Databricks Generative AI Engineer Associate: Assembling and deploying applications Practice Questions

HTTP 504 Gateway Timeout errors on complex multi-hop agent tasks are most commonly caused by (1) insufficient compute resources causing slow per-request execution, and (2) cold-start latency when scale-to-zero is enabled and the first request after a period of inactivity must wait for instance initialization. Increasing workload_size provides more CPU/memory for faster processing, and disabling scale-to-zero eliminates cold starts. Enabling inference tables captures telemetry but has no effect on the timeout threshold. max_provisioned_throughput is a parameter for Foundation Model API endpoints (token throughput), not custom model serving. Switching to GPU for LangGraph orchestration logic (Python control flow) would not meaningfully reduce latency since LangGraph coordination is CPU-bound, not GPU-acceleratable.

The correct MLflow client method for setting a named alias on a Unity Catalog registered model version is client.set_registered_model_alias(name, alias, version). This is the Unity Catalog-compatible API for alias management. client.transition_model_version_stage() is the legacy Workspace Registry API for lifecycle stages (Staging, Production, Archived) and is not supported in Unity Catalog, which uses aliases instead of stages. client.update_model_version() updates metadata like the description or tags, and does not accept an alias parameter. mlflow.register_model_alias() is not a valid top-level MLflow function — alias management is done through the MlflowClient, not a module-level function.

mlflow.langchain.log_model() is the correct flavor for logging LangChain-based RAG chains. It natively understands the LangChain interface, sets the correct model signature, and produces an artifact that Databricks Model Serving can directly serve. mlflow.pyfunc.log_model() is a generic flavor that requires manual wrapper implementation and does not natively handle LangChain specifics. mlflow.sklearn.log_model() is for scikit-learn estimators and is incompatible with LangChain chains. mlflow.transformers.log_model() is for Hugging Face Transformers pipelines, not LangChain chains.

MlflowClient().set_registered_model_alias(name, alias, version) is the correct MLflow Python API for assigning an alias to a specific registered model version in Unity Catalog. mlflow.set_registered_model_alias() is not a top-level MLflow function — it must be called on an MlflowClient instance. transition_model_version_stage() is from the legacy Workspace Model Registry API and deals with stages (Staging/Production), not aliases, and is not supported in Unity Catalog. mlflow.register_model() registers a new model version from a run; it does not assign aliases, and tags are metadata, not aliases.

In the Databricks Model Serving REST API, the served_entities block uses entity_name (the fully qualified Unity Catalog model name) and entity_version to specify which registered model version to serve. model_name and model_version are field names from older or incorrect API schemas not used in the served_entities structure. registered_model_name and model_version combine a valid concept with a wrong field name for the entity block. While workload_size is often specified, it is not strictly required alongside entity_name and entity_version — those two fields are the key identifiers for the model to serve.