EC2 placement groups: SAA-C03 performance scenarios

Cluster for latency, spread for isolation, partition for scale — and the blast-radius trade-off the exam keeps testing. Seven scenarios, SAA-C03, 2026 edition.

Try 7 free EC2 placement-group scenarios in the exam's own style — every answer explained, no sign-up.

Straight to the 7 free questions ↓

EC2 placement groups are a small topic with an outsized presence in the Design High-Performing Architectures domain, because they force the exam’s favourite tension into one screen: performance versus fault isolation. Three strategies, three different jobs — and the distractors are always the other two strategies offered where they do not belong. The diagram lays the three side by side; match the requirement word to the right panel and the answer follows.

Cluster: proximity, and its price

A cluster placement group packs instances onto hardware physically close together within one Availability Zone. That proximity is what delivers the lowest node-to-node latency and the highest throughput, so it is the answer whenever a scenario describes a tightly-coupled workload — HPC, MPI, big analytics shuffles — where inter-node network speed is the bottleneck. The catch is the one the exam loves to punish: everything sits in one AZ on nearby racks, so a single failure has a large, correlated blast radius. Cluster is a performance choice, never an availability one. If a stem pairs “cluster placement group” with “must survive failures”, that is the trap.

Spread: distinct hardware for a critical few

A spread placement group does the opposite — it puts each instance on separate underlying hardware, with separate power and network, so one hardware fault can affect at most one instance. It is built for a small number of individually critical instances: a licence server, a primary broker, a domain controller. The number to know is the ceiling: at most seven running instances per Availability Zone per spread group, because each needs its own rack. When a scenario wants distinct-hardware isolation for a handful of instances, it is spread — and when it asks for more than seven in one AZ, that ceiling is the constraint being tested.

Partition: rack isolation at scale

A partition placement group splits instances into partitions, each on its own set of racks that share no hardware with the others, and it exposes which partition each instance is in. That does two things a spread group cannot: it scales to many instances, and it lets a topology-aware application place replicas so no two copies land in the same failure boundary. This is why large distributed systems — Cassandra, HDFS, Kafka — use it: the loss of a rack is confined to one partition, and a well-placed replica set survives.

The requirement-matching method

Every placement-group question resolves to one of three tells:

Lowest latency / highest throughput between coupled nodes → cluster (accept the single-AZ blast radius).
A few critical instances that must not share hardware → spread (remember the seven-per-AZ cap).
Many nodes, rack-level fault tolerance, visible topology → partition.

When two options look plausible, find the requirement word the wrong one ignores — a spread answer offered for 200 nodes fails on the count; a cluster answer offered for a resilient tier fails on availability.

Enhanced networking as a complement

Placement is only half of the performance story. For the most demanding HPC and MPI traffic, a cluster placement group pairs with an Elastic Fabric Adapter (EFA) — enhanced networking that lowers latency and raises packet-per-second performance beyond a standard interface. The exam links cluster placement plus EFA as the combination for tightly-coupled traffic; an Application Load Balancer or a second Region is a distractor from a different problem.

Using the seven scenarios

Answer each scenario below before opening the explanation, and when you miss, name the requirement word that decided it — “latency”, “distinct hardware”, “many nodes”, “survive a rack”. Every explanation takes the other two strategies apart so you see why they miss. For the wider domain, see Design High-Performing Architectures.

Three panels contrasting EC2 placement-group strategies. Cluster: instances packed onto one rack in a single Availability Zone for lowest network latency and highest throughput, with a note that a rack failure takes them all down. Spread: a small number of instances each placed on distinct underlying hardware, up to seven per Availability Zone, so no two share a failure. Partition: groups of instances assigned to separate racks or partitions, each partition its own failure boundary, sized for large distributed systems.

7 free SAA-C03 practice questions

Answers and explanations — no email wall

SAA-C03 Question 1 of 7

A tightly-coupled HPC simulation runs across dozens of EC2 instances that exchange large messages constantly, and the bottleneck is node-to-node network latency. Which placement strategy gives the lowest latency and highest throughput between the instances?

Answer: A — A cluster placement group, which packs the instances close together in one Availability Zone.

A cluster placement group packs instances onto hardware physically close together within a single Availability Zone, giving the lowest inter-node latency and the highest per-flow throughput — exactly what a chatty, tightly-coupled HPC job needs. Why not the others: a spread group deliberately separates instances onto distinct hardware, which raises latency, the opposite of the requirement; a partition group isolates racks for fault tolerance, not proximity; and letting instances land across Availability Zones adds cross-AZ latency. When the requirement word is latency or throughput between coupled nodes, it is cluster.

SAA-C03 Question 2 of 7

An architect proposes putting a stateless web tier that must survive infrastructure failures into a cluster placement group for speed. Why is this the wrong choice for availability?

Answer: A — A cluster group concentrates instances in one Availability Zone on nearby hardware, so a single failure can take them all down together.

Cluster placement trades resilience for proximity: the instances share a single Availability Zone and nearby hardware, so a rack or AZ event can hit the whole group at once — a large, correlated blast radius that is wrong for a tier whose whole point is surviving failure. Why not the others: the seven-per-AZ cap belongs to spread groups, not cluster; placement groups do not impose any replication model; and a cluster group is single-AZ, never multi-Region. For availability, spread the tier across Availability Zones instead of packing it.

SAA-C03 Question 3 of 7

A team runs a handful of individually critical instances — a licence server, a primary broker, a KMS proxy — and needs assurance that no two ever share the same underlying host, so one hardware fault cannot take out more than one. Which placement group fits?

