Prepare and test your skills
Prepare and test your skills
Hybrid networking (opens in a new tab) connects an on-premises network to a Virtual Private Cloud (VPC) so traffic can move securely between the two environments. A VPC is a global private network with regional subnets, and hybrid links extend it with encrypted tunnels over the internet or with private circuits. Dynamic routing exchanges routes automatically so new subnets propagate without manual static routes.
Hybrid links to on-premises use Cloud VPN for encrypted tunnels over the internet and Cloud Interconnect for private circuits that do not traverse the internet. Traffic flows from the on-premises network through the encrypted tunnel or private circuit to the VPC network, and return traffic flows back through the same path. High Availability (HA) VPN is the production form of Cloud VPN with a 99.99% service level agreement (SLA), where availability is handled by Google infrastructure instead of a customer-managed virtual machine (VM) endpoint.
Cloud Router manages Border Gateway Protocol (BGP) routing between the VPC network and the on-premises router. Cloud Router advertises VPC routes outward and learns on-premises routes inward, while the tunnels provide encrypted transport. If the BGP session is not in the UP state, routes are not exchanged and traffic cannot be forwarded. An intermediary VM as a VPN endpoint creates a single point of failure because one VM failure can block all communication between environments.
Redundant production designs require four Cloud Interconnect connections across two Google Cloud regions. Connection pairs must sit in different metropolitan areas and in separate edge availability domains, with Cloud Router advertising routes to the on-premises router. If one metro, edge domain, or region is impaired, traffic continues over the remaining paths to the other region.
| Option | Network path | Capacity and setup |
|---|---|---|
| Cloud VPN | Private secure path over internet with IPsec | Lower bandwidth; choose for low cost or experimental use |
| Dedicated Interconnect | Direct connection to Google, no internet or provider | 10, 100, or 400-Gbps circuits up to 8 per connection; VLANs 50 Mbps to 50 Gbps; requires colocation equipment |
| Partner Interconnect | Dedicated path through supported provider, no internet | VLANs 50 Mbps to 50 Gbps; no colocation equipment |
Exam tip: A single Dedicated Interconnect connection and a single Partner Interconnect VLAN attachment both carry No Uptime SLA; 99.99% requires four connections across two regions with different metros and separate edge availability domains.
Dynamic routing with Cloud Router and HA VPN gateways links VPC networks to on-premises networks, other clouds, and to each other. Cloud Router is a regional service that belongs to one VPC network in one region and uses one Google autonomous system number (ASN). The Google ASN must be a private ASN in the range 64512 through 65534 or 4200000000 through 4294967294 and cannot be changed after creation. A BGP session runs over each VPN tunnel through a Cloud Router interface bound to that tunnel.
An HA VPN gateway is a regional gateway with two interfaces, each with its own automatically assigned external IP address. A tunnel connects one HA VPN interface to one peer interface, with a BGP session on top of each tunnel. Building the connection follows an ordered dependency:
A left-to-right process flow showing the five ordered steps to build an HA VPN connection, from VPC and gateway creation through Cloud Router, peer representation, tunnel pair, and BGP configuration.
0 to peer interface 0 and a second from interface 1 to interface 1, using matching pre-shared keys and IKEv2.By default both BGP sessions advertise unmodified base priorities, creating an active-active configuration where traffic flows over both tunnels. An active-passive configuration uses unequal priorities, for example priority 10 on tunnel 1 and priority 20 on tunnel 2, so traffic prefers tunnel 1 and uses tunnel 2 only after failure. Cross-Cloud Interconnect is the dedicated Google-provisioned link between a Google Cloud VPC and another provider, delivering up to 100 Gbps per link with paired primary and redundant connections.
Private communication between separate VPC networks keeps isolation while allowing controlled reachability. A VPC network is an isolated private partition, so workloads in different VPC networks cannot communicate by default even though subnets in different regions of the same VPC communicate privately over Google's backbone. Three private methods share routes, subnets, or individual services without using the internet.
| Method | Connectivity model | Isolation boundary |
|---|---|---|
| VPC Network Peering | Two VPC networks exchange routes for VM-to-VM reachability | Each VPC keeps its own firewall and Identity and Access Management (IAM); not transitive |
| Shared VPC | Service projects share subnets in one host-project VPC | Central network control with delegated workload deployment in one organization |
| Private Service Connect | Consumer VPC connects privately to a producer service | Only the published service is exposed; networks not merged |
VPC Network Peering merges route propagation so forwarding behaves as if VMs were in the same VPC, but it requires non-overlapping IPs and is not transitive. Spokes peered to the same hub cannot reach each other through the hub alone. Shared VPC lets central administrators control subnets, routes, and firewalls in the host project while service-project administrators deploy VMs into designated subnets. Private Service Connect (PSC) uses a producer-to-consumer model where the producer publishes a service and the consumer creates an endpoint in its own VPC that forwards to the producer.
