Models is the unified asset entry point for machine learning models in the DataBuddy data science module. Built on MLflow and Catalog, it manages the full lifecycle of models from registration, version control, and governance to deployment, taking in training results from model experiments and feeding model serving to provide inference capabilities.
Models are a type of asset in the Catalog that share the same permission, Tag, and lineage systems as tables and volumes, and are uniquely identified by the three-part name catalog.schema.model_name.
Data Organization Model
Models share the same four-level data organization model as tables and volumes:
Layer | Description |
Catalog | A data catalog that is visible across workspaces, serves as the outermost container for model registration, and has a specific type (Table / Model / Volume, and so on). |
Schema | A sub-namespace under Catalog, used to organize a group of related models. |
Model | A registered model asset whose full name is catalog.schema.model_name and that contains one or more versions. |
Model Version | A specific version of a model that corresponds to a training output and contains complete metadata such as artifacts, metrics, parameters, and lineage. |
Note:
Before registering a model, create a catalog of type Model in Catalog and create a Schema under it. For details, see Model Catalog. Key Concepts
To understand model management, you need to be familiar with the following core concepts:
Registered Model: a model asset governed by Catalog, with a three-part full name and an incrementing version sequence.
Model Version: the smallest deployable unit of a model, corresponding to a single training artifact. The version number increments automatically with each registration (1, 2, 3, ...) and cannot be specified manually.
Model Signature: the definition of model input and output fields. A signature is mandatory for registering a managed model, and registration fails if it is missing.
Tags: key-value pairs applied to models or versions, used to mark business attributes (such as team, business line, and framework) and lifecycle stages, and support filtering in lists.
Lineage: the relationships between model versions and upstream and downstream assets (such as feature tables, Notebooks, inference tables, and model serving), recorded centrally by Catalog.
References