Deployment

Deployment makes a selected release available in an execution environment. For a model service, the release includes the model, runtime, configuration, and required dependencies. Deployment does not train the model, and a deployed model need not immediately receive all production traffic.

From an accepted model to a running service

Suppose a recommendation model passes offline evaluation. The team prepares its runtime, verifies compatible feature definitions, starts workers, checks readiness, and decides how requests will reach them. SageMaker's documentation illustrates several inference deployment modes, including hosted services and batch execution.

Sending a small share of traffic to the new version can expose problems before a wider rollout. The comparison still needs representative requests and explicit quality and latency criteria. A healthy process is not enough if it returns the wrong product ranking.

Plan what can be reversed

Keep the known working release and the configuration needed to restore it. If the new version changed feature schemas or wrote incompatible data, reverting only the model file may not restore correct behavior.

Serving describes how the running system provides outputs; deployment describes the change that made a release available. Record the deployed version, its dependencies, and rollout decisions. Continue observing the system after deployment because real traffic and data can differ from the test conditions.

Reference: Amazon SageMaker: Deployment for inference.