Skip to main content

DynamoGuard Reliability

DynamoGuard runs on the shared platform and inherits its availability. Everything on this page is what DynamoGuard adds on top: serving guardrail models, and what a moderation request does when part of the path is unavailable.

Read Platform Reliability first. A guardrail policy cannot be more available than the API, datastores and object storage beneath it.

What Is Different About Guardrail Serving

Three properties separate a guardrail policy from an ordinary stateless workload.

PropertyEffect
Each policy is served by its own model processReplica counts, placement and recovery are per policy, not per deployment
A model must load weights before it can answerA replacement pod is not immediately useful, and the weights come from object storage
A custom policy is a trained artifactIt cannot be recreated in a second region by copying configuration

The Question That Matters

For the platform, the reliability question is whether a request gets a response. For DynamoGuard it is whether a request was actually governed. Those come apart: a moderation request can return success with fewer checks applied than were asked for.

That distinction drives how this section is organised.

PageAnswers
Guardrail Serving Across ZonesWhat a policy needs to survive a node or zone loss
Failure BehaviorWhat a request returns when a dependency is unavailable
Configuration RequirementsWhat a deployment and a calling application must set
Verifying A StandbyConfirming a second region can serve, rather than reporting that it can