In one sentence
A placement group controls the relative physical placement of EC2 instances: place them close for network performance or separate them to contain correlated failures.
Strategies
| Strategy | Goal | Scope and limit | Typical use |
|---|---|---|---|
| Cluster | Low latency and high throughput | One AZ; concentrated failure domain | HPC and tightly coupled workloads |
| Spread | Isolate a small number of critical instances | Up to 7 running instances per AZ per group | Critical nodes |
| Partition | Group many nodes into fault domains | Up to 7 partitions per AZ; many instances per partition | HDFS, Cassandra, Kafka |
- Cluster optimizes performance; it is not a Multi-AZ HA design.
- The “7” limit means instances for Spread and partitions for Partition.
- A partition is a set of isolated racks, not one rack.
- Placement groups are different from capacity reservations, dedicated hosts, and Auto Scaling groups.
Decision cues
- Extremely low latency and node-to-node throughput → Cluster.
- A few critical instances that should not share hardware → Spread.
- A large distributed system with replica-aware placement → Partition.
Remember
Cluster trades concentration for speed, Spread isolates individual instances, and Partition isolates groups at scale.