Google ACE Cheat Sheet
Quick-reference summary of the key facts for the ACE exam.
Exam at a glance
| Item | Detail |
|---|---|
| Fee | $125 USD (+ tax where applicable); renewal exam $75 |
| Length | 2 hours |
| Questions | 50–60 multiple choice and multiple select (Google publishes a range, not an exact count) |
| Passing score | Not published — pass/fail only |
| Validity | 3 years |
| Retakes | Max 4 attempts in 2 years; waits of 14 / 60 / 365 days after 1st / 2nd / 3rd failure; full fee each time |
| Delivery | Online proctored or Pearson VUE test center; English, Japanese, Spanish, Portuguese |
Section weights
| Section | Weight |
|---|---|
| Setting up a cloud solution environment | ~20% |
| Planning and implementing a cloud solution | ~30% |
| Ensuring the successful operation of a cloud solution | ~30% |
| Configuring access and security | ~20% |
Resource hierarchy
Organization → Folder → Project → Resource. Permissions and policies inherit downward. One organization per account. Project IDs are globally unique and permanent. Folders nest up to 10 levels. Organization policies restrict what is allowed (even for admins); IAM controls who can act.
Billing
One billing account pays for many projects; a project links to exactly one billing account. Budgets only alert — they never stop spending. Billing export to BigQuery enables SQL cost analysis (no backfill). Labels enable per-team cost attribution.
Compute choices
| Workload | Service |
|---|---|
| Full OS control, legacy apps, GPUs | Compute Engine |
| Container orchestration, microservices | GKE (Autopilot default; Standard for node control) |
| Containerized web service/API, scale to zero | Cloud Run |
| Short event-driven task | Cloud Functions |
| Fault-tolerant batch, lowest cost | Spot VMs (60–91% off, 30-second reclaim warning) |
| Unusual vCPU/memory ratio | Custom machine types |
Storage choices
| Need | Service / class |
|---|---|
| Hot object data | Cloud Storage Standard |
| Monthly / quarterly / yearly access | Nearline / Coldline / Archive (30 / 90 / 365-day minimums) |
| VM disks, survive zone failure | Regional Persistent Disk |
| Standard relational DB | Cloud SQL |
| Demanding PostgreSQL | AlloyDB |
| Global relational, strong consistency | Spanner |
| Document DB, mobile/web | Firestore |
| Massive-scale time-series / high throughput | Bigtable |
| Analytics over TBs with SQL | BigQuery (not a transactional DB) |
Bucket location and BigQuery dataset location cannot be changed after creation.
Networking
VPCs are global; subnets are regional. Custom mode VPC: you create subnets. Subnet ranges can expand, never shrink. Firewall rules are stateful; lowest priority number wins; implied deny ingress, implied allow egress. Target rules by service account (robust) over network tags. Global HTTP(S) → global external Application Load Balancer. Non-HTTP → Network Load Balancer. Internal-only → internal load balancer. Global load balancing requires Premium Tier. HA VPN: two tunnels, 99.99% SLA. VPC peering is non-transitive. Shared VPC: one host project owns the network, service projects use it.
Compute Engine operations
Instance templates are immutable; changes need a new template + rolling update. OS Login ties SSH to IAM (Compute OS Login / Compute OS Admin Login roles). Live migration is the default on host maintenance (not for GPU/Spot VMs). VM Manager = patch management + OS inventory. Snapshots back up disks (incremental, schedulable); images boot new VMs.
GKE
gcloud container clusters get-credentials before kubectl works. Regional cluster = control plane across 3 zones. Private cluster = nodes without external IPs. HPA scales replica count; VPA resizes CPU/memory requests. Images live in Artifact Registry; nodes need Artifact Registry Reader. Service type LoadBalancer = public IP; ClusterIP = internal only.
Cloud Run & Cloud Functions
Cloud Run services are private by default (grant run.invoker to allUsers for public). Revisions are immutable; traffic splitting enables canary + instant rollback. Minimum instances avoid cold starts; maximum instances cap cost/connections. Functions: short, stateless, event-triggered; design for idempotency (Pub/Sub delivers at least once). Eventarc routes multi-source events to Cloud Run.
Data
Cloud SQL regional = HA with automatic failover. Firestore mode (Native/Datastore) is permanent. Pub/Sub: push delivers to HTTPS; pull requires requests. Dataflow = batch + streaming ETL. Big transfers / recurring / from another cloud → Storage Transfer Service. Load pattern: upload to a bucket, then load from the bucket.
Monitoring & logging
Alerting policy = condition + notification channel. Alert on logs via a log-based metric first. Ops Agent collects VM logs + metrics. Sinks route logs to Storage / BigQuery / Pub/Sub / another project. Log buckets hold logs with configurable retention (default 30 days). Admin Activity audit logs are always on; Data Access logs must be enabled. Check the Status Dashboard before troubleshooting a multi-service outage.
IAM & service accounts
Least privilege always. Basic roles (Owner/Editor/Viewer) are almost always too broad — prefer predefined roles; custom roles when no predefined role fits. Grant at the highest level matching the scope; prefer groups over users. Service accounts are for applications. Service Account User role is needed to attach an account to a resource. Prefer impersonation and short-lived credentials over keys. External workloads → Workload Identity Federation. IAM conditions enable time-limited access.
Commands to know
gcloud services enable [API] · gcloud compute instances create · gcloud compute ssh [VM] · gcloud compute instances list · gcloud container clusters create · gcloud container clusters get-credentials [CLUSTER] · kubectl apply -f · kubectl get pods/nodes/services · gcloud run deploy · gcloud functions deploy --trigger-topic · gcloud storage cp · gcloud projects get-iam-policy · gcloud iam service-accounts create