View a markdown version of this page

Use fault isolation to protect your workload - AWS Well-Architected Framework

Use fault isolation to protect your workload

Everything fails eventually, from a single disk to the power and network feeding an entire site. A failure that reaches every part of a workload takes the whole workload down with it. Fault isolation limits the impact of a component or system failure to a defined boundary, so components outside that boundary keep working. Running a workload across several isolation boundaries turns a single point of failure into a contained event that most requests never notice.

This capability covers how an organization uses fault isolation to keep a failure from spreading: deploying compute and data across multiple isolated locations, automating recovery for the components that cannot leave a single location, and partitioning a workload into independent cells so one fault reaches only a fraction of the requests.

Capability intent

  • Production workloads run across at least two isolated locations, so a failure in one location does not end availability everywhere.

  • The location strategy matches the workload's resilience requirements, rather than defaulting to the most complex option.

  • Components that must stay in a single location have an automated, tested rebuild path that meets their recovery time and recovery point objectives.

  • Workloads that need containment beyond location boundaries are partitioned into independent cells that share no state.

  • Cell size is bounded and treated as a reliability control, and the workload grows by adding cells rather than by enlarging existing ones.

  • Traffic redirects to healthy locations and cells automatically when one becomes impaired.

Common issues to watch for

  • A production workload runs in a single Availability Zone, so the loss of that zone ends availability.

  • Teams reach for a multi-Region architecture when a multi-AZ design meets the requirements, and take on cost and complexity for no gain.

  • A component in one Region calls a component in another Region during normal operation, which couples the two and lets a failure cascade between them.

  • Deployments and data drift out of sync between locations, so the surviving location cannot actually take over.

  • A single-location component has no automated rebuild, and its recovery time stays unknown until an incident exposes it.

  • Cells grow without an upper bound or share state, so the fault domain widens and the isolation the architecture promised erodes.