kitefs 0.2.0a2

A Python feature store library for offline/online feature storage, registry, validation, and serving

KiteFS

KiteFS is a Python feature store library for machine learning. It manages the full lifecycle of ML features — defining feature groups as Python code, registering them in a versioned registry, storing historical data in Parquet files, and serving the latest values for real-time predictions.

KiteFS is library-first: no running server, no Docker, no infrastructure to manage. Install it with pip, define your features, and start building.

• SDK: from kitefs import FeatureStore
• CLI: kitefs init, kitefs apply, kitefs list, kitefs describe, and more
• Python 3.12+ required

Installation

KiteFS is currently in alpha. To install the latest pre-release version:

pip install --pre kitefs

Usage

Quick Start (CLI)

# 1. Initialize a new KiteFS project
kitefs init

# 2. Edit feature_store/definitions/example_features.py
#    to define your feature groups as Python code

# 3. Register definitions into the registry
kitefs apply

# 4. List registered feature groups
kitefs list

# 5. Inspect a specific feature group
kitefs describe <feature_group_name>

SDK

from kitefs import FeatureStore

# Initialize — finds kitefs.yaml by walking up from cwd
fs = FeatureStore()

# Register all feature group definitions
result = fs.apply()
print(f"Registered {result.group_count} group(s)")

# List all registered feature groups
groups = fs.list_feature_groups()

# Describe a specific feature group
details = fs.describe_feature_group("listing_features")

# Export as JSON
json_output = fs.list_feature_groups(format="json")

# Write JSON to a file
fs.describe_feature_group("listing_features", target="output.json")

Available Features

• Project scaffolding (kitefs init) — initialize a new KiteFS project with config, directory structure, and example definitions
• Feature group definitions — define feature groups as Python code using frozen dataclasses (FeatureGroup, Feature, EntityKey, EventTimestamp, Expect, etc.)
• Registry (kitefs apply, kitefs list, kitefs describe) — register, list, and inspect feature groups via a deterministic JSON registry suitable for Git versioning
• Local provider — local filesystem storage with Hive-style Parquet partitioning for the offline store
• Configuration — YAML-based project configuration (kitefs.yaml) with environment variable overrides

In Development

The following features are planned but not yet available:

• Data validation — schema and data validation engine with configurable strictness modes (ERROR, FILTER, NONE) to enforce quality constraints on ingested data
• Data ingestion (kitefs ingest) — ingest DataFrames, CSV, or Parquet files into the offline store with automatic partitioning and validation
• Historical retrieval — query the offline store with column selection and time-range filtering for training dataset generation
• Point-in-time joins — temporally correct joins across feature groups to prevent data leakage in training sets
• Materialization (kitefs materialize) — sync the latest feature values from the offline store to a SQLite-backed online store
• Online serving — single-key lookups for real-time feature serving from the online store
• AWS provider — S3 + DynamoDB backend for production deployments, installable via pip install kitefs[aws]
• Registry sync — push and pull the registry between local and remote storage
• Mock data generation (kitefs mock) — generate synthetic test data that respects schema and expectations
• Smart sampling (kitefs sample) — pull a representative data subset from a remote store for local development

Contributing

Prerequisites

• Python 3.12+
• uv — Python package manager
• just — command runner

Setup

git clone https://github.com/fedaipaca/kitefs.git
cd kitefs
uv sync

Run checks

# Lint + format check + type check + tests
just check