Security protection in Google Cloud filters traffic from the internet edge toward the VPC through successive controls. Traffic moves from the internet to the Google edge and then to protected backends, with enforcement inside points of presence where external load balancing occurs. VPC firewall rules are project-level controls bound to a single VPC, while hierarchical firewall policies attach at the organization or folder level and apply to many projects below.
A VPC rule targets workloads by IP ranges, network tags, or service accounts, with direction, priority, and matching deciding when the rule applies. Lower priority numbers are evaluated first. Firewall logging records connection metadata for auditing without changing the allow or deny decision. Google Cloud Armor is an edge security service that filters traffic before it reaches workloads, using ordered policies with match conditions on IP, geography, headers, cookies, URL path, and HTTP method.
Layer 3 and Layer 4 volumetric protection is always on for protected load balancers with automatic mitigation by Google. Layer 7 application protection requires a user-configured security policy attached to a backend service, with preconfigured Web Application Firewall (WAF) rules for SQL injection, cross-site scripting, and command injection. Identity-Aware Proxy (IAP) checks Google identity and IAM permission before traffic reaches applications, and IAP TCP forwarding reaches a VM with no public IP. Firewall policy still allows TCP port 22 from 35.235.240.0/20, which contains all IAP TCP-forwarding addresses.
Cloud Load Balancing distributes traffic across backends such as instance groups, network endpoint groups (NEGs), and on-premises endpoints. A global balancer with Premium Tier advertises a global anycast IP from all points of presence worldwide, with one backend service able to use backends in multiple regions. A regional balancer uses a regional external or internal IP, and all components must stay in that region. Google Front Ends send global traffic to healthy backends closest to the user, with overflow going to another region with capacity.
Private Google Access is a per-subnet setting that lets internal-only VMs reach supported Google APIs over the VPC internal path. Cloud NAT is a managed egress service that translates a private source address to a shared public address for internet destinations without assigning an external IP to each VM. PSC exposes Google APIs and cross-VPC services as internal endpoints, where a PSC endpoint serves a bundle of global APIs and a PSC backend serves a single API. Network Service Tiers controls the internet path, where Premium Tier routes over Google's global fiber entering at the point of presence closest to the user, while Standard Tier routes over transit providers entering near the resource region.
Continue with this study guide: Configuring network topologies
Data storage allocation (opens in a new tab) is the combined choice of where data lives, how fast it performs, and how it survives failure. Cloud Storage is object storage for files and backups, Persistent Disk is durable network block storage for VMs, and databases add replication and backup settings. Allocation also chooses bucket location for availability and storage class for cost as access cools.
Zonal and regional block storage allocation is the choice of where disk data lives and how it survives a zone outage. Persistent Disk stays independent from a VM, so it can be detached, reattached, and kept after the VM is deleted. A zonal disk stores data in one zone with built-in redundancy against equipment failure, while Regional Persistent Disk replicates every write synchronously across two zones in the same region. If the primary zone fails, a backup VM in the surviving zone can force-attach the regional disk and continue with near-zero recovery point objective (RPO).
Google Cloud Hyperdisk provides higher performance than Persistent Disk, and Hyperdisk Balanced lets input/output operations per second (IOPS) and throughput be provisioned separately and changed dynamically. Hyperdisk Balanced High Availability adds synchronous replication between two zones so the volume remains available during a single-zone outage. Hyperdisk Storage Pools share pre-provisioned capacity, throughput, and IOPS across many volumes and use thin provisioning and deduplication to lower cost. The pool tries to add capacity automatically to keep utilization at or below 80% of provisioned capacity.
Local SSDs are physically attached drives that give high throughput to a single VM, but data lasts only until the instance is stopped or deleted. Choose Local SSDs for temporary caches and scratch data that can be rebuilt, and choose Persistent Disk or Hyperdisk for data that must survive VM deletion. Snapshots capture point-in-time copies of Persistent Disk volumes, are incremental by default, can be taken from running instances, and can be copied across regions for disaster recovery.
