Serving

Serving is the responsibility of making data or model results available through an interface that consumers can use. The consumer may be a person, a dashboard, or another program. Serving can expose prepared data or compute a result when a request arrives.

A saved result still needs a usable path

Imagine a sales table that is correct but inaccessible to the reporting application. The processing job succeeded, but the intended consumer cannot use its output. A serving design specifies the query interface, access rules, accepted version, response format, and behavior when data is unavailable.

Model serving is a narrower case. TensorFlow Serving is an example of infrastructure for making trained models available for inference. Inference is the model calculation; the surrounding service may also validate inputs, retrieve features, route requests, and format responses.

Deployment prepares the service; serving operates it

Deployment places the selected code, model, and configuration into an execution environment. Serving is how consumers obtain results from the running system. A successful deployment does not establish acceptable response time or correct outputs.

For a report, serving may mean querying a prepared table. For a recommendation, it may mean an API response. Serving does not require that every result be newly calculated or that every consumer access a public internet endpoint.

Reference: TensorFlow: Serving models.