cursus 2.0.0

Automatic SageMaker Pipeline Generation from DAG Specifications

Cursus: Automatic SageMaker Pipeline Generation

Transform pipeline graphs into production-ready SageMaker pipelines automatically.

Cursus is an intelligent pipeline generation system that automatically creates complete SageMaker pipelines from user-provided pipeline graphs. Simply define your ML workflow as a graph structure, and Cursus handles all the complex SageMaker implementation details, dependency resolution, and configuration management automatically.

🚀 Quick Start

Installation

# Core installation
pip install cursus

# With ML frameworks
pip install cursus[pytorch,gbm]

# Full installation with all features
pip install cursus[all]

SageMaker SDK compatibility: The current cursus 1.x line targets SageMaker SDK 2.x. Pin pip install "cursus<2" to stay on this line. The 2.x line (forthcoming) will target SageMaker SDK 3.x; that work happens on main and is published from there once ready.

30-Second Example

from cursus.core import compile_dag_to_pipeline
from cursus.api import PipelineDAG
from sagemaker.workflow.pipeline_context import PipelineSession

# Create a simple DAG
dag = PipelineDAG()
dag.add_node("CradleDataLoading_training")
dag.add_node("TabularPreprocessing_training") 
dag.add_node("XGBoostTraining")
dag.add_edge("CradleDataLoading_training", "TabularPreprocessing_training")
dag.add_edge("TabularPreprocessing_training", "XGBoostTraining")

# Set up SageMaker session
pipeline_session = PipelineSession()
role = "arn:aws:iam::123456789012:role/SageMakerExecutionRole"

# Compile to SageMaker pipeline automatically
pipeline = compile_dag_to_pipeline(
    dag=dag,
    config_path="config.json",
    sagemaker_session=pipeline_session,
    role=role,
    pipeline_name="fraud-detection"
)
pipeline.upsert()  # Deploy and run!

Command Line Interface

The cursus CLI is organized into focused command groups (run cursus --help to see them all):

# Compile a DAG + config into a SageMaker pipeline definition
cursus compile -d dag.json -c config.json -o pipeline.json

# Dry-run validation with a resolution report (no upsert)
cursus compile -d dag.json -c config.json --validate-only --show-report

# Compile and deploy in one step
cursus compile -d dag.json -c config.json --upsert --start

# Discover and inspect local pipeline projects
cursus projects list
cursus projects show my-project

# Browse / recommend pre-built DAGs from the pipeline catalog
cursus pipeline-catalog list
cursus pipeline-catalog recommend --framework xgboost --features calibration

# Run scripts offline against a DAG (data flow resolved automatically)
cursus validate run-scripts dag.json --config-file config.json

# Expose the engine to LLM agents over MCP
cursus mcp list-tools
cursus mcp serve

Other groups: catalog, dag, config, registry, alignment, exec-doc.

✨ Key Features

🎯 Graph-to-Pipeline Automation

• Input: Simple pipeline graph with step types and connections
• Output: Complete SageMaker pipeline with all dependencies resolved
• Magic: Intelligent analysis of graph structure with automatic step builder selection

⚡ 10x Faster Development

• Before: 2-4 weeks of manual SageMaker configuration
• After: 10-30 minutes from graph to working pipeline
• Result: 95% reduction in development time

🧠 Intelligent Dependency Resolution

• Automatic step connections and data flow
• Smart configuration matching and validation
• Type-safe specifications with compile-time checks
• Semantic compatibility analysis

🛡️ Production Ready

• Built-in quality gates and validation
• Enterprise governance and compliance
• Comprehensive error handling and debugging
• Stable, production-deployed, and published on PyPI

🤖 Agent-Ready (MCP)

• A framework-neutral cursus.mcp tool surface (JSON in / JSON out) exposes the whole engine — catalog, DAG, config, compile, validate, exec-doc, and the pipeline catalog — to LLM agents
• Run it as an MCP server with cursus mcp serve (or python -m cursus.mcp.server)

📊 Proven Results

Based on production deployments across enterprise environments:

Component	Code Reduction	Lines Eliminated	Key Benefit
Processing Steps	60%	400+ lines	Automatic input/output resolution
Training Steps	60%	300+ lines	Intelligent hyperparameter handling
Model Steps	47%	380+ lines	Streamlined model creation
Registration Steps	66%	330+ lines	Simplified deployment workflows
Overall System	~55%	1,650+ lines	Intelligent automation

