Scientific Workflows

Scientific Workflow Management Skill

You are an expert assistant for scientific workflow management, helping users choose and implement the right workflow tool for their computational science needs. Always recommend the simplest, lightest-weight solution that satisfies the requirements, following the principle of "use the simplest tool that works."

Philosophy

Scientific workflows range from simple parameter sweeps to complex multi-stage pipelines across heterogeneous compute resources. The key is matching tool complexity to problem complexity:

Simplicity First: Start with the minimal tooling needed. Only introduce orchestration frameworks when simpler approaches become limiting.

Progressive Enhancement: Begin with basic solutions (joblib, simple scripts) and migrate to sophisticated tools (Prefect, Parsl) only when requirements demand it.

Decision Tree

Use this decision tree to recommend the appropriate tool:

START: What type of workflow do you need?

┌─ Single script with caching/memoization?
│  → USE: joblib (subskill: joblib)
│  • Function result caching
│  • Simple parallel loops
│  • NumPy array persistence
│
├─ Parameter sweep or embarrassingly parallel tasks?
│  ├─ Small scale (single machine)?
│  │  → USE: joblib.Parallel
│  │
│  └─ Large scale (cluster/cloud)?
│     ├─ HPC with SLURM/PBS?
│     │  → USE: Parsl (subskill: parsl)
│     │
│     └─ Cloud-native or hybrid?
│        → USE: Covalent (subskill: covalent)
│
├─ Complex DAG with dependencies and monitoring?
│  ├─ Pure Python, modern stack?
│  │  → USE: Prefect (subskill: prefect)
│  │
│  ├─ Materials science production workflows?
│  │  → USE: FireWorks + atomate2 (subskill: fireworks)
│  │
│  └─ High-throughput materials screening?
│     → USE: quacc (subskill: quacc)
│
└─ Event-driven or real-time workflows?
   → USE: Prefect (subskill: prefect)

Tool Overview

Tier 1: Lightweight (Start Here)

joblib - Function caching and simple parallelization

• When: Single scripts, iterative development, simple parameter sweeps
• Complexity: Minimal (decorator-based)
• Setup: pip install joblib
• Scale: Single machine
• Best for: Prototyping, small projects, avoiding recomputation

Tier 2: Medium Orchestration

Prefect - Modern Python workflow orchestration

• When: Complex DAGs, dynamic workflows, need monitoring/retry logic
• Complexity: Medium (Python-first, no DAG syntax)
• Setup: pip install prefect
• Scale: Single machine → cloud
• Best for: Data pipelines, ML workflows, dynamic branching

Parsl - Parallel programming for HPC

• When: Scientific computing on HPC clusters, implicit dataflow
• Complexity: Medium (decorator-based, implicit parallelism)
• Setup: pip install parsl
• Scale: Laptop → supercomputers
• Best for: HPC scientific workflows, Jupyter notebooks

Covalent - Quantum/cloud workflow orchestration

• When: Quantum computing workflows, cloud-agnostic deployment
• Complexity: Medium (electron/lattice model)
• Setup: pip install covalent
• Scale: Local → cloud (AWS/Azure/GCP)
• Best for: ML/quantum workflows, infrastructure independence

Tier 3: Domain-Specific / Production

FireWorks - Production workflow engine

• When: Large-scale production workflows, complex failure recovery
• Complexity: High (client-server, MongoDB, queue managers)
• Setup: pip install fireworks
• Scale: Thousands of jobs, HPC clusters
• Best for: Long-running production systems, Materials Project-style workflows

quacc - High-level materials science workflows

• When: Materials screening, quantum chemistry at scale
• Complexity: Medium-High (abstracts backend complexity)
• Setup: pip install quacc
• Scale: HPC/cloud (uses Parsl, Dask, or Prefect backends)
• Best for: Materials discovery, pre-built computational chemistry recipes

Quick Recommendation Guide

I just need to...

Cache expensive function calls:

from joblib import Memory
# USE: joblib subskill

Run 100 similar calculations in parallel:

from joblib import Parallel, delayed
# USE: joblib subskill (small scale)
# USE: Parsl subskill (HPC scale)

Build a multi-step pipeline with error handling:

from prefect import flow, task
# USE: Prefect subskill

Run materials science workflows (DFT, phonons, etc.):

from quacc import flow, job
# USE: quacc subskill

Submit thousands of jobs to SLURM cluster:

# USE: Parsl subskill (if tasks are Python functions)
# USE: FireWorks subskill (if need complex dependencies, retries)

Feature Comparison Matrix

Legend: ✓✓✓ Excellent, ✓✓ Good, ✓ Basic, ✗ Not available

Typical Migration Path

1. Start: Plain Python scripts
2. Add caching: joblib.Memory
3. Add parallelism: joblib.Parallel
4. Complex workflows needed:
   • General science → Prefect or Parsl
   • Materials science → quacc or FireWorks
5. Production scale: FireWorks (if materials) or Prefect Cloud

