Choose the serving boundary
A supported prebuilt container can reduce maintenance when its framework, version, artifact format, and inference behavior fit your model. A custom container offers control over dependencies, request preprocessing, postprocessing, and the server implementation, with additional responsibility for compatibility and operations. Agent Platform also documents custom inference routines that help package Python prediction logic. Verify the supported versions and target environment before choosing.
A team has a supported scikit-learn model and standard prediction needs. Compare a prebuilt container with a custom one. Name a requirement, such as specialized preprocessing, that could justify extra customization.
Implement the HTTP and artifact contract
For a custom prediction container, Agent Platform expects an HTTP server that handles configured prediction and health routes on the specified port. The platform exposes runtime configuration through variables such as AIP_HTTP_PORT and route variables; defaults and API behavior are documented in the current container requirements. If model artifacts are supplied when registering the model, AIP_STORAGE_URI identifies the platform-managed copy available to the container. Load only approved artifacts, handle startup failures, and use a runtime identity with the required access.
A container listens on port 9000, while the model configuration points to 8080. Predict what the liveness check observes and identify the configuration or server change needed to align them.
Distinguish liveness from readiness
A TCP liveness check can establish that a process is listening, but it does not prove that weights and preprocessing are loaded or dependencies are usable. A health response should reflect whether the instance can safely serve requests, according to the platform contract and application design. Test slow initialization, failed downloads, malformed inputs, and dependency timeouts. Avoid returning a healthy response prematurely; otherwise traffic can reach an instance before it is ready.
The web server opens its port immediately but needs two minutes to load model artifacts. What can the liveness probe establish, and what should a health check verify before accepting traffic?
Test the complete prediction path
Version the image, model artifact, preprocessing and postprocessing logic, schema, and runtime settings together in the release record. Test normal and boundary inputs, batch sizes, missing features, error responses, concurrency, latency, and permissions. Run an offline-versus-serving parity test using representative raw examples. Follow a staged deployment with monitoring and a tested rollback; the previous model may require its previous image and feature definitions as well as its weights.
Offline preprocessing normalizes a numeric field, but the container omits the transform. Design a parity test that detects the mismatch before a canary receives production traffic.
A custom server is listening on its configured port but has not loaded its model. What does a successful TCP liveness check prove?
Keep exploring
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.