mlprep-rust 0.3.1

High-performance no-code data preprocessing engine

mlprep 🚀

The fastest no-code data preprocessing engine for Machine Learning.
Powered by Rust & Polars.

🎬 Introduction Video

📺 Watch: mlprep Introduction (NotebookLM)

---

Stop writing slow, fragile pandas boilerplate.
Start defining robust, reproducible pipelines.

mlprep is a high-performance CLI tool and Python library that handles the dirty work of ML engineers: type inference, missing value imputation, complex joins, and feature engineering—all defined in a simple YAML config.

🔥 Why mlprep?

🚀 Blazing Speed

Built on Rust and Polars, mlprep processes gigabytes of data in seconds, not minutes. It leverages multi-threading and SIMD vectorization out of the box.

✨ Zero-Code Pipelines

Define your entire preprocessing workflow in pipeline.yaml. No more "spaghetti code" notebooks that no one can read.

🛡️ Quarantine Mode

Don't let dirty data crash your training. mlprep isolates invalid rows (schema mismatch, outliers) into a separate "quarantine" file, so your pipeline stays green and your models stay clean.

🔄 Build Once, Run Anywhere

fit your feature engineering steps (scaling, encoding) on training data and transform production data with exact reproducibility. No more training-serving skew.

---

⚡️ Quick Start

1. Install

pip install mlprep-rust

2. Define your pipeline (pipeline.yaml)

inputs:
  - path: "data/raw_users.csv"
    format: csv

steps:
  # ETL
  - fillna:
      strategy: mean
      columns: [age, income]
  - filter: "age >= 18"
  
  # Data Quality Check
  - validate:
      mode: quarantine # Bad rows go to 'quarantine.parquet'
      checks:
        - name: email
          regex: "^.+@.+\\..+$"

  # Feature Engineering
  - features:
      config: features.yaml

outputs:
  - path: "data/processed_users.parquet"
    format: parquet
    compression: zstd

3. Run it

mlprep run pipeline.yaml

Result: A clean, highly-compressed Parquet file ready for training. 🚀

---

🆚 Comparison

Feature	Pandas	mlprep
Speed	🐢 Single-threaded	🐆 Multi-threaded (Rust)
Pipeline	Python Script	YAML Config
Validation	Manual .loc[] checks	Built-in Quality Engine
Bad Data	Crash or Silent Fail	Quarantine Execution
Memory	Bloated Objects	Zero-Copy Arrow

---

⚡️ Performance

mlprep is designed for speed, leveraging Rust's ownership model and Polars' query engine.

Operation	vs Pandas	Note
CSV Read	~3-5x Faster	Multi-threaded parsing
Pipeline	~10x Faster	Lazy evaluation & query optimization
Memory	~1/4 Usage	Zero-copy Arrow memory format

Benchmarks run on 1GB generated dataset. To run your own benchmarks:

python scripts/benchmark.py --size 1.0 --compare-pandas

---

🗺️ Roadmap

We are actively building MVP (Phase 1). Check out our documentation:

• Implementation Plan & Roadmap
• Technical Specification

---

📚 Use Cases & Examples

Explore full examples in the examples/ directory:

1. Basic ETL Pipeline

• Scenario: Filter, select columns, and convert CSV to Parquet.
• Key Features: filter, select, write_parquet.

2. Data Validation

• Scenario: Ensure data quality before training.
• Key Features: Schema validation, quarantine mode for invalid rows.

3. Feature Engineering

• Scenario: Generate features for ML training.
• Key Features: fit (train) / transform (prod) pattern, standard_scaler, one_hot_encoding.

4. Scikit-Learn Integration

• Scenario: Use mlprep as a preprocessing step in a Scikit-Learn pipeline.
• Key Features: Seamless integration with Python ML ecosystem.

5. MLflow Experiment Tracking

• Scenario: Track preprocessing parameters and artifacts in MLflow.
• Key Features: Reproducibility and experiment management.

6. Airflow DAG

• Scenario: Schedule and monitor mlprep run as part of an Airflow DAG.
• Key Features: Production-friendly orchestration with BashOperator.

7. DVC Pipeline

• Scenario: Version control processed datasets with a DVC stage that calls mlprep.
• Key Features: Reproducible data artifacts (dvc repro + mlprep run pipeline.yaml).

---

🤝 Contributing

We welcome contributions! Please see the issue tracker for good first issues.

📄 License

MIT