Model Deployment Strategies in MLOps¶

Introduction¶

Model deployment is the process of making a trained model available to a user, system, job, or API. The strategy should match latency needs, risk, and rollback requirements.

Why This Matters¶

A high-quality model can still fail if released badly. Deployment strategy controls blast radius, traffic exposure, validation, and recovery.

Core Concepts¶

Strategies include batch inference, real-time inference, shadow deployment, canary deployment, blue/green deployment, and rollback.

Practical Example¶

A canary can start with one replica of the new version:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: churn-api-v2
spec:
  replicas: 1
  template:
    spec:
      containers:
      - name: api
        image: registry.example.com/churn-api:v2

How This Fits in a Production Workflow¶

Choose the strategy before release. APIs need health checks, latency SLOs, error-rate alerts, model-version logs, and a rollback command.

Common Mistakes¶

• Using real-time serving when batch scoring is enough.
• Sending all traffic to a new model without rollback.
• Comparing canary models without logging model version.
• Ignoring business metrics after a technically successful release.

Quick Checklist¶

• Is inference batch or real-time?
• Is model version logged?
• Is rollback tested?
• Are latency and error-rate alerts active?
• Is canary or shadow evaluation needed?

Related Guides¶

• Batch vs Real-Time Inference Explained
• Deploy an ML Model with FastAPI
• Kubernetes for MLOps

Summary¶

Learn practical model deployment strategies including batch inference, real-time APIs, shadow deployments, canaries, blue/green releases, and rollback.