPC as a consumable service for fully routed connectivity.
  • DDeploy an AWS Transit Gateway as a central hub, attach every VPC to it, and use Resource Access Manager to share the gateway so all accounts route through one managed attachment point. Correct
AWS Transit Gateway provides transitive hub-and-spoke routing that scales connectivity across many VPCs and accounts without a peering mesh. VPC peering is non-transitive, so connecting many VPCs needs a full mesh whose connection and route count grows quadratically. Transit Gateway acts as a single regional hub offering transitive routing to all attached VPCs, scales to thousands of attachments, and shares across accounts via RAM, so adding a VPC means one attachment rather than many new peerings.

Why A is wrong: VPC peering is non-transitive and a full mesh grows at roughly n squared, so dozens of connections and route entries must be maintained by hand as VPCs are added, which is the operational burden to avoid.

Why B is wrong: Direct Connect provides private connectivity between on-premises networks and AWS, not a hub for routing traffic among many VPCs, so it does not solve VPC-to-VPC scaling.

Why C is wrong: PrivateLink exposes a specific service endpoint rather than full bidirectional VPC routing, so it cannot provide general any-to-any connectivity across dozens of growing VPCs.

Why D is correct: Transit Gateway is a regional hub that gives transitive any-to-any routing, scales to thousands of attachments, and can be shared across accounts with RAM, removing the mesh of peering connections and per-pair routes.

Other domains in this exam

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

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