PCA question patterns
Find a pattern
Each row is wording that keeps showing up in questions, with the answer it usually points to in green. Tap a row to see why.
In this list
No pattern matches that. Try a service name, such as CodeDeploy or Config.
Wording that decides the answer
-
Managed and serverless services: Cloud Run, GKE Autopilot, Cloud SQL, Spanner, BigQuery, Dataflow.
Self-managed VMs and clusters add patching, scaling and failure handling the team must run.
-
Predefined roles granted to groups, service accounts without keys, private connectivity, and organization policy guardrails.
Basic roles, shared keys and public endpoints widen the blast radius.
-
The cheapest DR pattern that meets the targets: backup and restore, pilot light, warm standby, or hot.
Active-active everywhere is correct for near-zero targets only; it is waste for hours-level targets.
-
Hybrid connectivity (Interconnect or HA VPN) and integrate; do not force migration.
The scenario fixes the constraint; answers that move the system break it (EHR insurer interfaces).
Databases and storage
-
Spanner (multi-region for 99.999%).
Cloud SQL and AlloyDB scale writes vertically within a region.
-
Cloud SQL, with HA for zone failure and Database Migration Service for the move.
Compatible engines, least change.
-
Bigtable with a row key that avoids hotspots.
Wide-column store built for scale; not relational.
-
BigQuery, with partitioning and clustering to cut scanned bytes.
Warehouse, not a transactional database.
-
Firestore.
Document model with client SDKs and offline sync.
-
Cloud Storage with Autoclass, plus lifecycle rules for deletion.
Autoclass moves objects between classes by actual access; no minimum-duration guesswork.
-
Retention policy with Bucket Lock (and object holds if needed).
Locking makes the retention period irreversible, meeting WORM rules.
-
Filestore.
Managed NFS; Cloud Storage is object storage.
Compute and containers
-
Cloud Run.
Per-request billing, no cluster operations.
-
GKE Autopilot (Standard if special hardware or node customisation is needed).
Autopilot manages nodes and bills per pod.
-
Spot VMs, with checkpointing and a shutdown script.
60 to 91% cheaper but reclaimable with 30 seconds notice.
-
Committed use discounts, flexible if shapes will change.
Discount for commitment; Spot is wrong for steady critical load.
-
Google Cloud VMware Engine.
Fastest data centre exit with unchanged VMs and tools.
-
GKE Enterprise fleets with Config Sync and Cloud Service Mesh.
One control plane for policy, config and mesh across clusters.
-
Managed instance groups with new templates (rolling update) and VM Manager patch deployments.
Replace instances from a golden image; automate patches for the rest.
Networking
-
Global external Application Load Balancer with Cloud CDN and Cloud Armor.
One anycast IP, edge caching and protection.
-
HA VPN with two tunnels and Cloud Router (BGP).
Deployable in hours; 99.99% SLA with the redundant setup.
-
Dedicated Interconnect (or Partner Interconnect if you cannot reach a Google facility); redundant circuits in two metros for 99.99%.
VPN over the internet cannot guarantee throughput or latency.
-
MACsec, or HA VPN over Cloud Interconnect.
Interconnect is private but not encrypted by default.
-
Network Connectivity Center for hub and spoke; Private Service Connect for service publishing across overlapping ranges.
VPC peering is not transitive and cannot have overlapping ranges.
-
Shared VPC with a host project and service projects.
Central control of subnets and firewalls with delegated app teams.
-
Cloud NAT, plus Private Google Access for Google APIs.
NAT is outbound only.
Security and compliance
-
VPC Service Controls perimeter with access levels, tested in dry run.
IAM alone allows the copy; the perimeter blocks cross-boundary API calls.
-
IAP TCP forwarding with OS Login.
IAM-based access, audited, no exposed ports.
-
Workload Identity Federation.
Exchange the external token for short-lived Google credentials.
-
Service account impersonation (Token Creator), or Privileged Access Manager for just-in-time roles.
No shared keys, short-lived tokens, full audit trail.
-
CMEK in Cloud KMS; HSM if hardware-backed; EKM if keys must stay outside Google.
Default encryption is Google-managed and cannot be revoked by the customer.
-
Sensitive Data Protection: inspect, then de-identify (mask, tokenise).
Built-in infotypes and transformations, works on Cloud Storage, BigQuery and streams.
-
Organization policy constraint gcp.resourceLocations at the organization or folder.
Inherited guardrail; includes new projects.
-
Binary Authorization with attestations from Cloud Build, plus Artifact Registry scanning.
Deploy-time enforcement of the supply chain.
-
Aggregated log sink at the organization to a dedicated project's bucket or Cloud Storage with a locked retention policy; enable Data Access logs where needed.
Per-project sinks miss new projects; Data Access logs are off by default.
-
Model Armor in front of the model, with Sensitive Data Protection for leakage.
Purpose-built screening for prompts and responses on any model.
-
Sign the BAA for HIPAA, use covered services, restrict locations, VPC Service Controls, audit logs, minimise data and consent flows for COPPA.
Law dictates the controls; a Google certificate does not make your app compliant.
Migration and data movement
-
Transfer Appliance.
Shipping beats months of transfer time.
-
Storage Transfer Service (with agents on premises).
Managed, scheduled, checksummed.
-
Database Migration Service with continuous replication, then cut over.
Replication keeps source live until cutover.
-
Datastream.
Serverless change data capture.
-
Migration Center.
Assessment and TCO before waves are planned.
-
Migrate to Virtual Machines (test clones, then cut over).
Replicates while the source runs.
AI and data platform
-
A prebuilt API (Vision, Speech, Translation, Document AI) before a custom model.
Fastest and cheapest; custom training last.
-
Agent Builder with grounding on your data (RAG), with Gemini.
Reduces hallucinations and needs no model training.
-
Human-in-the-loop review step with approve, reject, edit before the catalogue update.
Risk control for generative output; named in the Cymbal case.
-
Pub/Sub into Dataflow, output to BigQuery.
Managed, exactly-once streaming; Dataproc is for existing Spark.
-
TPUs or GPUs on AI Hypercomputer; Spot with checkpoints or flex-start for scarce capacity.
Matches accelerator and consumption model to the job.
Process, delivery and operations
-
Canary rollout (Cloud Run traffic split or Cloud Deploy) watched by SLO metrics.
Small exposure first, rollback on signal.
-
Apigee.
API products, developer portal, quotas, monetisation; API Gateway is simpler and lacks this.
-
Alert on user-facing symptoms and SLO burn rate, route by severity, add runbooks.
Cause-based, noisy alerts breed alert fatigue.
-
Use the error budget: when spent, freeze risky releases and fix reliability.
SRE trade-off mechanism agreed with product owners.
-
Mitigate first (roll back, fail over), communicate, then root cause; blameless postmortem with owned action items.
Restore service before investigating.
-
Terraform blueprints in a service catalog, project factory, guardrails by organization policy.
Consistent and fast without manual review of each request.
-
Chaos experiments with a hypothesis and small blast radius, DR drills, and load tests above peak.
Reliability is demonstrated, not assumed.
-
Labels and billing export, Recommender right-sizing, delete idle resources, autoscale, lifecycle rules, Spot and committed use.
Remove waste before negotiating price.
-
Executive sponsor, pilot group, training before go-live, communication plan, phased rollout.
People causes most migration failures; big-bang cutovers are wrong.
Now spot them in real questions. Practice sets use the same wording.
Practise 20 questions