01 · GCP Architecture Framework¶

Google publishes the Architecture Framework — five pillars for evaluating any GCP design: Operational Excellence, Security/Privacy/ Compliance, Reliability, Cost Optimization, and Performance Optimization. This module isn't new CLI syntax; it's the lens for reviewing everything built in Levels 1-3, plus the tooling GCP gives you to audit against it.

The five pillars, applied¶

Pillar	Question it asks	GCP tooling
Operational Excellence	Can we deploy, observe, and recover without heroics?	Cloud Build, Cloud Monitoring, Cloud Logging, runbooks
Security, Privacy, Compliance	Least privilege? Data protected at rest/in transit/in use?	IAM, VPC-SC, SCC, CMEK, Assured Workloads
Reliability	Does it survive a zone/region failure? What's the RTO/RPO?	Multi-region, health checks, SLOs, chaos testing
Cost Optimization	Are we paying for capacity we use?	Committed use discounts, rightsizing recommender, budgets
Performance Optimization	Does it scale to 10x load without a redesign?	Load testing, autoscaling, caching, CDN

Every design decision trades between these — a single-region deployment optimizes cost and simplicity at the expense of reliability; that's a valid choice if stated explicitly, not a default made by omission.

Using the Architecture Review checklist in practice¶

Google's gcloud recommender surfaces automated suggestions mapped to several pillars directly from your actual usage:

gcloud recommender recommendations list \
  --project=my-project \
  --recommender=google.compute.instance.MachineTypeRecommender \
  --location=us-central1-a

# NAME                                          PRIMARY_IMPACT
# .../recommendations/abcd-1234                 COST: reduce cost by resizing e2-standard-8 to e2-standard-4

gcloud recommender recommendations list \
  --project=my-project \
  --recommender=google.iam.policy.Recommender \
  --location=global
# .../recommendations/efgh-5678   SECURITY: remove unused role roles/editor from user:dev@example.com

Recommenders don't auto-apply — each recommendation includes an associatedInsights reference and an operation you review and apply explicitly:

gcloud recommender recommendations mark-claimed \
  --recommendation=abcd-1234 \
  --recommender=google.compute.instance.MachineTypeRecommender \
  --location=us-central1-a \
  --etag=ETAG_VALUE

Gotcha — recommenders need history to generate signal. A newly created project or resource often shows zero recommendations for weeks — the underused-resource and IAM recommenders both need a baseline observation window (typically ~28 days) before they have enough data to recommend anything. Don't read "no recommendations" as "this is optimal" for a young project.

A worked trade-off: the same service, two ways¶

Design A — cost-optimized: single-region Cloud Run, Cloud SQL with no replica, no CDN. Cheapest, RTO measured in hours (restore from backup), no protection from a regional outage.

Design B — reliability-optimized: two-region Cloud Run behind a global LB, Spanner multi-region, Cloud CDN. RTO near-zero, but Spanner's per-node cost and cross-region write latency are the price, and the design is meaningfully harder to operate (more moving parts to monitor).

Neither is "correct" in isolation — Design A is right for an internal tool where a few hours of downtime is a non-event; Design B is right for a customer-facing payments path. The Architecture Framework's value is forcing that justification to be explicit and reviewed, rather than inherited from whatever the last engineer happened to build.

Landing zone concept¶

A landing zone is the pre-built org structure — folders, projects, baseline IAM, baseline org policies, logging sinks — that new workloads land into, so every new project starts compliant instead of getting audited into compliance after the fact.

gcloud resource-manager folders create --display-name="Production" --organization=123456789012
gcloud resource-manager folders create --display-name="Non-Production" --organization=123456789012
gcloud resource-manager folders create --display-name="Shared-Services" --organization=123456789012

gcloud resource-manager folders list --organization=123456789012
# DISPLAY_NAME        PARENT_ID
# Production           organizations/123456789012
# Non-Production       organizations/123456789012
# Shared-Services      organizations/123456789012

Org policies (Level 3, Module 04) and default IAM bindings attached at the folder level inherit automatically to every project created underneath — so "Production" can enforce compute.vmExternalIpAccess: denyAll and every project in it inherits that without per-project setup. Module 05 of this level covers folder/project structuring in depth.

Cheat sheet¶

Command	Purpose
`gcloud recommender recommendations list`	Pull automated cost/security/performance suggestions.
`gcloud recommender recommendations mark-claimed`	Record that you're acting on a recommendation.
`gcloud resource-manager folders create`	Build landing-zone folder structure under an org.
Five pillars (Ops, Security, Reliability, Cost, Performance)	The lens for every architecture review.

How It Actually Works¶

The Well-Architected Framework's five pillars aren't independent checklists — they're competing forces that architecture decisions actually trade off against each other in the underlying infrastructure. Adding redundancy for reliability (multi-zone, multi-region deployment) mechanically increases cost (more running resources) and often operational complexity (more failure modes to reason about, more replication lag to manage), which is why a well-architected system explicitly documents which pillar it's optimizing for a given component rather than trying to maximize all five simultaneously. Performance efficiency and cost optimization intersect at the resource- sizing level in a very literal way: GCP bills machine types and disk tiers on fixed pricing, so the "efficient" configuration is the smallest resource that meets your latency SLA, discoverable only by load-testing against real traffic patterns — architectural review at this level is fundamentally an exercise in reading utilization metrics against SLA targets, not applying abstract principles.

Exercise¶

Take the multi-tier CI/CD project from Level 3 Module 10 and score it against all five pillars — one sentence each on what it does well and one gap. Then run (or describe, if quota-limited) gcloud recommender recommendations list against a real or sample project for both the MachineTypeRecommender and IamRecommender, and note what pillar each returned suggestion maps to.