Core concept
Cluster: same rack, high bandwidth low-latency. HPC + tightly-coupled apps. Single AZ.
Advertisement
How it works
Partition: distinct partitions of racks. HDFS, Cassandra. Partition failure isolated. Up to 7 partitions per AZ.
Advertisement
Trade-offs + gotchas
Spread: each instance on separate hardware. Max availability. Small groups (7 per AZ). Max avoid correlated failures.