Hard Difficulty Questions

Do not automate — the manual task is only 10 hours per month and the 4-week build cost is too high to justify

Why wrong: SRE philosophy strongly favors eliminating toil through automation. 10 hours/month of toil is significant, and the automation investment pays back within ~16 months and then provides permanent relief. The long-term value far exceeds the short-term cost.

Hire an additional junior engineer to perform the manual task more efficiently instead of automating

Why wrong: Hiring more people to do toil is the opposite of SRE philosophy. It scales costs proportionally with growth and doesn't address the root cause. SRE explicitly avoids 'throwing people at the problem' when automation is feasible.

The team cannot make this decision without knowing the exact annual salary cost of the engineers who perform the manual task

Why wrong: While cost calculations are relevant, the SRE framework provides a clear principle: eliminate automatable toil. The decision doesn't require precise salary data — the direction (build automation) is clear from SRE principles.

Hybrid cloud

Why wrong: Hybrid cloud combines on-premises infrastructure with public cloud. Using two different public cloud providers (AWS + Google Cloud) is multi-cloud, not hybrid — there's no on-premises component described.

Multi-region

Why wrong: Multi-region typically describes using multiple geographic regions within the same cloud provider. Using two different providers (AWS and Google) is multi-cloud.

Distributed cloud

Why wrong: Distributed cloud extends a public cloud provider's services to edge locations or customer premises — a specific architecture, not the multi-provider strategy described.

Lower insurance premiums because cloud infrastructure costs less than data centers.

Why wrong: Infrastructure cost is one factor, but the competitive advantage described is primarily innovation speed and operational efficiency — not just infrastructure economics.

Better regulatory compliance because cloud providers have more compliance certifications.

Why wrong: Compliance certifications help with regulatory requirements but are not the competitive advantage in this scenario — innovation speed and operational efficiency are.

Access to more insurance actuarial data than the traditional insurer.

Why wrong: Data volume access is not inherently a cloud advantage. The advantage is in the ability to analyze and act on data faster with AI/ML, not simply having more data.

Cloud Logging — search logs for error messages across all 15 services.

Why wrong: Log searching finds errors after they occur but doesn't correlate cross-service request timelines or show which service hop added latency in a distributed trace.

Cloud Monitoring dashboards — create per-service CPU utilization graphs.

Why wrong: CPU graphs show resource utilization but don't correlate individual request flows across services or show which service contributed to a specific slow request.

Security Command Center — scan for misconfigurations causing performance issues.

Why wrong: SCC identifies security findings and misconfigurations, not application performance issues or distributed request tracing.

Deploy identical origin servers in every Google Cloud region globally.

Why wrong: Replicating full origin servers in every region is operationally expensive and complex to synchronize. CDN achieves the same performance benefit by caching at edge PoPs without full origin replication.

Use Cloud VPN to route viewer traffic through a direct tunnel to the origin servers.

Why wrong: VPN creates encrypted tunnels between networks — it doesn't cache content or improve CDN-style performance. VPN adds overhead rather than reducing latency.

Increase the origin servers' network bandwidth to handle more simultaneous viewer connections.

Why wrong: More bandwidth on the origin doesn't reduce geographic latency for distant viewers. CDN fundamentally solves the distance problem by moving content closer to users.

DevOps is a specific programming language designed for cloud applications.

Why wrong: DevOps is a philosophy and set of practices, not a programming language. It describes how development and operations teams collaborate and automate software delivery.

DevOps means developers take over all IT operations responsibilities, eliminating operations teams.

Why wrong: DevOps is about collaboration and shared responsibility, not eliminating operations. Ops expertise remains critical for reliability, capacity planning, and incident response.

DevOps is only applicable to software startups — traditional enterprises use ITIL for operations.

Why wrong: DevOps principles apply to enterprises of all sizes. ITIL and DevOps can coexist — many large enterprises successfully blend both frameworks.

BigQuery ML — train a custom language model on company policy documents.

Why wrong: BigQuery ML builds traditional ML models (regression, classification) from tabular data. It's not designed for large language model training or conversational AI from documents.

Cloud Natural Language API — it reads and summarizes documents automatically.

Why wrong: The Natural Language API analyzes text (sentiment, entity extraction, syntax). It doesn't build conversational AI assistants or perform RAG-based question answering.

Cloud Translation API — it translates policy documents into the user's language.

