ftlite 0.2.1

A lightweight, zero-infrastructure feature store for machine learning.

FTLite

The simplest local-first feature store for Python.

Define features once. Train models without data leakage. Serve features in milliseconds.

No Kubernetes. No Redis. No Cloud. Just Python.

---

Why do I need a Feature Store?

Without FTLite	With FTLite
❌ Feature engineering logic is duplicated across train & serve	✅ Define features once and reuse them everywhere
❌ Training and inference compute different feature values	✅ Ensure consistency between offline & online models
❌ Data leakage silently hurts real-world model performance	✅ Prevent data leakage with temporal point-in-time joins
❌ Feature versions are difficult to manage and swap	✅ Track lineage & version features out-of-the-box

---

FTLite vs. Feast

Capability	FTLite	Feast
Local-first / Zero-infrastructure	✅	⚠️ Complex
Kubernetes Required	❌	Often
Redis Required	❌	Usually
DuckDB (Offline Joins)	✅	❌
SQLite (Online Serving)	✅	❌
Point-in-Time Joins	✅	✅
Feature Versioning	✅	✅
Zero Setup	✅	❌

---

Architecture

---

Installation

pip install ftlite

To enable native Polars DataFrame outputs:

pip install ftlite[polars]

---

2-Minute Quick Example

1. Define and Register Features

from ftlite import Entity, Feature, FeatureView, FtliteClient

client = FtliteClient()

# Identify entity
customer = Entity(name="customer_id", value_type="INT64")
client.register_entity(customer)

# Define feature view mapping to local Parquet file
fv = FeatureView(
    name="customer_stats",
    entities=[customer],
    features=[
        Feature(name="balance", dtype="double"),
        Feature(name="active_days", dtype="int64")
    ],
    source_path="customer_data.parquet",
    timestamp_field="timestamp"
)
client.register_feature_view(fv)

2. Historical Join (Prevent Data Leakage)

# Pass raw entity observations and timestamps
obs_df = pd.DataFrame({
    "customer_id": [1001],
    "timestamp": ["2026-07-14T12:00:00"]
})

hist_df = client.get_historical_features(
    entity_df=obs_df,
    features=["customer_stats:balance", "customer_stats:active_days"]
)

Visualizing Point-in-Time Join:

Input (Entity Observations):
| customer_id | timestamp           |
| ----------- | ------------------- |
| 1001        | 2026-07-14T12:00:00 |

Output (PIT Correct Joined Features):
| customer_id | timestamp           | customer_stats:balance | customer_stats:active_days |
| ----------- | ------------------- | ---------------------- | -------------------------- |
| 1001        | 2026-07-14T12:00:00 | 5250.75                | 14                         |

3. Materialize & Serve Low-Latency Online Features

import datetime

# Sync last 30 days offline features to SQLite serving store
client.materialize(
    start_time=datetime.datetime(2026, 6, 15),
    end_time=datetime.datetime(2026, 7, 15)
)

# Fetch low-latency prediction features in <1ms
online_features = client.get_online_features(
    entity_keys=[1001],
    features=["customer_stats:balance"]
)
# Returns: [{'entity_id': 1001, 'customer_stats:balance': 5250.75}]

---

Feature List

• ✅ Point-in-Time Correct Joins: Built on DuckDB temporal ASOF joins to prevent data leakage.
• ✅ Low-Latency Online Serving: SQLite-backed serving engine fetching features in sub-milliseconds.
• ✅ One-Line Feature Versioning: Track feature versions explicitly with automatic fallbacks.
• ✅ Zero-Configuration Caching: Automatically cache historical query computations.
• ✅ Feature Lineage Tracing: View recursive upstream dependencies easily via CLI or Python API.
• ✅ Optional Polars Support: Perform lightning-fast queries with native Polars outputs.
• ✅ Python-First API: Simple object-oriented structures with 100% type annotations.

---

Complete ML Examples

We provide fully self-contained ML examples inside the examples/ directory:

• 📊 Customer Churn: End-to-end model training and online inference.
• 🛡️ Fraud Detection: Dynamic query-time On-Demand transformations.
• 🏠 House Price Prediction: Zip-code statistics using temporal joins.
• 🔍 Recommendation System: Version fallback and lookup syntaxes.
• 📈 Time Series Forecasting: Target lag feature generations.
• ⚡ Performance Benchmarks: Cache hit vs. miss speedups, and Pandas vs. Polars extraction timings.

---

Detailed Documentation

For comprehensive guides on On-Demand Transformations, versioning patterns, cache management, lineage tracking, and command-line scripts, refer to the Detailed Guide & API Reference.

---

License

MIT License.