์Šคํ‚ฌ ml-pipeline-workflow
๐Ÿ”„

ml-pipeline-workflow

์•ˆ์ „

Build Production ML Pipelines

๋˜ํ•œ ๋‹ค์Œ์—์„œ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค: wshobson

Building reliable ML pipelines requires orchestrating data preparation, training, validation, and deployment stages while maintaining reproducibility and monitoring. This skill provides proven patterns and best practices for creating end-to-end MLOps workflows that scale from prototype to production.

์ง€์›: Claude Codex Code(CC)
๐Ÿฅ‰ 75 ๋ธŒ๋ก ์ฆˆ
1

์Šคํ‚ฌ ZIP ๋‹ค์šด๋กœ๋“œ

2

Claude์—์„œ ์—…๋กœ๋“œ

์„ค์ • โ†’ ๊ธฐ๋Šฅ โ†’ ์Šคํ‚ฌ โ†’ ์Šคํ‚ฌ ์—…๋กœ๋“œ๋กœ ์ด๋™

3

ํ† ๊ธ€์„ ์ผœ๊ณ  ์‚ฌ์šฉ ์‹œ์ž‘

ํ…Œ์ŠคํŠธํ•ด ๋ณด๊ธฐ

"ml-pipeline-workflow" ์‚ฌ์šฉ ์ค‘์ž…๋‹ˆ๋‹ค. Design a basic ML pipeline for image classification

์˜ˆ์ƒ ๊ฒฐ๊ณผ:

Here is a 4-stage pipeline architecture: Stage 1 - Data Ingestion (load images, split train/val/test), Stage 2 - Feature Engineering (augmentation, normalization), Stage 3 - Model Training (train CNN with hyperparameter logging), Stage 4 - Deployment (package model, deploy to serving endpoint). Each stage has independent retry logic and outputs versioned artifacts.

"ml-pipeline-workflow" ์‚ฌ์šฉ ์ค‘์ž…๋‹ˆ๋‹ค. How do I implement canary deployment for my model?

์˜ˆ์ƒ ๊ฒฐ๊ณผ:

Canary deployment strategy: 1) Deploy new model to 10% of traffic using weighted routing, 2) Monitor key metrics (latency, error rate, prediction quality) for 1-2 hours, 3) If metrics stay within thresholds, gradually increase to 50%, then 100%, 4) If anomalies detected, automatically rollback to previous model. Use MLflow or similar to track model versions and performance.

๋ณด์•ˆ ๊ฐ์‚ฌ

์•ˆ์ „
v1 โ€ข 2/25/2026

Static analysis detected 17 potential issues (external_commands and weak cryptographic patterns), all of which are false positives. The flagged backtick patterns are markdown code formatting for file path references (e.g., `resources/implementation-playbook.md`), not shell execution. The cryptographic algorithm flags do not correspond to actual cryptographic code in the skill file. This is a documentation-only skill with no executable code, no network access, no file system operations, and no security risks.

1
์Šค์บ”๋œ ํŒŒ์ผ
260
๋ถ„์„๋œ ์ค„ ์ˆ˜
1
๋ฐœ๊ฒฌ ์‚ฌํ•ญ
1
์ด ๊ฐ์‚ฌ ์ˆ˜
๋‚ฎ์€ ์œ„ํ—˜ ๋ฌธ์ œ (1)
SKILL.md:22SKILL.md:73SKILL.md:81SKILL.md:90-103SKILL.md:200-211SKILL.md:215-219SKILL.md:223-227
False Positive: Markdown Code Formatting
Static analyzer flagged 10 instances of 'Ruby/shell backtick execution' at lines 22, 73, 81, 90, 103, 200, 211, 215, 219, 223. These are markdown code spans using backticks to format file paths like `resources/implementation-playbook.md` and `assets/`, not actual shell execution commands. This is standard markdown syntax with no security implications.
๊ฐ์‚ฌ์ž: claude

ํ’ˆ์งˆ ์ ์ˆ˜

38
์•„ํ‚คํ…์ฒ˜
100
์œ ์ง€๋ณด์ˆ˜์„ฑ
87
์ฝ˜ํ…์ธ 
50
์ปค๋ฎค๋‹ˆํ‹ฐ
100
๋ณด์•ˆ
100
์‚ฌ์–‘ ์ค€์ˆ˜

๋งŒ๋“ค ์ˆ˜ ์žˆ๋Š” ๊ฒƒ

Build New ML Pipeline

Create a complete ML pipeline from scratch with proper orchestration, validation, and deployment stages for a production machine learning system.

