mlops-lifecycle-manager_skill

Overview

This skill manages the ML model lifecycle from development through production with experiment tracking, versioning, monitoring, and automated retraining workflows. It provides prescriptive recommendations and runnable artifacts to move models from notebook experiments to robust, governed deployments. Use it to define pipelines, monitoring, governance, and cost-aware deployment patterns.

How this skill works

The skill analyzes model type, deployment target, scale and governance requirements, and existing stack to select appropriate tools and patterns (MLflow, Kubeflow, KServe, Vertex AI, SageMaker, etc.). It produces tiered deliverables: quick assessment and checklist (T1), production pipeline designs and manifests (T2), and comprehensive governance, retraining automation, and observability playbooks (T3). Outputs include tracking URIs, model registry schema, CI/CD pipeline templates, Kubernetes manifests, and monitoring dashboard queries.

When to use it

• Transitioning a model from experiment/notebook to production deployment
• Detecting production performance degradation that requires retraining or rollback
• Establishing experiment tracking and reproducible versioning for a team
• Designing an automated end-to-end pipeline from data to inference
• Implementing governance, audit trails, and compliance for regulated use cases
• Optimizing deployment patterns for scale, latency, or edge constraints

Best practices

• Validate inputs: model_type, deployment_target, scale_requirements, governance_level, and optionally existing_stack before recommending a stack
• Start simple for small teams: MLflow + basic deployment patterns, then evolve to Kubeflow/managed services as maturity grows
• Enforce model registry metadata: semantic version, git SHA, data hash, hardware/inference requirements
• Monitor both data drift and model performance with thresholded alerts and labeled-sample checks
• Automate retraining triggers (schedule, drift, performance drops) and require conditional champion-challenger promotion with approval gates
• Optimize cost: autoscale serving, use spot instances for training, and apply quantization/pruning for latency-sensitive targets

Example use cases

• Quick assessment and checklist to deploy an NLP classifier to online KServe endpoints with MLflow tracking
• Designing a production pipeline for a recommendation system with Feast feature store, autoscaling, and canary rollouts
• Implementing automated retraining for time-series forecasting with scheduled retrains and drift-triggered jobs using Kubeflow Pipelines
• Creating governance artifacts for a regulated model: model card, lineage diagrams, RBAC, and audit logs
• Migrating ad-hoc batch scoring jobs into a reproducible CI/CD flow with model registry, monitoring, and cost optimization