Answer: A — A spread placement group, which places each instance on distinct hardware.

A spread placement group puts each instance on separate physical hardware with separate power and network, so a single host or rack failure can affect at most one — the right tool for a small set of individually critical instances. Why not the others: a cluster group does the opposite, packing them together; a single partition holds instances that share a partition's hardware, so they are not isolated from one another; and one instance per cluster group gives no isolation guarantee and misuses the construct. Distinct-hardware isolation for a few critical instances is spread.

SAA-C03 Question 4 of 7

An architect designs a spread placement group but needs to guarantee distinct hardware for 15 critical instances within a single Availability Zone. What is the constraint they will hit?

Answer: A — A spread placement group supports at most seven instances per Availability Zone, so 15 will not fit in one.

A spread placement group is capped at seven running instances per Availability Zone, because each must sit on its own distinct rack — so 15 instances in one AZ exceeds the limit and the launches will fail. The fix is to spread across multiple Availability Zones or reconsider the design. Why not the others: spread groups do not restrict you to a single instance type, do not require dedicated hosts, and are not bound to one subnet. The seven-per-AZ ceiling is the number the exam expects you to know for spread.

SAA-C03 Question 5 of 7

A large distributed data platform — think a Cassandra ring or an HDFS cluster with dozens of nodes — needs its replicas placed so that the loss of one rack cannot take out multiple replicas of the same data. Which placement strategy is designed for this?

Answer: A — A partition placement group, which spreads instances across separate partitions on distinct racks.

A partition placement group divides instances into partitions, each on its own set of racks that share no hardware with other partitions, and exposes the partition each instance is in — so a rack failure is confined to one partition and a well-placed replica set survives. It scales to far more instances than a spread group, which is why big distributed systems like Cassandra, HDFS and Kafka use it. Why not the others: spread groups top out at seven per AZ and are for a few critical instances, not dozens of nodes; a cluster group maximises the blast radius rather than isolating racks; and no placement group gives no rack-isolation guarantee at all.

SAA-C03 Question 6 of 7

A requirement reads: 'up to 200 nodes, must tolerate the loss of a rack without losing a quorum, rack topology must be visible to the application for replica placement.' Which construct matches, and why not the alternatives?

Answer: A — A partition placement group — it scales to many instances, isolates failures per partition, and surfaces partition metadata for topology-aware placement.

The requirement has three tells — many nodes, rack-level fault tolerance, and visible topology for replica placement — and only a partition placement group answers all three: it supports large fleets, gives each partition an independent hardware footprint, and reports each instance's partition so the application can place replicas across partitions. Why not the others: a spread group caps at seven per AZ, far short of 200; a cluster group concentrates rather than isolates; and plain multi-AZ Auto Scaling gives no rack-level partition metadata for the application to reason about. Match every requirement word, not just one.

SAA-C03 Question 7 of 7

A cluster placement group already gives instances low-latency proximity, but an MPI workload needs even lower latency and higher packet-per-second performance for its inter-node traffic. What complements the placement group to achieve this?

Answer: A — Attach an Elastic Fabric Adapter (EFA) to enhance networking for tightly-coupled traffic.

An Elastic Fabric Adapter is an enhanced-networking device that gives HPC and MPI workloads lower and more consistent latency and higher throughput than standard ENIs, and it pairs with a cluster placement group — proximity plus EFA is the combination the exam links to tightly-coupled traffic. Why not the others: a spread group increases separation and works against low latency; an Application Load Balancer is for distributing application requests, not accelerating node-to-node MPI traffic; and a second Region adds latency and solves a resilience problem the question did not ask about. Cluster placement plus EFA is the performance answer.

That is exactly how every question in the course works — answer, explanation, why-not. The real set continues in the practice player: 10 free questions, no sign-up.

Continue with the 10 free questions →

Those 7 questions were the start.

The exam does not test whether you recognise a term — it tests whether you can rule out three plausible answers under time pressure. That is what the explanations above are for, and there are 780 more questions built exactly like them.

Collecting questions yourself

Scattered across forums, of unknown age
Answer keys without reasoning
No idea which domain you are weak in

Practising with a system

780 questions in 12 full tests, SAA-C03 (2026)
Every option explained — including the wrong ones
Readiness per exam domain, and drills for your weakest

Start free — 10 questions

Straight into the player. No account, no email.

Frequently asked questions

Cluster, spread or partition — how do I tell them apart on the exam?

By the requirement word. Cluster packs instances close in one Availability Zone for lowest latency and highest throughput, at the cost of a large blast radius. Spread puts each instance on distinct hardware (max seven per Availability Zone) for a few individually critical instances. Partition isolates instances into rack-level partitions and scales to many nodes, for large distributed systems like Cassandra, HDFS and Kafka.

Does a cluster placement group reduce availability?

Yes. A cluster group concentrates instances in a single Availability Zone on nearby hardware, so a rack or AZ event can take the whole group down together — a correlated failure the exam expects you to flag. It is a performance choice, not a resilience one; for availability, spread the workload across Availability Zones.

Are these real exam questions?

No. They are original scenarios written for this page in the exam's style and difficulty. Reproducing real exam items violates the AWS certification agreement and can cost a candidate their certification — ours come with the reasoning attached instead.

Updated for SAA-C03 (July 2026). The sample questions above are our own work in the style of the exam — not real exam items. The exam itself is set and marked by the certification body.