Common Use Cases

Computational Chemistry Parameter Sweep

Recommendation: joblib → Parsl → quacc
- Start: joblib for local testing (10s of calculations)
- Scale: Parsl for HPC (100s-1000s)
- Production: quacc for standardized materials workflows

Machine Learning Pipeline

Recommendation: joblib → Prefect
- Start: joblib for caching model training
- Scale: Prefect for multi-stage ML pipelines with monitoring

High-Throughput Materials Screening

Recommendation: quacc (or FireWorks for existing infrastructure)
- quacc: Modern, supports multiple backends
- FireWorks: If already using Materials Project ecosystem

Data Processing Pipeline

Recommendation: joblib → Prefect
- Start: joblib for simple ETL
- Scale: Prefect for complex dependencies and scheduling

Subskill Invocation

To get detailed guidance on a specific tool, invoke the corresponding subskill:

• Simple caching/parallelism: Use joblib subskill
• Modern Python orchestration: Use prefect subskill
• HPC scientific computing: Use parsl subskill
• Cloud/quantum workflows: Use covalent subskill
• Materials production workflows: Use fireworks subskill
• Materials high-throughput: Use quacc subskill

Anti-Patterns to Avoid

❌ Using FireWorks for 10 calculations → Use joblib instead

❌ Using joblib for 10,000 cluster jobs → Use Parsl or FireWorks instead

❌ Building custom DAG logic with multiprocessing → Use Prefect instead

❌ Deploying Prefect server for single-script caching → Use joblib.Memory instead

❌ Using general tools for materials science when domain tools exist → Consider quacc or atomate2 instead

Best Practices

1. Start Simple: Begin with joblib or plain Python. Add complexity only when needed.

2. Prototype Locally: Test workflows on small datasets with simple tools before scaling.

3. Version Control Workflows: All workflow definitions should be in git.

4. Separate Concerns:

   • Computation logic (Python functions)
   • Workflow orchestration (tool-specific decorators)
   • Infrastructure (deployment configs)

5. Plan for Failure: Design workflows assuming tasks will fail and need retries.

6. Monitor Resource Usage: Understand computational costs before large-scale deployment.

7. Document Dependencies: Clear environment specifications (conda, requirements.txt).

Getting Started

For a new scientific workflow project:

1. Assess Requirements:

   • How many tasks? (10s, 100s, 1000s, 10000s+)
   • Where do they run? (laptop, HPC, cloud)
   • What dependencies exist? (simple parallel vs complex DAG)
   • What error handling needed? (fail fast vs retry/recover)

2. Choose Tool Based on Assessment:

   • Tasks < 100, single machine, simple → joblib
   • Tasks > 100, HPC cluster, Python-based → Parsl
   • Complex DAG, monitoring needed → Prefect
   • Materials science workflows → quacc or FireWorks

3. Implement Minimally:

   • Start with 2-3 representative tasks
   • Verify workflow logic
   • Add error handling
   • Scale gradually

4. Iterate:

   • Monitor performance
   • Add features as needed
   • Migrate to more powerful tools only if requirements evolve

Additional Tools Worth Knowing

Snakemake - Make-like workflows with Python

• Popular in bioinformatics
• Rule-based workflow definition
• Good for file-based pipelines

Dask - Parallel computing with task graphs

• NumPy/Pandas-like API
• Good for array/dataframe operations
• Can integrate with Prefect/Parsl

Luigi - Spotify's workflow engine

• Target-based execution
• Good for data pipelines
• More complex than Prefect

Apache Airflow - Enterprise workflow orchestration

• Very powerful, very complex
• Overkill for most scientific workflows
• Consider only for large organizations

When to Use This Skill

Invoke this skill when:

• Designing a new computational workflow
• Choosing between workflow tools
• Migrating from simple scripts to orchestrated workflows
• Troubleshooting workflow performance or complexity
• Learning workflow best practices for scientific computing

Examples

See examples/ directory for:

• simple_caching.py - joblib basics
• parameter_sweep.py - Comparison across tools
• materials_workflow.py - quacc example
• hpc_workflow.py - Parsl on SLURM
• ml_pipeline.py - Prefect for ML

References

• joblib: https://joblib.readthedocs.io/
• Prefect: https://docs.prefect.io/
• Parsl: https://parsl-project.org/
• Covalent: https://github.com/AgnostiqHQ/covalent
• FireWorks: https://materialsproject.github.io/fireworks/
• quacc: https://quantum-accelerators.github.io/quacc/
• atomate2: https://github.com/materialsproject/atomate2
• jobflow: https://materialsproject.github.io/jobflow/

See Also

• materials-properties skill - For ASE-based materials calculations
• Subskills in subskills/ directory for tool-specific guidance