aprender-registry 0.29.0

Table of Contents

• Overview
• Architecture
• Installation
• Usage
• CLI
• Features
• Benchmarks
• Testing
• Security
• Contributing
• License

Overview

Pacha is a unified registry for machine learning artifacts -- models, datasets, and training recipes -- with full lineage tracking, semantic versioning, and cryptographic integrity verification. It provides content-addressed storage with BLAKE3 hashing for deduplication, tamper detection, and efficient delta storage.

Key Capabilities

• Model Registry - Register, version, and stage ML models with metadata and metrics
• Data Registry - Track datasets with schema validation and provenance
• Recipe Registry - Store training configurations with hyperparameters and environment specs
• Lineage Tracking - Full dependency graph from data to deployed model
• Content-Addressed Storage - BLAKE3-based deduplication and integrity verification
• Cryptographic Signing - Ed25519 signatures for artifact authenticity
• Experiment Tracking - Record training runs with metrics, parameters, and artifacts

Architecture

+------------------+     +------------------+     +------------------+
|  Model Registry  |     |  Data Registry   |     | Recipe Registry  |
|  (.apr files)    |     |  (.ald files)    |     | (TOML configs)   |
+--------+---------+     +--------+---------+     +--------+---------+
         |                        |                        |
         +------------------------+------------------------+
                                  |
                      +-----------+-----------+
                      |   Content-Addressed   |
                      |   Storage (BLAKE3)    |
                      +-----------+-----------+
                                  |
                      +-----------+-----------+
                      |  SQLite Metadata DB   |
                      |  (~/.pacha/registry)  |
                      +-----------------------+

Installation

Add to your Cargo.toml:

[dependencies]
pacha = "0.2.5"

Or install the CLI:

cargo install pacha

Usage

use pacha::prelude::*;

fn main() -> Result<()> {
    let registry = Registry::open(RegistryConfig::default())?;

    // Register a model with documentation
    let card = ModelCard::builder()
        .description("Fraud detection model")
        .metrics([("auc", 0.95), ("f1", 0.88)])
        .build();

    registry.register_model(
        "fraud-detector",
        &ModelVersion::new(1, 0, 0),
        &model_bytes,
        card,
    )?;

    // Retrieve and inspect
    let model = registry.get_model(
        "fraud-detector",
        &ModelVersion::new(1, 0, 0),
    )?;
    println!("Stage: {}", model.stage);

    Ok(())
}

Data Registry

use pacha::data::*;

// Register a dataset with schema
let schema = DataSchema::new(vec![
    Column::new("feature_1", DataType::Float64),
    Column::new("label", DataType::Int32),
]);

registry.register_data(
    "training-set-v2",
    &DataVersion::new(2, 0, 0),
    &data_bytes,
    schema,
)?;

Experiment Tracking

use pacha::experiment::*;

let experiment = Experiment::builder()
    .name("fraud-detection-v3")
    .model("fraud-detector", &ModelVersion::new(1, 0, 0))
    .dataset("training-set-v2")
    .hyperparams([("lr", "0.001"), ("epochs", "50")])
    .build();

registry.log_experiment(experiment)?;

CLI

# Initialize a registry
pacha init

# Model operations
pacha model register fraud-detector model.apr -v 1.0.0
pacha model list
pacha model stage fraud-detector -v 1.0.0 -t production
pacha model inspect fraud-detector -v 1.0.0

# Data operations
pacha data register training-set data.ald -v 1.0.0
pacha data list

# Registry statistics
pacha stats

Features

Enable all features:

[dependencies]
pacha = { version = "0.2.5", features = ["full"] }

Benchmarks

Run benchmarks with:

cargo bench

Content-addressing operations (BLAKE3 hashing, storage, retrieval) are benchmarked using Criterion. See benches/content_address.rs for benchmark definitions.

Testing

468 tests passing with zero warnings.

# Unit tests
cargo test --lib

# All tests (unit + integration)
cargo test

# All features
cargo test --all-features

# With nextest (faster)
cargo nextest run

# Quality gates
make tier1   # Fast feedback: fmt, clippy, check
make tier2   # Pre-commit: tests + clippy
make tier3   # Pre-push: full validation

# Coverage
make coverage

# Mutation testing
cargo mutants --no-times --timeout 300

Recent Fixes (v0.2.5)

• Non-atomic manifest write fixed: uses temp file + rename for crash safety
• find_best_run handles empty input gracefully instead of panicking

Security

• Cryptographic Integrity: All artifacts are content-addressed with BLAKE3
• Ed25519 Signing: Optional artifact signing for authenticity verification
• Encryption at Rest: Optional ChaCha20-Poly1305 encryption
• Dependency Auditing: cargo-deny and cargo-audit in CI pipeline
• No Unsafe Code: #![deny(unsafe_code)] enforced project-wide

To report a security vulnerability, please email security@paiml.com.

Contributing

Contributions welcome! Please follow the PAIML quality standards:

1. Fork the repository
2. Ensure all tests pass: cargo test
3. Run quality checks: cargo clippy -- -D warnings && cargo fmt --check
4. Submit a pull request

MSRV

Minimum Supported Rust Version: 1.75

See Also

• Cookbook — 7 runnable examples

License

MIT - see LICENSE for details.

Part of the Aprender monorepo — 70 workspace crates.