What's Next for Operations and Platform Builders - October 2023

Three Categories of Workloads

AI Workloads

• Vertex AI—foundation model consumers
• AI IaaS—foundation model providers
  • Compute Engine, Kubernetes Engine (GKE) etc.
  • Roll you own AI stack

GKE for AI

• Open standard API
  • Port workloads from data scientist workstation to cloud—reliably reproduce results
• Hugely horizontally scalable
• Autopilot—opinionated config
  • Increase productivity—don’t need to worry about infrastructure
• TPU v4—GA
• TPU v5e—preview

Modern Workloads

• By 2027—90% production apps on containers
  • In post-VM era
• Choices—right tool for job
  • Cloud Run
    • Deploy microservices in seconds
    • No infrastructure
    • Rapid autoscaling
    • Sidecars support
  • Kubernetes Engine (GKE)

Enterprise Workloads

• Challenges:
  • Multiple environments/teams
  • Increase risk of compliance
  • Speed

GKE Enterprise

• RBAC across clusters
• Vulnerability scanning
• Policies/guardrails—GitOps

GKE Interactive Troubleshooting Playbooks

• SRE practices
• Opinionated path to diagnose problems

Loveholidays

• Uptime is an antipattern
• 200+ Compute Engine instances created per hour
• Peaks—150x normal traffic
• Ripley—tool to replay realistic HTTPS traffic
• OwlBot—simulated autoscaling overnight
  • Identifies bottlenecks
• FinOps metric—$ cost to serve 1000 users

Continuous Disaster Recovery

• Create copies of prod—GKE fleets
• Balance load between clusters in multiple regions—Gateway API
• Scaling clusters—on-demand clusters