The GCP Problem

§ 00 — BEFORE GCPGoogle's internal infrastructure made public

GCP isn't a cloud that Google built — it's Google's own infrastructure that Google eventually opened to the world. Every major GCP service traces back to a paper, a system, or a problem that Google had to solve at planetary scale first.

§ 01 — THE QUESTIONDesign a globally distributed data platform on GCP

§ 02 — GLOBAL INFRASTRUCTUREThe Google backbone advantage

GCP's defining edge: traffic on the Premium Tier never leaves Google's private fiber — it enters at the nearest PoP and rides the backbone all the way to the destination region.

§ 03 — COMPUTEDecision tree: what are you running?

GCP compute choice flows from workload shape — not vendor defaults. Start with the question, not the service.

§ 04 — STORAGEGCS tiers and the storage decision matrix

GCS storage classes are priced on retrieval frequency — not access speed. Archive retrieval takes milliseconds, not hours (unlike AWS Glacier).

§ 05 — BIGQUERYGCP's crown jewel

BigQuery's secret: compute and storage are completely separate. A 100TB query doesn't saturate storage nodes — Dremel fan-out across thousands of workers happens in parallel on Jupiter's fabric.

§ 06 — NETWORKINGGlobal VPC — the biggest differentiator from AWS

In AWS, a VPC is regional — you need VPC peering or Transit Gateway to connect regions. In GCP, a single VPC spans all regions globally. One VPC, one firewall ruleset, subnets in every region.

§ 07 — SERVERLESS & MESSAGINGPub/Sub, Cloud Run, Dataflow

GCP's serverless stack is three distinct layers: Pub/Sub for durable messaging, Cloud Run for HTTP containers, Dataflow for stream/batch pipelines. Each is independently scalable.

§ 08 — IAMGCP's resource hierarchy and identity model

GCP IAM is a hierarchy — policies set at a parent propagate down. You cannot remove a permission granted at a higher level from a lower level.

§ 09 — Q&A12 interview questions

1. When would you choose GCP over AWS?

Choose GCP when BigQuery analytics is central (no AWS equivalent at the same scale-to-zero price model), when global network latency matters (premium tier keeps traffic on Google's backbone), or when your ML workload benefits from TPUs and Vertex AI. Also when your org is Google Workspace-native — IAM integration is seamless.

2. What makes BigQuery's architecture unique?

Dremel separates compute (workers) from storage (Colossus) completely — Jupiter's petabit network fabric means no I/O bottleneck. You can query 100TB without provisioning anything: Dremel fans out to thousands of slots on demand, and you pay only for bytes scanned. No indexes, no vacuuming, no cluster sizing.

3. GCP VPC vs AWS VPC — key difference?

GCP VPC is global: one VPC spans all regions; subnets are regional but belong to the same global network. AWS VPC is regional — cross-region requires Transit Gateway or VPC peering. GCP firewall is tag-based (not per-instance security groups), stateful, and applied globally to the VPC.

4. What is Cloud Spanner and when do you use it?

Cloud Spanner is a globally distributed, externally consistent RDBMS — the only one with a 99.999% SLA across regions. Use it when you need relational semantics (SQL, foreign keys, transactions) AND horizontal scale AND global consistency. Example: global financial ledger, inventory system that must never show negative stock globally.

5. Pub/Sub vs Kafka — when to choose each?

Choose Pub/Sub when you want zero infrastructure management, global delivery, and auto-scaling at any throughput — it's the right default on GCP. Choose Kafka (via Confluent or self-managed) when you need ordering guarantees per partition with consumer group offsets, log compaction, or Kafka Streams — or when migrating an existing Kafka architecture.

6. What is Anthos?

Anthos is GCP's multi-cloud/on-prem management platform. It runs GKE clusters on AWS, Azure, or bare metal, using Config Sync (GitOps), Policy Controller (OPA), and Service Mesh (Istio) to provide consistent operations across environments. Use it when you need a single control plane for workloads that span clouds.

