Implement Robust Model Versioning Strategies

This skill covers machine learning model versioning strategy - lifecycle management, deployment patterns, and MLOps best practices. It adapts semantic versioning (MAJOR.MINOR.PATCH) for models specifically: MAJOR for breaking changes to API, input/output schema, or architecture; MINOR for new features or backward-compatible improvements; PATCH for bug fixes and hotfixes. Comprehensive metadata tracking per version covers training data version and hash, feature-engineering pipeline version, hyperparameters, training metrics, framework versions, training duration and compute resources, and deployment environment requirements.

It provides a Git-based versioning implementation (a deterministic hash combining model bytes and sorted config for reproducible model identification) and an MLflow-backed model registry (registering a new version with logged parameters and metrics, retrieving a version by number or stage, and comparing two versions' metrics side by side). Deployment versioning covers blue-green Kubernetes deployment (version-labeled deployments and pods for zero-downtime releases) and weighted A/B testing (a router that deterministically assigns a user to a model version based on configured traffic weights, so the same user always sees the same version). Model lineage tracking uses a dependency graph connecting each model version to its parent models and data sources, supporting impact analysis - finding every downstream model affected by a change to a given component. An automated rollback manager deploys a new version, monitors its health over a configurable window, and automatically reverts to the prior production version and alerts if health checks fail.

def generate_model_hash(model_path: str, config: Dict) -> str:
    """Generate deterministic hash for model + config"""
    with open(model_path, 'rb') as f:
        model_bytes = f.read()
    
    config_str = json.dumps(config, sort_keys=True)
    combined = model_bytes + config_str.encode()
    return hashlib.sha256(combined).hexdigest()[:12]

Use this skill when versioning ML models for production - applying semantic versioning and metadata tracking, setting up an MLflow registry, implementing blue-green or A/B-tested deployments, tracking model lineage for impact analysis, or building automated rollback on deployment failure.

It is not a fit for the model training or experimentation phase itself - it assumes you already have a trained model and focuses on how to version, deploy, monitor, and roll it back safely once it's ready for production.

Inputs are a trained model artifact, its training metadata, and its position in the model lineage graph. Outputs are a versioned, hashed model registered in a model registry, blue-green or A/B-tested deployment configuration, a lineage graph supporting impact analysis, and an automated rollback system with health monitoring.

MLOps engineers managing production model deployments who need concrete versioning and rollback infrastructure - semantic versioning adapted for models, an MLflow-backed registry, weighted A/B routing, and automated health-monitored rollback - following eight best practices: version everything (models, data, code, environment), never modify existing versions, use gradual canary rollouts, test comprehensively, document changelogs, monitor continuously with rollback triggers, define retention policies, and apply version-specific access controls.