Exam tip: Regional Persistent Disk protects against a zone outage by synchronous replication between two zones, while snapshots protect against data loss and region outage by storing a copy that can be copied across regions.
Cloud Storage allocation combines bucket location for availability with storage class for cost. Regional buckets replicate data synchronously across multiple zones in one region, so they survive a zone outage but not a region outage. Dual-region and multi-region buckets store data in at least two geographic locations, where metadata is written synchronously and object data is replicated asynchronously. Turbo replication on dual-region buckets tightens that copy to a 15-minute RPO.
Storage class and lifecycle rules control cost as access cools:
Object Lifecycle Management (OLM) configuration lives on the bucket, the service evaluates each object, and only after all conditions are met does it apply the transition or delete action. Object Versioning preserves an overwritten live object as a noncurrent version, so lifecycle rules must handle both live and noncurrent states. Locked retention policies enforce compliance where Bucket Lock applies one minimum duration to every object and Object Retention Lock sets an individual retain-until time that cannot be shortened once locked.
Data migration tool selection chooses how to move data into Cloud Storage without changing it. Storage Transfer Service handles online transfers over the network, the gcloud storage command handles small online uploads, and Transfer Appliance handles offline moves with shipped hardware. Online transfer flows from the source over Cloud Interconnect or the internet to a bucket, while offline transfer flows from the data center to the appliance and then after shipping from the appliance to Cloud Storage.
| Transfer scenario | Size or bandwidth condition | Recommendation |
|---|---|---|
| Another cloud to Cloud Storage | Any size | Storage Transfer Service |
| Bucket to bucket | More than 1 TB | Storage Transfer Service |
| On-premises to Cloud Storage | Less than 1 TB | gcloud storage command |
| On-premises to Cloud Storage | More than 1 TB without enough bandwidth | Transfer Appliance |
Jobs with billions of small files are typically QPS-bound by per-object request rate, while jobs with large objects are bandwidth-bound by network throughput. Storage Transfer Service can complete around 1000 tasks per transfer job per second when bandwidth allows, and splitting one large job into multiple smaller jobs with include and exclude prefixes can exceed the single-job limit. Total latency sums last-mile, Google frontend and edge point of presence, Compute Engine region latency, and app processing time. Co-locating buckets and VMs in the same region keeps reads on the internal network, and adding global load balancing with Cloud CDN serves repeats from the edge.
Disaster recovery planning designs storage to resume within a Recovery Time Objective (RTO) and lose no more than a Recovery Point Objective (RPO). RTO is maximum acceptable time to restore service, while RPO is maximum acceptable data loss in time. Live replication across zones or regions keeps a current second copy for fast failover with low RPO, while backups protect against corruption and deletion but define longer RTO for restore.
Cloud Storage dual-region and multi-region buckets store objects in at least two places for 99.99% availability during large disruptions. Standard replication is asynchronous with no promised completion time, while turbo replication on dual-region buckets replicates 100% of new objects within 15 minutes. Cloud SQL protection requires automated backups plus binary logging, then recovery replays logs onto a fresh instance that applications must be pointed to after restore. Spanner consistent backups last up to 1 year and restore mounts the backup without copying data, while Firestore in Datastore mode supports scheduled backups up to 14 weeks and point-in-time recovery with 1-minute granularity.
Data encryption and key management controls how data at rest is protected and who controls the unlock keys. Google default encryption uses Google-owned keys with no customer action, while Customer-managed encryption keys (CMEK) are customer-controlled keys in Cloud Key Management Service (Cloud KMS) used through envelope encryption. The storage service calls Cloud KMS to encrypt and decrypt its data keys, so the customer key never leaves Cloud KMS but still governs access. Disabling or destroying the active key blocks access within minutes to hours, and data is automatically deleted if the key stays unavailable for more than 30 days.
VPC Service Controls creates a service perimeter around projects and explicitly restricted services, so resources inside can communicate but API calls that copy data outside are denied. The boundary applies to all user and service account identities, which limits exfiltration even if credentials are compromised. Identity and Access Management (IAM) controls who can act inside the perimeter, where a policy binds a principal, a role, and a resource together. Least privilege grants only needed roles at the smallest scope, such as roles/storage.objectViewer for read-only viewing and roles/storage.objectAdmin for full object control without bucket management.
