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.
downloads
projects
features servedWhy BoxLink for Feature Engineering
Traditional feature engineering fails during model deployment. BoxLink bridges the gap.
Performance Benchmark
How BoxLink stack measures up against cloud-heavy traditional feature store infrastructure.
BoxLink (Local-first)
Local SQLite/DuckDB retrieval times
Traditional Store (Feast)
Cloud infrastructure retrieval latency
Architecture
DuckDB powers point-in-time offline historical joins, and SQLite delivers sub-millisecond serving online.
BoxLink vs. Feast
A comparison of features, infrastructure requirements, and local development velocity.
| Capability | BoxLink | Feast |
|---|---|---|
| Local-first / Zero-infrastructure | Yes | Complex |
| Kubernetes Required | No | Often |
| Redis Required | No | Usually |
| DuckDB (Offline Joins) | Yes | No |
| SQLite (Online Serving) | Yes | No |
| Point-in-Time Joins | Yes | Yes |
| Feature Versioning | Yes | Yes |
| Zero Setup | Yes | No |
2-Minute Quick Example
Register features, perform historical temporal joins, and materialize to SQLite for sub-millisecond serving.
from boxlink import Entity, Feature, FeatureView, BoxLinkClient
client = BoxLinkClient()
# 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)Complete ML Examples
Browse full end-to-end setups implementing BoxLink in standard machine learning pipelines.
Customer Churn
End-to-end model training, temporal offline evaluation, and low-latency online inference serving.
Fraud Detection
Dynamic query-time On-Demand transformations, sub-millisecond lookups, and feature pipelines.
House Price Prediction
Aggregating zip-code statistics and spatial-temporal features using correct point-in-time joins.
Recommendation System
Version fallback handling, dynamic entity lookups, and flexible recommendation-oriented APIs.
Time Series Forecasting
Generating target lag features, shifting window features, and correct temporal data alignments.
Performance Benchmarks
Testing cache hit vs. miss speedups, and analyzing Pandas vs. Polars extraction timings.
Latest From The Blog
Deep dives into local-first feature engineering, real-time machine learning, and data pipelines.
Building Local-First ML Pipelines with DuckDB & SQLite
How to utilize embedded databases to solve offline feature generation and online serving without standing up heavy network infrastructure.
Understanding Point-in-Time Joins and Data Leakage
A detailed breakdown of temporal joins, why traditional joins cause target leakage, and how BoxLink automates AS-OF joins.
From Laptop to Production: Deploying BoxLink in Serverless
Learn how BoxLink's zero-dependency design makes it the perfect fit for AWS Lambda, Google Cloud Run, and edge runtimes.