Managing and provisioning infrastructure

Network topologies

Hybrid connectivity

Facts that get tested:

• Cloud Router speaks BGP and advertises subnets and learns on-premises prefixes dynamically. Without it you rely on static routes.
• Global dynamic routing makes a region’s Cloud Router routes visible to all regions; regional mode confines them.
• Interconnect availability: two connections in one metro gives 99.9%, four connections across two metros gives 99.99%.
• Interconnect does not encrypt. If the requirement says “encrypted in transit over the dedicated link”, add MACsec (Dedicated) or run HA VPN over Cloud Interconnect.
• Reaching Google APIs privately from on premises: Private Service Connect for Google APIs or private.googleapis.com / restricted.googleapis.com virtual IPs advertised through the hybrid link. restricted.googleapis.com supports only services covered by VPC Service Controls.
• Network Connectivity Center connects many sites (VPN, Interconnect, SD-WAN appliances) and VPCs as spokes of one hub.

Multicloud and Google Cloud to Google Cloud

• Between two Google Cloud VPCs: VPC peering (private, not transitive), Shared VPC (one network, many projects), Network Connectivity Center VPC spokes (scales to many VPCs), Private Service Connect (consume a published service across organizations and overlapping ranges), or HA VPN between VPCs.
• Overlapping IP ranges rule out peering. PSC or NAT solves it.
• To another cloud: Cross-Cloud Interconnect for volume, HA VPN for speed to deploy, with consistent identity (Workload Identity Federation) and a single control plane for Kubernetes (GKE Enterprise, with fleet management).

Network security

• VPC firewall rules are stateful, allow or deny, with priorities (0 highest, 65535 lowest), implied rules (deny ingress, allow egress). Target by network tag or, better, by service account, which cannot be set by anyone with only instance edit access.
• Network firewall policies can be attached at organization, folder or VPC level (hierarchical) and support FQDN, geo-location and threat intelligence rules. Cloud NGFW is the managed next generation firewall; its Enterprise tier adds intrusion prevention with TLS inspection.
• Cloud IDS detects intrusions by mirroring traffic (detection only, powered by Palo Alto technology). Packet Mirroring sends copies of traffic to your own collector or IDS appliance.
• Cloud Armor is the WAF and DDoS defence for external Application Load Balancers (and network load balancers with Cloud Armor Network edge policy). Features: preconfigured OWASP rules, IP and geography allow or deny, rate limiting, adaptive protection, bot management with reCAPTCHA. It applies to the backend service, so it sits at the edge, not on VMs.
• Identity-Aware Proxy controls access to web apps and SSH/RDP without a VPN. VPC Service Controls protect Google APIs (covered in the security domain).
• Cloud NAT gives outbound internet to private VMs and GKE nodes. There is no inbound path through it.
• Private Google Access enables VMs without external IPs to reach Google APIs. It is configured per subnet.

VPC design and load balancing

• Use custom mode VPCs with planned, non-overlapping ranges (including secondary ranges for GKE pods and services). Subnets can be expanded but not shrunk, and ranges cannot be moved, so plan growth.
• Separate environments (production, non-production) in different VPCs or projects. Centralise with Shared VPC; give each team a service project.
• Network service tiers: Premium carries traffic on Google’s backbone (global load balancers need it); Standard uses the public internet for the long haul and is cheaper, regional only.
• Load balancer settings to know: health checks (their probe ranges must be allowed through the firewall), backend services with balancing mode (utilisation, rate, connections), session affinity, connection draining, URL maps for path or host routing, serverless NEGs for Cloud Run or Cloud Run functions behind a load balancer, backend buckets for static content with Cloud CDN, SSL policies and certificate management (Google-managed certificates, Certificate Manager), HTTP to HTTPS redirect, and backend timeouts.
• Access to “cloud adjacent services” (Cloud SQL, Memorystore, Filestore, Vertex) is via private services access (a peered producer VPC with an allocated range) or Private Service Connect endpoints.

