3.1.1. Choosing a Compute Platform
💡 First Principle: The exam's compute-selection questions are almost always a translation exercise — a scenario describes a workload's shape (stateful vs. stateless, containerized vs. not, event-triggered vs. always-on) and asks you to match it to the platform built for that shape, not to pick whichever platform sounds most modern.
Compute Engine fits workloads that need OS-level control, specific licensing, or persistent local state — think legacy applications or specialized software with strict environment requirements. Google Kubernetes Engine (GKE) fits containerized workloads that need orchestration at scale — many services, complex networking between them, or fine-grained scheduling control — with GKE Autopilot trading some of that control for Google managing node provisioning entirely, versus GKE Standard where you manage node pools yourself. Cloud Run fits stateless containerized services that scale to zero and back, without needing full Kubernetes orchestration. Cloud Run functions fits small, event-triggered, source-based code that doesn't warrant packaging as a container at all. Agent Runtime on the Gemini Enterprise Agent Platform is the newest addition, purpose-built for hosting AI agents rather than general-purpose application code.
⚠️ Exam Trap: A scenario mentioning "we don't want to manage any servers or Kubernetes concepts at all, just deploy our container" is pointing at Cloud Run, not GKE Autopilot — Autopilot still exposes Kubernetes concepts (Pods, node pools indirectly) even though it manages the nodes themselves, while Cloud Run abstracts Kubernetes away entirely.
Reflection Question: A team has a stateless containerized web API that receives unpredictable, spiky traffic and wants to pay nothing when there's no traffic at all. Which compute platform fits that description best, and what about GKE Standard makes it a worse fit for this specific scenario?
