mloda-plugins: Building mloda Plugins

Check the registry index first

Always do this before writing a plugin: it may already exist. mloda-registry publishes community plugins covering common {col}__{op} transforms (aggregation, window/scalar/frame aggregate, scalar/point arithmetic, rank, offset, percentile, binning, datetime, string ops, time bucketization, ffill, ema, sessionization, resample); see its plugins table for the full list. None of it ships with plain pip install mloda; install what you need (pip install mloda-community for all, or one package like pip install "mloda-community-rank[pandas]"), then confirm what's actually loaded:

from mloda.user import PluginLoader
from mloda.steward import get_feature_group_docs

PluginLoader.all()
for fg in get_feature_group_docs():
    print(fg.name, fg.description)

An empty or unrelated result means it isn't installed yet, not that it doesn't exist; see 02-discover-plugins.md for more on installed-vs-available. (The mloda skill covers the same check for requesting data instead of writing a plugin.)

The three plugin types

Type	Role	Full guide
FeatureGroup	A data transformation, the unit you'll write most often	09-create-feature-group.md
ComputeFramework	The execution backend a transformation runs on (Pandas, PyArrow, Polars, DuckDB, ...)	10-create-compute-framework.md
Extender	Hooks for logging, tracing, validation around feature calculation	11-create-extender.md

Reach for a ComputeFramework only when adding a new execution backend, or an Extender for cross-cutting hooks.

Minimal example

A primary-source FeatureGroup (no inputs):

from mloda.provider import FeatureGroup

class CustomerScoring(FeatureGroup):
    @classmethod
    def calculate_feature(cls, data, features):
        return {"customer_score": 100}

A chained one, reusable via an input__operation feature name (see "FeatureGroup: key decisions" below):

from mloda.provider import FeatureChainParserMixin, FeatureGroup

class MyOp(FeatureChainParserMixin, FeatureGroup):
    PREFIX_PATTERN = r".*__([\w]+)_my_op$"
    PROPERTY_MAPPING = {...}

    @classmethod
    def calculate_feature(cls, data, features): ...

Register either via PluginLoader.all() (installed package) or a direct import; see 03-create-plugin-in-project.md.

The plugin journey

Progression from using to sharing, each guide hosted in mloda-registry:

1. Use an existing plugin / Discover plugins - start here, before writing anything.
2. Create a plugin in your project - add a FeatureGroup inline, no separate package.
3. Create a plugin package - scaffold a standalone installable package from mloda-plugin-template.
4. Share with your team (private git) / Publish to the community registry / Contribute to official plugins / Become an official plugin.

Full progression: docs/guides/index.md.

FeatureGroup: key decisions

Full decision tree: guide 09. The recurring ones:

• Loads external data (file, DB, API)? Root feature pattern, see 01-root-features.md.
• Transforms 1+ existing features? Derived feature, see 02-derived-features.md.
• Should it be reusable via an input__operation naming pattern? FeatureChainParserMixin, see 03-chained-features.md (2+ inputs: 04-multi-input-features.md).
• Multiple output columns? 05-multi-output-features.md (feature~0, feature~1).
• Fitted/trained state between runs? 06-artifact-features.md.
• Time ordering, group-by, or joins across feature groups? 07-index-features.md / 08-links-joins.md.
• Group-identity option vs. runtime-only metadata? 11-options.md - a required context option that isn't forwarded through input_features() breaks chained requests; see 26-input-feature-forwarding.md for the forwarding mechanism itself.
• Standard column transform (binning, window/scalar/frame aggregate, scalar/point arithmetic, rank, offset, percentile, string op, time bucketization, ffill, ema, resample, sessionization)? See the data operation patterns index before writing one from scratch.
• Feature defined via JSON/config instead of a naming pattern (e.g. to back the mloda skill's LLM Tool Function pattern)? 22-feature-config.md.
• Ready to test? 10-testing-guide.md.

Full pattern index (27 guides, covering filters, validators, versioning, streaming, realtime execution, and more): feature-group-patterns/index.md.

ComputeFramework: key decisions

Full decision tree: guide 10.

• Needs a connection/session object? Data lake table format (Iceberg, Delta, Hudi) -> 05-data-lake.md; otherwise -> 03-stateful-connection.md (DuckDB, Spark).
• Lazy evaluation? 02-stateless-lazy.md (Polars Lazy, Ibis).
• Eager, no external deps? 04-zero-dependency.md; with deps -> 01-stateless-eager.md (Pandas, PyArrow).
• Cross-framework conversion? 08-framework-transformer.md.
• Joins/merges or filters? 06-merge-engine.md / 07-filter-engine.md.
• Ready to test? 09-testing-guide.md.

Full pattern index: compute-framework-patterns/index.md.

Extender: key decisions

Full guide: guide 11. Wraps feature calculation or input/output validation; set a custom priority (lower runs first, default 100) if execution order matters; use class-level storage for state under ParallelizationMode.MULTIPROCESSING (must be pickle-safe).

Reference

• Registry (community/enterprise plugins, guides): https://github.com/mloda-ai/mloda-registry
• Plugin template (scaffold for a standalone package): https://github.com/mloda-ai/mloda-plugin-template
• Core docs: https://mloda-ai.github.io/mloda/