Skip to main content

Week 10 — MLOps on GCP Part 2: Deploy & Monitor

Week 9 covered the foundations — GCP setup, Vertex AI training pipelines, and experiment tracking. This week we focus on the deployment and operations side of MLOps: getting models into production, keeping them healthy, reacting to events automatically, managing container artifacts securely, and doing all of this cost-effectively.

Pages​

#PageDescription
1Cloud Run for MLDeploy FastAPI model servers, GPU support, traffic splitting, autoscaling
2Model MonitoringVertex AI Model Monitoring, drift & skew detection, alerting
3Pub/Sub + Cloud FunctionsEvent-driven triggers, automated retraining, Cloud Scheduler
4Artifact RegistryDocker image registry, CI/CD push, vulnerability scanning
5Cost OptimizationFree tier, committed use discounts, spot instances, budget alerts

Learning Outcomes​

By the end of this week you will be able to:

  • Deploy a FastAPI-based ML model server to Cloud Run with autoscaling and traffic splitting
  • Configure Vertex AI Model Monitoring to detect feature drift and prediction skew
  • Build event-driven pipelines using Pub/Sub, Cloud Functions, and Cloud Scheduler
  • Manage Docker images in Artifact Registry with versioning and vulnerability scanning
  • Optimize GCP spending through free tier usage, spot instances, budget alerts, and cost dashboards

Prerequisites​

  • GCP account with billing enabled (Week 9)
  • Familiarity with Docker (Week 2)
  • Understanding of Vertex AI pipelines (Week 9)
  • Basic Python and FastAPI knowledge (Week 2)
Lab Connection

Lab 8 — Full MLOps on GCP ties everything together: BigQuery ML training, Cloud Run serving, monitoring, and automated retraining. Start it after completing pages 1-3.

Architecture Overview​

code
┌─────────────┐     ┌──────────────┐     ┌──────────────┐
│  Cloud Run   │◄────│   Pub/Sub    │◄────│ Cloud        │
│  (ML Server) │     │  (Events)    │     │ Scheduler    │
└──────┬───────┘     └──────┬───────┘     └──────────────┘
       │                    │
       ▼                    ▼
┌──────────────┐     ┌──────────────┐
│   Model      │     │   Cloud      │
│  Monitoring  │     │  Functions   │
│  (Drift/Skew)│     │ (Retraining) │
└──────────────┘     └──────────────┘
       │
       ▼
┌──────────────┐
│   Artifact   │
│  Registry    │
│  (Images)    │
└──────────────┘

All components are managed GCP services — no servers to maintain, pay only for what you use.