Python Config Style

Python Config Style

This is the user's house style for configuration, command-line interfaces, and how classes receive their settings in Python. Apply it whenever you write or refactor code that touches any of those — new config models, a script's CLI, a class constructor, or the wiring that builds an application from parts.

It rests on four pillars:

1. Config is plain, validated data — every component's settings live in a pydantic model (BaseModel), documented with attribute docstrings.
2. Whole-application wiring is declarative — Hydra YAML with _target_ selects and composes components; instantiate turns the tree into live objects.
3. Small scripts get config from the command line — pydantic-settings CliApp turns one config model into a validated CLI. No argparse, no Hydra ceremony.
4. One config in, one config stored — every __init__ (and every top-level worker function) takes a single config parameter and reads fields off it.

Why this shape

The point is uniform construction. A component never invents its own ad-hoc keyword arguments; it declares one config type and reads from it. That makes every object buildable the same way — from Python, from a YAML file, or from the command line — and makes configs serializable and round-trippable (you can dump a live object back to the YAML that would rebuild it). Validation happens once, at the boundary, so the rest of the code trusts its config.

Keep this reasoning in mind rather than applying the rules mechanically: when a situation isn't covered below, choose the option that preserves "settings are validated data that flow in as one object."

Which tool for which job

The Hydra-vs-pydantic-settings line is about scope: Hydra is for assembling an application out of interchangeable parts (and for sweeps/experiments); pydantic-settings is for a single script you run with a few flags. When unsure, ask "is the user composing a system, or running one tool?"

Pydantic config classes

The canonical shape — study it, then read the rules:

from typing import Dict, Optional

from pydantic import BaseModel, ConfigDict, NonNegativeFloat, NonNegativeInt, field_validator


class DataCollectionConfig(BaseModel, frozen=True):
    """Configuration for the data collection."""

    model_config = ConfigDict(use_attribute_docstrings=True, extra="forbid")

    update_rate: NonNegativeFloat = 0
    """The maximum update rate for the managers; 0 means as fast as possible."""

    fsm: DemonstrateFSMConfig
    """The finite state machine config."""

    managers: Dict[str, Optional[ManagerConfigBasis]] = {}
    """Managers that drive the demonstration actions."""

    log_metrics: int = -1
    """Log metrics every N seconds; -1 disables it."""

    @field_validator("managers", mode="after")
    def drop_empty_managers(cls, v: dict):
        return {name: m for name, m in v.items() if m is not None}

The rules, with the reasoning:

• Name it <Thing>Config. Abstract bases end in ConfigBasis; a top-level CLI/entry config may end in Args. The suffix makes "this is data, not behavior" obvious at a glance.

• Subclass BaseModel and default to frozen=True. A config is a fact about how a component was built; freezing it prevents a component from mutating shared settings out from under whoever else holds them, and lets pydantic cache attribute access. Drop frozen only when something genuinely needs to rewrite its own config.

• Document every field with an attribute docstring — the triple-quoted string below the field — and turn on use_attribute_docstrings=True so pydantic adopts that text as the field's description. This is the single source of truth for field docs: it shows up in --help (pydantic-settings) and in any generated schema, with no duplication. Do not restate the same text in Field(description=...).

• Reach for Field() only for behavior, not for docs — i.e. constraints (Field(min_length=1)), a non-trivial default, or aliases. If a field only needs a type, a default, and a docstring, don't wrap it in Field().

• Prefer pydantic's constrained types (NonNegativeFloat, NonNegativeInt, PositiveInt, NonNegativeInt, …) over a bare int/float plus a hand-written validator. The type states the constraint declaratively and shows up in the schema.

• Set model_config deliberately. Common choices:

  • use_attribute_docstrings=True — adopt attribute docstrings as descriptions (use nearly always).
  • extra="forbid" — reject unknown keys, so a typo in YAML or on the CLI fails loudly instead of being silently ignored.
  • arbitrary_types_allowed=True — when a field holds a live instance (a device handle, another component) rather than plain data.

• Validators run mode="after" and return self / the value. Use @field_validator for one field, @model_validator(mode="after") for cross-field invariants; annotate the model validator's return as Self. Validate at the boundary so the rest of the code doesn't re-check.

• Compose configs by nesting. A field whose type is another *Config builds a tree; this is what lets Hydra assemble and instantiate rebuild the whole thing. Mutable defaults (= {}, = []) are safe here — pydantic deep-copies them per instance.

For the fuller catalog — enums with StrEnum + auto(), kebab-case aliases, the NoConfig placeholder, extra="forbid" trade-offs, and more worked examples — read references/pydantic-config.md.

One config in, one config stored

Every class and every top-level worker function takes its settings as a single config argument and reads fields off it. Don't explode a config into a long kwargs list, and don't scatter self.lr, self.batch_size, … across the instance — keep self.config and read self.config.lr.

class DataSampler(DataSamplerBasis):
    def __init__(self, config: DataSamplerConfig):
        self.config = config

def reorganize_data_by_episode(config: ReorganizeDataByEpisodeConfig) -> ExecutionSummary:
    source_root = config.source_root.expanduser().resolve()
    ...

Why: one parameter means one construction protocol. The same object can be built in Python (DataSampler(cfg)), from YAML (Hydra instantiate), or from the CLI, because all three just need to produce one config. Keeping the whole config on self (instead of copying fields out) keeps the object honest about what it was configured with, which is what makes dump()/round-tripping work.

In Hydra-wired projects (those built on ConfigurableBasis / the InitConfigMeta metaclass from the cfgable package), the base class is the bridge: it lets Hydra pass the YAML fields as flat kwargs and assembles them into your config model automatically before __init__ runs, then calls config_post_init(). So:

• declare def __init__(self, config: YourConfig) and set self.config = config (or rely on the base to set it),
• put derived / post-construction state in config_post_init(), not in __init__, so it runs no matter how the object was built.

See references/hydra-app-config.md for how the bridge maps YAML → config, and what to do in plain projects that don't use it.

Small-script CLIs

For a single script with a few options, the CLI is a pydantic config model, parsed by pydantic_settings.CliApp. Attribute docstrings become --help text; the worker takes the validated config:

from pathlib import Path
from typing import List, Optional

from pydantic import BaseModel
from pydantic_settings import CliApp


class ReorganizeConfig(BaseModel):
    model_config = {"use_attribute_docstrings": True}

    source_root: Path = Path("data/aao_data")
    """Source root in the format <task>/<episode>."""

    dry_run: bool = False
    """Log planned operations without writing anything."""


def main(cli_args: Optional[List[str]] = None) -> None:
    config = CliApp.run(ReorganizeConfig, cli_args=cli_args)
    do_the_work(config)


if __name__ == "__main__":
    main()

Full pattern — kebab-case flag aliases, parse_config/main(cli_args=None) scaffolding, custom argument normalization for multi-value flags — in references/pydantic-settings-cli.md.

Before you finish — quick checklist

• Settings live in a *Config BaseModel, frozen=True unless mutation is needed.
• Every field has an attribute docstring; use_attribute_docstrings=True is set; no duplicated description=.
• Constrained types used where they fit; validators are mode="after".
• Each class / worker function takes a single config and stores it as self.config.
• Whole-app wiring went through Hydra (_target_ + instantiate); a lone script got a CliApp CLI — not the other way around.
• Derived state in config_post_init() (in ConfigurableBasis projects), not __init__.