Why wrong: Cloud Translation converts text between languages. Building a Q&A assistant over documents requires LLM + retrieval infrastructure, not translation.

Economies of scale — lower per-unit costs as production volume increases.

Why wrong: Economies of scale describe cost reductions with volume, not how value increases for all users as more users join. Network effects are about value creation, not cost reduction.

Monopoly pricing power — larger platforms can charge higher prices.

Why wrong: Monopoly pricing describes a market power outcome, not the value creation mechanism. Network effects explain why platforms grow in value — pricing is a separate consideration.

Marginal cost reduction — digital goods can be replicated at near-zero marginal cost.

Why wrong: Marginal cost reduction applies to digital content reproduction. Network effects describe the value increase from additional users, not the cost of serving them.

Cloud Interconnect — it connects on-premises to Google Cloud.

Why wrong: Cloud Interconnect provides network connectivity between on-premises and GCP. It doesn't manage workload deployment or provide unified container management.

Google Distributed Cloud — runs Google Cloud services inside the customer's data center.

Why wrong: Google Distributed Cloud (GDC) brings Google Cloud hardware and services on-premises. It's an extreme option for air-gapped environments. Anthos is the broader hybrid management platform for most use cases.

Cloud Deployment Manager — deploys resources via infrastructure-as-code templates.

Why wrong: Cloud Deployment Manager is GCP's native IaC service (similar to CloudFormation). It doesn't manage cross-environment container workloads.

Microservices always cost less than monolithic applications to run.

Why wrong: Microservices can actually cost more due to service-to-service communication overhead, multiple deployment pipelines, and distributed system complexity. The benefit is agility and independent deployability, not necessarily cost.

Microservices eliminate the need for testing because each service is small enough to be bug-free.

Why wrong: Microservices still require testing — arguably more complex testing due to distributed system interactions. Service size doesn't prevent bugs.

Microservices allow applications to run on any hardware without modification.

Why wrong: Hardware portability is a benefit of containers, not specifically microservices architecture. Microservices are an architectural pattern for service decomposition, not hardware abstraction.

Zones are continents; regions are individual countries within a continent.

Why wrong: This is incorrect. Regions are metropolitan geographic areas (city/state level). Zones are isolated physical data center areas within a region.

Zones and regions are different terms for the same thing — Google uses them interchangeably.

Why wrong: Zones and regions have distinct, specific meanings in GCP. Confusing them leads to architectural mistakes in availability design.

A zone is a global resource; a region is a local data center.

Why wrong: This is backwards. Regions are geographic areas (local); zones are isolated failure domains within a region. Global resources (like global load balancers) exist independently of both.

`compute/instance/cpu/utilization` per instance group.

Why wrong: CPU utilization shows compute load per instance group but not request traffic distribution from the load balancer perspective.

`networking/vm_flow/egress_bytes_count` per VM.

Why wrong: Network flow egress bytes measure outbound data volume, not request count distribution by load balancer backend.

`logging/log_entry_count` filtered by region.

Why wrong: Log entry count measures logging volume, not load balancer traffic distribution.

Looker Studio (free BI dashboards)

Why wrong: Looker Studio is a free dashboarding tool suitable for simple reports. It lacks Looker's semantic layer (LookML), enterprise governance, embedded analytics API, and centralized metric definition.

BigQuery — it provides natural language querying via BQML.

Why wrong: BigQuery is a data warehouse for SQL analytics. BigQuery ML builds ML models. Natural language querying for business users and semantic governance are Looker's domain.

Vertex AI — it builds ML models that answer business questions.

Why wrong: Vertex AI is an ML platform for building predictive models. Enterprise BI with semantic layers, natural language querying, and embedded analytics is Looker's purpose.

E-government — providing digital access to government services online.

Why wrong: E-government is about digital citizen service delivery (online permits, payments). Smart city is about using IoT and data to optimize physical city operations and infrastructure.

Digital twin — creating virtual replicas of city infrastructure.

Why wrong: Digital twins are one component a smart city might use, but the broader concept of real-time IoT data + automated optimization is the smart city framework.

Edge computing — processing data locally at each sensor to reduce cloud bandwidth.

Why wrong: Edge computing is a technique that may be part of a smart city architecture, but the described scenario (centralized cloud analytics and optimization) is primarily a smart city use case.

Send monthly reminders to all teams via email to add labels to their resources.