Storage systems

Allocation and sizing

• Disks: Persistent Disk is zonal or regional (synchronous replication across two zones). Hyperdisk lets you provision IOPS and throughput independently of size, and is the newer default for high performance. Local SSD is very fast but ephemeral: lost when the VM stops or is deleted, not for the only copy of data. Disk size limits IOPS for Persistent Disk, so a larger disk can be the cheapest way to get more performance.
• File storage: Filestore tiers (Basic, Zonal, Regional, Enterprise) for NFS with different scale and availability. Parallelstore and Managed Lustre feed GPUs and HPC.
• Object storage: pick location type and class. Move compute next to the data (same region) to avoid egress and latency.
• Databases: size by workload. Cloud SQL storage can auto-increase; Spanner scales by processing units or nodes (autoscaler available); Bigtable scales by nodes (autoscaling per cluster) and needs a good row key to avoid hotspots; BigQuery is serverless, priced on-demand per bytes scanned or by slots reservations (editions: Standard, Enterprise, Enterprise Plus).

Processing and compute near the data

• Run analytics where the data lives (BigQuery, Dataflow in the same region). Cross-region reads cost egress and add latency.
• For GPU training that reads big datasets, use Cloud Storage with Anywhere Cache, Parallelstore, or Hyperdisk ML so accelerators are not starved.

Security and access

• Cloud Storage: IAM at project or bucket level, uniform bucket-level access (recommended, turns off ACLs), public access prevention, signed URLs for time-limited access, CMEK for encryption.
• Databases: private IP only, IAM database authentication, Cloud SQL Auth Proxy or connectors, no public IP when avoidable, SSL enforced.
• BigQuery: dataset and table IAM, column-level and row-level security, policy tags, authorised views, data masking.

Transfer and latency

• Online transfer: Storage Transfer Service, gcloud storage with parallel composite uploads. Offline: Transfer Appliance. Hybrid ingestion: Interconnect.
• Cloud Storage buckets in multi-region or dual-region lower latency for geographically spread readers; Cloud CDN caches reads at the edge.
• Choose regions close to users, and keep chatty services in the same region.

Retention and lifecycle

• Object Lifecycle Management rules: change class after N days, delete after N days, delete old noncurrent versions. Soft delete keeps deleted objects for a window (default 7 days). Retention policies and Bucket Lock enforce a minimum retention (locking is irreversible) for WORM compliance; object holds for legal hold.
• BigQuery table and partition expiration; long-term storage pricing applies after 90 days without modification (automatic).
• Logs: log buckets with set retention (default 30 days for _Default, 400 days for _Required), sinks to Cloud Storage or BigQuery for longer retention.

Growth and protection

• Plan growth with quotas, autoscaling storage (Cloud SQL, Spanner autoscaler, Bigtable autoscaling) and partitioned tables. Review capacity through monitoring metrics and quota alerts.
• Protection: snapshots, backups with retention schedules, cross-region copies, point-in-time recovery, Backup and DR Service, and regular restore tests. Cloud SQL HA protects against zone failure; backups protect against corruption. They are not the same thing.

Compute systems

Compute Engine provisioning

• Instance templates define a VM shape; managed instance groups (MIGs) create, heal, scale and update identical VMs from a template. Regional MIGs spread across zones. Stateful MIGs preserve disks and names. Rolling updates with surge and max-unavailable, and canary updates with two templates.
• Images and image families for golden images; Packer or Image Builder pipelines create them. OS Config / VM Manager handles patch management (patch jobs and deployments, inventory, OS policies) across fleets.
• Spot vs standard: Spot VMs are preemptible with no max duration, for stateless, batch and fault-tolerant work. Standard VMs for anything that cannot be interrupted. Mix them: a MIG with a standard baseline and Spot for burst, or GKE node pools with Spot. Use a shutdown script (30 seconds) to checkpoint. Flex-start and reservations offer assured capacity for scarce GPUs.
• Reservations guarantee capacity in a zone; committed use discounts lower price; they are different things.
• Sole-tenant nodes for compliance or BYOL licensing; Confidential VMs for encrypted memory; Shielded VMs (secure boot, vTPM, integrity monitoring) on by default for supported images.
• Metadata and startup scripts; service accounts attached to VMs (use a dedicated least-privilege one, never the default Compute Engine account with Editor).

