Skip to content

Pydantic Policies & Validation

To make development reliable and prevent silent runtime crashes, mlflow-experience validates all configuration settings using Pydantic v2 (src/core/context/contexts.py and src/core/context/policies.py).

Why Use Pydantic?

In traditional ML projects with YAML files, a simple typo (such as strategy: meaan or omitting a required column) would only fail deep inside Scikit-Learn code after 10 minutes of execution.

With Pydantic: - Instant Failure: Your configuration is checked before any data is loaded or processed. - Helpful Error Messages: It tells you the exact line, field name, and expected values if something is wrong. - Automatic Type Conversion: String numbers are parsed into integers or floats automatically where appropriate.

How It Works

When main.py runs, it calls:

context = RunContext.from_yaml_dict(config)
  1. Pydantic parses the dictionary against RunContext.
  2. A model validator (cast_all_preprocessing_params) checks each step in preprocessing.
  3. The step's type looks up the corresponding policy class from PreprocessingRegistry.policy_class.
  4. The dictionary is instantiated into a strongly-typed policy object (e.g. MissingValuePolicy, ScalingPolicy, EncodingPolicy).

Policy Reference

RenamePolicy

Validates column rename mappings: - mapping (dict[str, str]): Must not be empty.

MissingValuePolicy

Validates missing value strategies: - strategy: Must be one of "mean", "median", "mode", or "constant". - fill_value: Required if strategy: constant.

EncodingPolicy

Validates encoding configurations: - strategy: Must be one of "one_hot", "ordinal", "label", or "target". - order: A list of strings specifying the exact order for ordinal encoding (e.g. ['Low', 'Medium', 'High']).

ScalingPolicy

Validates scaling strategies: - strategy: Must be one of "standard", "minmax", or "robust".

Example Error Output

If you provide an invalid strategy in your YAML:

- name: bad_step
  type: missing
  params:
    columns:
      - name: age
        strategy: invalid_strategy

Pydantic immediately catches it:

pydantic_core._pydantic_core.ValidationError: 1 validation error for MissingValueConfig
strategy
  Input should be 'mean', 'median', 'mode' or 'constant' [type=literal_error, input_value='invalid_strategy', input_type=str]