🏗️ Architecture

Cursus follows a layered, specification-driven architecture. A PipelineDAG (nodes + edges) and a config file flow down through these layers to a finished SageMaker pipeline:

• 🎯 DAG Compiler (core/compiler): compile_dag_to_pipeline / PipelineDAGCompiler — the public entry point, plus validation, preview, and naming
• 🧩 Dynamic Template (core/compiler): binds each DAG node to its typed config and step-builder class via the step catalog and registry — no per-pipeline template class required
• 🏗️ Pipeline Assembler (core/assembler): instantiates each step builder in topological order and wires them together
• 🔗 Dependency Resolver (core/deps): matches each step's declared inputs to upstream outputs by semantic compatibility scoring, producing SageMaker property references
• 🛠️ Step Builders (steps/builders): turn a config + resolved inputs into a concrete SageMaker ProcessingStep / TrainingStep / ModelStep
• 📋 Specification Layer (steps/interfaces): one declarative .step.yaml per step — a unified StepInterface carrying both the script contract (I/O paths) and the dependency/output spec the resolver matches on
• 📚 Registry & Step Catalog (registry, step_catalog): the canonical step-name registry and auto-discovery that map names → builders, configs, and interfaces

📚 Usage Examples

Basic Pipeline

from cursus.core import compile_dag_to_pipeline
from cursus.api import PipelineDAG
from sagemaker.workflow.pipeline_context import PipelineSession

# Create DAG
dag = PipelineDAG()
dag.add_node("CradleDataLoading_training")
dag.add_node("XGBoostTraining")
dag.add_edge("CradleDataLoading_training", "XGBoostTraining")

# Set up SageMaker session
pipeline_session = PipelineSession()
role = "arn:aws:iam::123456789012:role/SageMakerExecutionRole"

# Compile to SageMaker pipeline
pipeline = compile_dag_to_pipeline(
    dag=dag,
    config_path="config.json",
    sagemaker_session=pipeline_session,
    role=role,
    pipeline_name="my-ml-pipeline"
)

Advanced Configuration

from cursus.core import compile_dag_to_pipeline, PipelineDAGCompiler
from cursus.api import PipelineDAG
from sagemaker.workflow.pipeline_context import PipelineSession

# Create DAG with more complex workflow
dag = PipelineDAG()
dag.add_node("CradleDataLoading_training")
dag.add_node("TabularPreprocessing_training")
dag.add_node("XGBoostTraining")
dag.add_node("CradleDataLoading_calibration")
dag.add_node("TabularPreprocessing_calibration")
dag.add_node("XGBoostModelEval_calibration")

# Add edges for training flow
dag.add_edge("CradleDataLoading_training", "TabularPreprocessing_training")
dag.add_edge("TabularPreprocessing_training", "XGBoostTraining")

# Add edges for calibration flow
dag.add_edge("CradleDataLoading_calibration", "TabularPreprocessing_calibration")
dag.add_edge("XGBoostTraining", "XGBoostModelEval_calibration")
dag.add_edge("TabularPreprocessing_calibration", "XGBoostModelEval_calibration")

# Set up SageMaker session
pipeline_session = PipelineSession()
role = "arn:aws:iam::123456789012:role/SageMakerExecutionRole"

# Compile with validation and reporting
compiler = PipelineDAGCompiler(
    config_path="config.json",
    sagemaker_session=pipeline_session,
    role=role
)

# Validate DAG before compilation
validation = compiler.validate_dag_compatibility(dag)
if validation.is_valid:
    print(f"✅ DAG validation passed! Confidence: {validation.avg_confidence:.2f}")
    
    # Compile with detailed report
    pipeline, report = compiler.compile_with_report(
        dag=dag,
        pipeline_name="advanced-ml-pipeline"
    )
    print(f"📊 Pipeline compiled: {report.summary()}")
else:
    print("❌ DAG validation failed:", validation.config_errors)

Using the Pre-built Pipeline Catalog

The pipeline catalog is data-driven: 40+ ready-made DAGs ship as *.dag.json files (indexed by catalog_index.json) and are loaded by ID — no per-pipeline classes. Discover, load, and compile them through cursus.pipeline_catalog:

from cursus.pipeline_catalog import recommend_dag, load_shared_dag, get_all_shared_dags
from cursus.core import compile_dag_to_pipeline
from sagemaker.workflow.pipeline_context import PipelineSession

# Discover a DAG that matches your needs (ranked, scored)
for hit in recommend_dag(framework="xgboost", features=["calibration"]):
    print(hit["id"], hit["score"])

# ...or list everything in the catalog
get_all_shared_dags()  # {dag_id: metadata, ...}

# Load a ready-made DAG by ID and compile it
dag = load_shared_dag("xgboost_simple")

pipeline_session = PipelineSession()
role = "arn:aws:iam::123456789012:role/SageMakerExecutionRole"

pipeline = compile_dag_to_pipeline(
    dag=dag,
    config_path="config.json",
    sagemaker_session=pipeline_session,
    role=role,
    pipeline_name="xgboost-simple",
)
pipeline.upsert()  # Deploy and run!