Orchestrate ML Workflows

Design and implement DAG-based workflow automation for existing ML components using tools like Airflow, Dagster, or Kubeflow Pipelines.

Deploy Models to Production

Implement safe deployment strategies including canary releases, blue-green deployments, and automated rollback mechanisms for ML models.

์ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•ด ๋ณด์„ธ์š”

Create Basic Pipeline 
Design a simple ML pipeline that processes data, trains a model, and deploys it. Include the key stages and dependencies.
Configure DAG Orchestration 
Help me set up a DAG-based ML workflow using [Airflow/Dagster/Kubeflow]. I have [data processing/training/validation] stages that need to run in sequence.
Implement Deployment Strategy 
Design a deployment strategy for my ML model that includes canary testing and automated rollback. The model serves [description] predictions.
Set Up Continuous Training 
Create a continuous training pipeline that retrains my model when data drift is detected. Include monitoring triggers and validation gates.

๋ชจ๋ฒ” ์‚ฌ๋ก€

  • Design each pipeline stage to be modular and independently testable for easier debugging and maintenance
  • Implement idempotent stages so re-running any part of the pipeline is safe and produces consistent results
  • Version all artifacts including datasets, models, and configurations for full reproducibility and rollback capabilities

ํ”ผํ•˜๊ธฐ

  • Avoid monolithic pipeline stages that combine multiple responsibilities, as this makes debugging difficult and reduces reusability
  • Do not skip data validation between stages, as invalid data can cause silent failures downstream that are hard to diagnose
  • Never deploy models directly to 100% production traffic without validation testing or gradual rollout mechanisms

์ž์ฃผ ๋ฌป๋Š” ์งˆ๋ฌธ

What orchestration tool should I use for my ML pipeline?
Choose based on your needs: Airflow for mature DAG workflows with large community support, Dagster for asset-based pipelines with strong data lineage, Kubeflow for Kubernetes-native ML workflows, or Prefect for modern Pythonic workflow automation. This skill supports all four options.
How do I handle data versioning in my pipeline?
Use DVC (Data Version Control) or similar tools to track dataset versions alongside code. Create unique version identifiers for each processed dataset, and log these versions with your trained models for full lineage tracking.
What is the difference between batch and real-time feature pipelines?
Batch pipelines process data on a schedule (e.g., daily) and are simpler to implement and debug. Real-time pipelines use stream processing to compute features on-demand for low-latency predictions. Start with batch, then add real-time features only if your use case requires sub-second prediction latency.
How do I monitor model performance after deployment?
Track prediction latency, error rates, and business metrics. For prediction quality, use ground truth labels when available to compute accuracy/F1/MAE. Set up alerts for data drift detection and performance regression. Most teams use MLflow, Weights & Biases, or custom observability tools.
When should I use canary vs blue-green deployment?
Use canary deployment to test new models on a small percentage of traffic while keeping the old model running. Use blue-green deployment when you need to switch traffic completely between versions. Canary is safer for model changes with uncertain performance impact; blue-green works well for infrastructure updates.
How do I implement automated rollback triggers?
Define thresholds for key metrics (latency p95 < 100ms, error rate < 1%, prediction accuracy within 5% of baseline). Configure monitoring to check these metrics every 5-10 minutes after deployment. If any metric exceeds threshold for 2 consecutive checks, automatically revert to previous model version using your orchestration tool's rollback mechanism.

๊ฐœ๋ฐœ์ž ์„ธ๋ถ€ ์ •๋ณด

์ž‘์„ฑ์ž

sickn33

๋ผ์ด์„ ์Šค

MIT

๋ฆฌํฌ์ง€ํ† ๋ฆฌ

https://github.com/sickn33/antigravity-awesome-skills/tree/main/skills/ml-pipeline-workflow

์ฐธ์กฐ

main

ํŒŒ์ผ ๊ตฌ์กฐ

๐Ÿ“„ SKILL.md