7. How does Cloud Run differ from Cloud Functions?

Cloud Run deploys any OCI container — you control the runtime, dependencies, and can handle multiple concurrent requests per instance (up to 1000). Cloud Functions deploys source code into a managed runtime — simpler, but less control. Cloud Run Gen 2 and Cloud Functions Gen 2 share the same underlying infrastructure (Cloud Run); Functions is now a packaging layer on top.

8. What is a TPU and when would you use one?

A Tensor Processing Unit is Google's custom ASIC designed specifically for matrix multiply at neural-network scales. Use TPUs when training or fine-tuning large models in JAX or TensorFlow — they offer 10–100× better throughput/$ vs GPUs for large batch matrix ops. Not useful for inference at small batch size or for non-TF/JAX frameworks.

9. How do you design a BigQuery schema for cost optimization?

Partition by event date (eliminates full-table scans), cluster by the highest-cardinality filter column (usually user_id or entity_id), select only needed columns (BQ is columnar — never SELECT *), and use materialized views for repeated aggregations. Reserve slots for production workloads to cap costs; use on-demand for exploration.

10. What is VPC Service Controls?

VPC Service Controls creates a security perimeter around GCP API access — even a user with IAM permissions cannot exfiltrate data if their request originates outside the perimeter (wrong VPC, wrong IP range, wrong device). Essential for PCI, HIPAA, and FedRAMP workloads where data residency and exfiltration prevention are hard requirements.

11. How does GCP's premium network tier work?

When a packet enters GCP's network, it's routed onto Google's private fiber at the nearest PoP and carried entirely on Google's backbone to the destination region — never touching the public internet. Standard tier exits to the public internet immediately. Premium tier reduces latency (especially cross-continental) and avoids internet routing instabilities, but costs ~25% more per egress GB.

12. What is Workload Identity Federation?

Workload Identity Federation lets external identities — AWS IAM roles, Azure managed identities, GitHub Actions OIDC tokens — exchange their native credential for a short-lived GCP access token via an OIDC/SAML trust. Zero service account keys needed. This is the right pattern for CI/CD pipelines, cross-cloud integrations, and any workload that can't use Workload Identity (the K8s-specific variant).

§ 10 — COMMON MISTAKESWhat interviewers actually hear

These are the three misconceptions that immediately signal a candidate hasn't worked deeply with GCP — each one has a specific correction that demonstrates real understanding.

§ 11 — WHY NOT?Honest trade-offs — when to choose GCP and when not to

A strong candidate knows when GCP wins and when it doesn't. Defaulting to GCP for everything signals lack of real-world experience as much as defaulting to AWS does.

§ 12 — ONE-MINUTE ANSWERWhen the interviewer says "just quickly — what makes GCP different?"

This is the answer you should be able to give in under 90 seconds — specific, technical, and honest about trade-offs.

Question

"What makes GCP different from AWS?"

Answer

"GCP's differentiation comes from two sources: Google's internal infrastructure made public, and ML/AI leadership. BigQuery came from Google's internal Dremel query engine. Kubernetes came from Google's internal Borg scheduler. The global VPC architecture reflects how Google itself operates — a single network spanning all regions rather than isolated regional silos. For data and AI workloads, GCP frequently wins on price-performance because you're using the same infrastructure Google uses for Search, YouTube, and Maps. The tradeoff: smaller ecosystem than AWS, fewer enterprise integrations, and Google's history of sunsetting products creates enterprise hesitancy."

§ 13 — INTERVIEWER'S MINDWhat they're actually probing with each question

GCP interviewers are not testing service name recall. They are testing whether you understand the why behind each architectural choice. Here are the four most common probes and what depth looks like.

§ 14 — THE EVOLUTIONGCP's architectural arc: 2003 → 2023

Understanding GCP's evolution explains why its services are designed the way they are. Each milestone is Google solving a real problem at scale — and then opening that solution to the world.

§ 15 — WHAT'S NEXT?The frontier GCP is pushing into

GCP's next act isn't just more services — it's collapsing the boundary between the data layer and the AI layer. Every major service is being retrofitted to become AI-native infrastructure.

§ 16 — SUMMARYWeak vs Strong answer