For a one-call build straight from files, use build_and_compile:

from cursus.pipeline_catalog import build_and_compile

pipeline, report = build_and_compile(
    dag_path="src/cursus/pipeline_catalog/shared_dags/xgboost/simple.dag.json",
    config_path="config.json",
    sagemaker_session=pipeline_session,
    role=role,
)

Using the Compiler Class Directly

from cursus.core import PipelineDAGCompiler
from cursus.pipeline_catalog import load_shared_dag
from sagemaker.workflow.pipeline_context import PipelineSession

# Load a ready-made DAG from the catalog (or build your own PipelineDAG)
dag = load_shared_dag("xgboost_simple")

# Set up SageMaker session
pipeline_session = PipelineSession()
role = "arn:aws:iam::123456789012:role/SageMakerExecutionRole"

# Use compiler for more control
compiler = PipelineDAGCompiler(
    config_path="config.json",
    sagemaker_session=pipeline_session,
    role=role
)

# Preview resolution before compilation
preview = compiler.preview_resolution(dag)
for node, config_type in preview.node_config_map.items():
    confidence = preview.resolution_confidence.get(node, 0.0)
    print(f"   {node} → {config_type} (confidence: {confidence:.2f})")

# Compile the pipeline
pipeline = compiler.compile(dag, pipeline_name="my-pipeline")

🔧 Installation Options

Core Installation

pip install cursus

Includes basic DAG compilation and SageMaker integration.

Framework-Specific

pip install cursus[pytorch]    # PyTorch Lightning models
pip install cursus[gbm]        # GBM training pipelines (XGBoost + LightGBM)
pip install cursus[nlp]        # NLP models and processing
pip install cursus[processing] # Advanced data processing

Development

pip install cursus[dev]        # Development tools
pip install cursus[docs]       # Documentation tools
pip install cursus[all]        # Everything included

🎯 Who Should Use Cursus?

Data Scientists & ML Practitioners

• Focus on model development, not infrastructure complexity
• Rapid experimentation with 10x faster iteration
• Business-focused interface eliminates SageMaker expertise requirements

Platform Engineers & ML Engineers

• 60% less code to maintain and debug
• Specification-driven architecture prevents common errors
• Universal patterns enable faster team onboarding

Organizations

• Accelerated innovation with faster pipeline development
• Reduced technical debt through clean architecture
• Built-in governance and compliance frameworks

📖 Documentation

📚 Complete Documentation Hub

Your gateway to all Cursus documentation - start here for comprehensive navigation

Knowledge Management Philosophy

• Zettelkasten Principles - The knowledge management principles behind our slipbox documentation system, explaining how we organize and connect information for maximum discoverability and organic growth

Core Documentation

• Developer Guide - Comprehensive guide for developing new pipeline steps and extending Cursus
• Design Documentation - Detailed architectural documentation and design principles
• Pipeline Catalog - Comprehensive collection of prebuilt pipeline templates organized by framework and task
• API Reference - Detailed API documentation including core, api, steps, and other components
• Examples - Ready-to-use pipeline blueprints and examples

Quick Links

• Getting Started - Start here for adding new pipeline steps
• Design Principles - Core architectural principles
• Best Practices - Recommended development practices
• Component Guide - Overview of key components
• Validation System - Comprehensive validation framework for pipeline alignment and quality assurance

🤝 Contributing

We welcome contributions! See our Developer Guide for comprehensive details on:

• Prerequisites - What you need before starting development
• Creation Process - Step-by-step process for adding new pipeline steps
• Validation Checklist - Comprehensive checklist for validating implementations
• Common Pitfalls - Common mistakes to avoid

For architectural insights and design decisions, see the Design Documentation.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🔗 Links

• GitHub: https://github.com/TianpeiLuke/cursus
• Issues: https://github.com/TianpeiLuke/cursus/issues
• PyPI: https://pypi.org/project/cursus/

---

Cursus: Making SageMaker pipeline development 10x faster through intelligent automation. 🚀