Storage scalability is the ability to handle growing volume and speed without loss of performance. BigQuery separates analytical storage in Colossus from query execution in Dremel, storing compressed columnar data while slots bundling CPU and RAM execute queries. Projects are limited by default to 2,000 slots under on-demand pricing. Cloud Storage has no lower or upper limit and scales automatically, which fits unpredictable growth, static content, backups, and data lakes.
| Service | Capacity limit | Scaling behavior |
|---|---|---|
| Persistent Disk | 10 GiB to 64 TiB per disk, up to 257 TiB per VM | Scale up, add disks, autoscale |
| Google Cloud Hyperdisk | 4 GiB to 64 TiB per disk, up to 512 TiB per VM | Scale up with dynamic resizing |
| Local SSD | 375 GiB per disk, up to 12 TiB per VM | Not scalable, fixed at creation |
Data lifecycle and tiering keeps fresh data in fast storage while aging data moves to cheaper tiers. BigQuery partitioning divides large tables by date, timestamp, or integer column so queries scan only matching partitions, and clustering on up to four columns sorts data inside partitions to skip blocks. Capacity planning sizes storage, memory, and compute for current plus future growth, where Cloud SQL needs at least 20% free space for maintenance and growth controls set autogrow before thresholds hit. Relief flows in order: first scale up disk, vCPU, and RAM, then defragment and tune queries, then scale reads to replicas, then re-architect with partitioning or sharding.
Continue with this study guide: Configuring individual storage systems
Compute resource provisioning (opens in a new tab) matches a workload to the right Google Cloud service based on control, packaging, and scaling needs. Compute Engine provides configurable VM and bare metal instances where you manage the guest operating system. Managed platforms reduce operational effort by having Google manage infrastructure while you deploy code or containers.
Selection of compute services balances flexibility against operational effort while supporting availability and scalability. Infrastructure as a Service (IaaS) gives the most control because you manage the operating system, middleware, and applications while Google manages virtualization, servers, storage, and networking. Traffic flows from users to your VMs, and your VMs depend on the networks, disks, and images you attach to them.
| Service | What you manage | Choose when |
|---|---|---|
| Compute Engine | Guest OS, middleware, applications, machine type, disks, and accelerators | The workload needs full OS control, custom networking, GPUs or TPUs, or lift-and-shift compatibility |
| Google Kubernetes Engine (GKE) | Container workloads, scaling policies, and in Standard mode node configuration | The workload is containerized and needs orchestration across many services |
| Cloud Run | Container image and request logic; no servers or clusters | The workload is request- or event-driven and should scale automatically |
Google Kubernetes Engine (GKE) deploys and manages containerized applications with orchestration, with Autopilot managing nodes automatically and Standard leaving node pools to the user. Cloud Run is fully managed serverless containers that scale on requests or CPU without managing servers, while Batch schedules queued asynchronous jobs on provisioned Compute Engine resources without managing VMs directly. Only stateless managed instance groups support autoscaling, because stateful groups must preserve persistent disks during repair.
Exam tip: Choose Compute Engine when OS-level control is required, GKE when container orchestration across services is required, and Cloud Run when you want serverless containers without managing servers or clusters.
Standard and Spot VMs differ in cost, capacity assurance, and lifecycle stability. Standard VMs use on-demand pricing, run until you stop or delete them, and support live migration and automatic restart during host events. Spot VMs use excess Compute Engine capacity at discounts of up to 91% on vCPUs, memory, GPUs, and Local SSD disks, but they offer only best-effort availability with no SLA.
| Model | Pricing | Capacity and lifecycle |
|---|---|---|
| Standard VMs | Standard on-demand rates while in use | Runs until stopped or deleted; supports live migration and automatic restart |
| Spot VMs | Discounted up to 91% while in use | Best-effort from excess capacity; stopped by default or deleted at any time |
When preemption occurs, the Spot VM moves from running to stopped or deleted based on its termination action, and local data such as Local SSD content is lost. Spot VMs suit stateless, fault-tolerant work such as batch processing, data analytics, high performance computing, media encoding, and distributed training. Choose standard VMs for critical production, inference and model serving, non-replicated databases, and stateful servers that must stay available. A mixed architecture keeps a stable base on standard VMs and adds volatile capacity on Spot VMs, with node taints and tolerations in GKE keeping critical work off Spot nodes.
