3.5. Reflection Checkpoint

Key Takeaways

• Compute platform selection is a spectrum from full control (Compute Engine) to fully managed (Cloud Run functions) — match the workload's control needs and operational tolerance to the right point on that spectrum, don't default to whichever option is newest.
• Data product selection narrows quickly once you answer transactional-vs-analytical, structured-vs-flexible, and single-region-vs-global for the workload described.
• The VPC-is-global-but-subnets-are-regional split is the foundation for reasoning about every other networking decision in this domain, including why adding a region means adding a subnet, not a new VPC.
• Firewall rules are stateful by default, and match criteria can target tags or service accounts instead of brittle IP addresses for dynamically-scaled infrastructure.
• Infrastructure as Code and AI-assisted tooling solve different problems — repeatability/auditability versus acceleration of the planning and authoring process itself.

Connecting Forward

Phase 4 shifts from planning these same resource categories to operating them day to day — managing running compute, storage, and networking resources, and building the monitoring and logging visibility to know when something needs attention.

Self-Check Questions

• Without re-reading 3.3.1, can you explain why Shared VPC and VPC Network Peering solve superficially similar problems but aren't interchangeable?
• If a colleague assumes Spot VMs are simply "Compute Engine but cheaper" with no other implications, what would you correct about that understanding using 3.1.2?