Networking for compute

• Compute Engine: internal IPs, optional external IPs (avoid), alias IP ranges, multiple NICs, Cloud NAT, Private Google Access, OS Login for IAM-based SSH, IAP TCP forwarding for administration without public IPs.
• GKE: VPC-native clusters, private clusters, authorized networks, Dataplane V2 network policies, Workload Identity Federation for GKE to map Kubernetes service accounts to Google service accounts (never node keys), Gateway API.
• Serverless: Serverless VPC Access connectors or Direct VPC egress (Cloud Run) to reach private resources; ingress settings (internal, internal and load balancer, all); put Cloud Run behind a load balancer with Cloud Armor; Private Service Connect to reach producers.
• Google Cloud VMware Engine: dedicated VMware clusters (vSphere, vSAN, NSX) in Google Cloud, connected by private service access; run VMs unchanged for fast data centre exit, integrate with VPC through the VMware Engine network.

Infrastructure orchestration and configuration

• Terraform (with Google’s provider and the Cloud Foundation Toolkit modules) is the standard IaC choice; state in a Cloud Storage backend with locking. Infrastructure Manager is Google’s managed Terraform service. Deployment Manager is deprecated. Config Connector and Config Sync manage resources from Kubernetes manifests (GitOps).
• Service Catalog (private catalogue of approved solutions) and Terraform modules give teams safe, pre-approved building blocks.
• Patch management: VM Manager patch jobs, rolling MIG updates with new images (replace rather than patch for immutable infrastructure), GKE release channels and node auto-upgrade, maintenance windows and exclusions.

Container orchestration

GKE essentials: regional clusters for HA; release channels (Rapid, Regular, Stable); cluster, node auto-provisioning and horizontal pod autoscaling; multi-cluster with fleets; Binary Authorization for signed images; Workload Identity; Cloud Service Mesh. Use GKE Enterprise for consistent policy, config and mesh across clusters on Google Cloud, on premises and other clouds.

Serverless

• Cloud Run: containers, per-request concurrency (default 80, up to 1000), min and max instances, CPU always allocated or only during requests, revisions with traffic splitting for canary and rollback, jobs for batch, Cloud Run functions for source-based functions. Scale to zero, pay per use.
• Cloud Run functions: event-driven (Eventarc, Pub/Sub, Cloud Storage), pick when logic is small and single-purpose.
• App Engine standard and flexible: still present, but new builds usually pick Cloud Run.
• Serverless suits spiky, event-driven, low-operations work; use GKE or Compute Engine when you need long-lived connections beyond limits, custom kernel, special hardware or strict portability.

ML workflows on the Agent Platform

• Agent Platform Pipelines (Vertex AI Pipelines) run Kubeflow Pipelines or TFX definitions as serverless, repeatable ML workflows: data prep, training, evaluation, registration, deployment. Pair with Model Registry, Feature Store (now in BigQuery), Experiments and Model Monitoring for drift and skew. CI/CD triggers retraining.
• Data integration: BigQuery, Cloud Storage and Dataflow feed training. Managed datasets, BigQuery ML for in-warehouse models, and Dataplex for governance.
• Training: custom training jobs in containers, hyperparameter tuning, distributed training, and managed Ray. Serve with online prediction endpoints (autoscaling, traffic split between model versions) or batch prediction.
• AI Hypercomputer: pick GPUs (NVIDIA A3 and newer; flexible for many frameworks, best for most training and fine-tuning and for broad model support) or TPUs (Google’s accelerators, strong cost-performance for large transformer training and inference with JAX, TensorFlow or PyTorch/XLA). Orchestrate on GKE or Slurm (Cluster Director). High-bandwidth networking and parallel file systems keep accelerators busy.
• Consumption models: on-demand, committed use discounts, Spot for tolerant training (with checkpointing), reservations for guaranteed capacity, and Dynamic Workload Scheduler (flex-start and calendar mode) to queue for scarce GPUs and TPUs. Large-scale training needs checkpointing, topology-aware scheduling and fast storage.
• Use Cloud Run functions to trigger inference or glue steps from events.

Prebuilt AI solutions and APIs

• Gemini Enterprise is the workplace product: AI agents, search across company data, and NotebookLM for source-grounded research and summaries. Use it to boost workflows for employees without building a platform.
• Model Garden lets you discover, test and deploy foundation models (Gemini, Imagen, partner models such as Claude, and open models such as Gemma and Llama), either as managed APIs (pay per token) or self-deployed endpoints. Prefer a managed API for speed and low operations, self-deployed for strict control or custom tuned open models.
• Rule: prebuilt API first, then a Gemini prompt with grounding, then tuning, and a custom model last.