Why wrong: Email reminders are not enforceable — teams may forget or deprioritize. Automated policy enforcement is the scalable solution across 5 teams and 20 projects.

Manually add labels to all existing and new resources through the Cloud Console.

Why wrong: Manual labeling at scale (20 projects, ongoing new resources) is error-prone, inconsistent, and doesn't scale. Automation through IaC and CI/CD is required.

Grant only project owners permission to create resources, and rely on them to enforce labeling.

Why wrong: Restricting resource creation to project owners limits development velocity. Consistent labeling requires automated policy enforcement, not gatekeeping by individuals.

BigQuery ML and Looker Studio, to analyze and visualize the extracted diagnosis codes

Why wrong: These tools analyze already-structured data. The challenge described is extracting structured information from unstructured documents — a data capture problem, not an analytics problem.

Vertex AI Pipelines and Cloud Dataflow, to orchestrate machine learning training jobs on document data

Why wrong: These are ML pipeline orchestration and data processing tools. While they could be used in a custom build, using Google's pre-built Document AI models is far more appropriate for a standard document extraction use case.

Cloud Translation API and Natural Language API, to translate and analyze the text content of medical records

Why wrong: Translation and natural language understanding are useful but don't address the primary challenge: extracting structured information from scanned documents with mixed handwritten and printed content. Document AI is the right starting point.

Multi-cloud requires purchasing separate hardware for each cloud provider's environment

Why wrong: Multi-cloud refers to using multiple public cloud providers' managed services — no hardware purchases are involved. This option describes a fundamental misunderstanding of cloud.

Cloud providers refuse to allow customers to use competing providers simultaneously

Why wrong: Cloud providers do not contractually prevent multi-cloud use. Many large enterprises use multiple providers, and providers compete for workloads rather than preventing multi-cloud strategies.

Multi-cloud makes it impossible to use any managed services because applications must be portable across providers

Why wrong: Multi-cloud does not require avoiding all managed services. Organizations often use managed services from each provider for workloads specific to that provider, accepting some lock-in for specific services while maintaining portability for others.

IAM permissions that restrict users from accessing competitor projects

Why wrong: IAM controls access within your own organization's projects. You have no ability to grant or deny IAM permissions in a competitor's Google Cloud project. IAM cannot prevent copying data to external projects.

Cloud DLP, by scanning and redacting sensitive data before it can be stored

Why wrong: Cloud DLP identifies and redacts sensitive data but doesn't prevent authorized users from copying data to external destinations. It addresses data content, not data movement paths.

Organization Policy constraints that prevent resource creation in competitor accounts

Why wrong: Organization policies apply within your own Google Cloud organization. You have no ability to define policies for competitor organizations. Organization Policy cannot control what happens in accounts you don't own.

Loosely coupled applications are always faster because messages are passed in memory rather than over the network

Why wrong: Loose coupling typically introduces network overhead (API calls, message queue latency) compared to in-process communication. The benefit is resilience and flexibility, not raw performance.

Loosely coupled architectures require less developer expertise and are easier to build than monolithic applications

Why wrong: Distributed, loosely coupled systems are generally more complex to design, test, and operate than monoliths. The benefits are resilience and scalability, not development simplicity.

Loose coupling reduces cloud costs because fewer network calls are made between services

Why wrong: Loose coupling through message queues and APIs often increases the number of network calls and associated data transfer costs. Cost reduction is not the primary motivation.

Advantage: private cloud is always cheaper than public cloud. Trade-off: private cloud provides less storage capacity

Why wrong: Private cloud is rarely cheaper than public cloud at equivalent scale. The opposite is usually true — without global scale economics, private cloud costs more per unit. Storage capacity can match requirements with sufficient investment.

Advantage: private cloud provides automatic scaling to unlimited capacity. Trade-off: private cloud requires purchasing hardware every time capacity is needed

Why wrong: Unlimited automatic scaling is a public cloud characteristic. Private cloud has a fixed capacity ceiling determined by the infrastructure installed. This option describes the trade-off as an advantage and vice versa.

Advantage: private cloud services are managed by the cloud provider, reducing operational burden. Trade-off: customers cannot customize private cloud configurations

Why wrong: This describes a managed public or hosted cloud service, not private cloud. Private cloud's defining characteristic is that the organization owns and operates its own infrastructure — the opposite of provider-managed.