Daglite

A lightweight, type-safe Python framework for building and executing DAGs (Directed Acyclic Graphs) with explicit data flow and composable operations.

📚 Documentation | 🚀 Getting Started | 💡 Examples

[!WARNING] This project is in early development. The API may change in future releases. Feedback and contributions are welcome!

Quick Start

Installation

uv pip install daglite

# With CLI support
uv pip install daglite[cli]

Your First DAG

from daglite import task, evaluate

@task
def fetch_data(url: str) -> dict:
    """Fetch data from an API."""
    return {"url": url, "data": [...]}

@task
def process(data: dict) -> list:
    """Process the fetched data."""
    return [item.upper() for item in data["data"]]

@task
def save(items: list, path: str) -> None:
    """Save results to a file."""
    with open(path, "w") as f:
        f.write("\n".join(items))

# Build and execute the DAG
result = evaluate(
    fetch_data(url="https://api.example.com")
    .then(process)
    .then(save, path="output.txt")
)

Why Daglite?

Daglite is built for computational work in restricted environments.

Originally designed for operations research analysts working on air-gapped, Windows-only systems, Daglite solves a specific problem: building workflows that are easy to analyze, share with colleagues, and re-run—even after returning to a project months later.

The Core Philosophy

No infrastructure required. Daglite runs anywhere Python runs—no databases, no containers, no cloud services, no servers. Install it with uv pip, define your tasks, and execute them. When you need more (like distributed execution or advanced serialization), plugins extend functionality without adding mandatory dependencies.

Explicit over implicit. Every data dependency is visible in your code. The DAG structure is static and analyzable before execution. Type checkers catch errors before runtime. This makes workflows self-documenting and maintainable.

Type-safe and modular. Full support for mypy, pyright, and other type checkers means your IDE provides autocomplete and catches type mismatches. Compose simple functions into complex pipelines using familiar Python patterns.

When to Use Daglite

Perfect for:

• ETL scripts and data transformations
• Machine learning pipelines (feature engineering, training, evaluation)
• Computational science workflows
• Analysts and data scientists who need reproducible workflows
• Air-gapped or restricted environments
• CLI tools with workflow orchestration
• Local development and prototyping
• Projects where simplicity and type safety matter

Not ideal for:

• Production job scheduling with cron-like triggers → Use Airflow, Prefect
• Real-time streaming data → Use Kafka, Flink
• Distributed computing at massive scale → Use Spark, Dask
• Multi-tenant orchestration platforms → Use Dagster

Daglite complements these excellent tools. Think of it like Flask vs Django—we give you ownership of the toolchain for local, explicit workflows, while respecting the power and sophistication of infrastructure-heavy frameworks for production orchestration.

Key Features

• Type-Safe Task Composition - Complete type checking support with mypy, pyright, pyrefly, and ty. Your IDE catches errors before runtime.

• Fluent API - Chain operations naturally with .then(), .map(), .join(). Build complex pipelines with readable code.

• Lightweight Core - No mandatory infrastructure—runs anywhere Python runs. Optional plugins add capabilities when you need them.

• Async Execution - Built-in support for threading and multiprocessing backends. Run tasks in parallel without changing your code structure.

• Composable Patterns - Mix and match patterns: Inline pipelines, fan-out/fan-in, map-reduce, parameter sweeps, pairwise operations.

• Testable - Pure functions make DAGs easy to test and debug. No mocking infrastructure or database connections.

• CLI Support - Define pipelines once, run them from the command line with argument parsing included.

Core Concepts

Tasks

Functions decorated with @task become composable DAG nodes:

@task
def process_data(input: str, param: int = 10) -> dict:
    """Tasks are just functions with explicit inputs/outputs."""
    return {"result": input * param}

Lazy Evaluation

Tasks return futures—they don't execute until you call evaluate():

# Create a future (lazy evaluation)
future = process_data(input="hello", param=5)

# Execute when ready
result = evaluate(future)

Composition Patterns

Common Patterns

Inline Pipeline

@task
def load_config(path: str) -> dict:
    return json.load(open(path))

@task
def init_model(config: dict) -> Model:
    return Model(**config)

@task
def train(model: Model, data: pd.DataFrame) -> Model:
    model.fit(data)
    return model

# Chain operations
result = evaluate(
    load_config(path="config.json")
    .then(init_model)
    .then(train, data=training_data)
)

Parallel Fan-Out

@task
def fetch_user(user_id: int) -> dict:
    return api.get(f"/users/{user_id}")

@task
def save_all(users: list[dict]) -> None:
    db.bulk_insert(users)

# Process multiple users in parallel
result = evaluate(
    fetch_user.product(user_id=[1, 2, 3, 4, 5])
    .join(save_all)
)

Map-Reduce

@task
def square(x: int) -> int:
    return x ** 2

@task
def double(x: int) -> int:
    return x * 2

@task
def sum_all(values: list[int]) -> int:
    return sum(values)

# Fan-out, transform, reduce
result = evaluate(
    square.product(x=[1, 2, 3, 4])
    .map(double)
    .join(sum_all)
)
# Result: 60 = (2 + 8 + 18 + 32)

Async Execution

# Run DAG with threading backend
result = evaluate(my_dag, use_async=True)

# Per-task backends
@task(backend_name="threading")
def io_bound_task(url: str) -> bytes:
    return requests.get(url).content

@task(backend_name="multiprocessing")
def cpu_bound_task(data: np.ndarray) -> np.ndarray:
    return expensive_computation(data)

CLI Pipelines

from daglite import pipeline

@pipeline
def ml_pipeline(model_path: str, data_path: str, epochs: int = 10):
    """Train a machine learning model."""
    data = load_data(path=data_path)
    model = train_model(data=data, epochs=epochs)
    return save_model(model=model, path=model_path)

Run from command line:

daglite run ml_pipeline --model-path model.pkl --data-path data.csv --epochs 20

Documentation

Full documentation is available at cswartzvi.github.io/daglite

• Getting Started Guide
• User Guide
• Plugins
• API Reference
• Examples

Community

🤝 Contributing

Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

💬 Discussions

Join the conversation on GitHub Discussions.

🐛 Issues

Found a bug or have a feature request? Open an issue.

License

MIT License - see LICENSE for details.

Acknowledgments

Inspired by the design patterns and philosophies of:

• Apache Airflow - DAG orchestration at scale
• Prefect - Modern workflow design
• Dagster - Data pipeline architecture
• Dask - Lazy evaluation and graph execution
• itertools - Composable Python operations

Each of these projects excels in their domain. Daglite aims to complement them by providing a lightweight alternative for local, type-safe workflows.