Preemption handling runs cleanup during the short window before shutdown, typically stopping gracefully and uploading a checkpoint file to Cloud Storage. Shutdown script handling uses 0s default with best-effort cleanup up to 30 seconds, while in-workload handling uses a 120s notice duration where workload code reacts during the notice period. Resilience increases with regional managed instance groups that spread instances across zones and with autohealing health checks.
VPC networking for Compute Engine and GKE defines how VMs and Kubernetes Pods communicate securely inside a VPC network. Every Compute Engine VM and GKE node sits in a VPC subnet and receives an internal IP address from the subnet range, and resources in the same VPC reach each other over internal addresses when firewall rules permit. VMs with only internal addresses can still reach Google APIs through Private Service Connect or Private Google Access, while Cloud NAT handles outbound traffic.
Shared VPC centralizes administration when Compute Engine and GKE live in separate projects. A host project holds the VPC network and subnets, while service projects attach and deploy VMs or clusters into approved subnets. Cloud Next Generation Firewall (Cloud NGFW) enforces stateful filtering on each VM, including GKE nodes, with hierarchical policies permitting IAP TCP forwarding from 35.235.240.0/20 and load balancer health checks from 35.191.0.0/16 and 130.211.0.0/22.
GKE VPC-native networking makes Pod addresses part of the VPC through alias IP ranges, so Pod traffic is native VPC traffic. Kubernetes NetworkPolicy controls east-west traffic between Pods by selecting Pods by label and restricting ingress and egress, while VPC firewall rules protect GKE nodes at Layer 3 or Layer 4. External north-south access flows from internet clients to Cloud Load Balancing to Ingress to the Service to selected Pods. Serverless connectivity uses Direct VPC egress directly into the VPC or a Serverless VPC Access connector with its own range such as 10.8.0.0/28, and Google Cloud VMware Engine (GCVE) runs VMware on dedicated bare metal connected through private services access peering with migration by VMware HCX.
Automated resource configuration defines compute once and lets Google Cloud create identical VMs automatically. A Compute Engine instance template is an immutable blueprint that standardizes machine type, boot image, network, labels, service account, scheduling, and metadata for every VM in the group. A managed instance group (MIG) depends on that template, keeps the group at its targetSize, names VMs from its baseInstanceName, and recreates failed VMs from the same blueprint.
Deploying a standardized tier follows an ordered workflow:
targetSize such as 2.http:80 so the backend service routes to the correct port.create_before_destroy for template updates.VM Manager manages operating systems across large fleets, where Patch applies updates, OS inventory management collects package data, and OS policies maintains desired state. A patch job moves from pre-patch script to retrieving and applying updates to reboot to post-patch health check, with zone-by-zone staggering and disruption budgets. Terraform is the recommended declarative tool where configuration declares desired state, with Infrastructure Manager running Terraform from Google Cloud without a separate automation server and Config Connector managing resources as Kubernetes objects.
Google Kubernetes Engine (GKE) is a managed service to deploy, manage, and scale containerized workloads on Kubernetes. GKE Standard and Autopilot modes define who manages nodes and how billing works, where Standard bills for VM infrastructure backing each node and Autopilot bills only for CPU, memory, and ephemeral storage requested by Pods. A regional cluster runs control plane replicas across multiple zones, while a zonal cluster runs a single control plane in one zone. All Autopilot clusters are regional, and regional clusters are recommended for production.
| Autoscaler | Reacts to | Scales what |
|---|---|---|
| Horizontal Pod Autoscaler | Current load above or below target | Number of Pod replicas |
| Vertical Pod Autoscaler | Historical usage too high or too low | CPU and memory requests per Pod |
| Cluster Autoscaler | Pending Pods or underutilized nodes | Number of nodes in node pools |
In a surge, the flow moves in order: first the Horizontal Pod Autoscaler increases replicas, then the scheduler attempts placement, and only after Pods remain Pending does the Cluster Autoscaler add nodes. Cloud Run services are request-driven deployments that scale to zero by default when idle, while minimum instances can keep instances warm to reduce cold starts. Cloud Run jobs are finite executions that run batch work to completion, and Cloud Run functions respond to cloud events where data flows from the event source through Eventarc or Pub/Sub to the function. Temporary files inside a container disappear when the instance stops, so persistent state depends on Cloud Storage, Firestore, and Cloud SQL.
Continue with this study guide: Configuring compute systems
Vertex AI Pipelines (opens in a new tab) is the serverless orchestrator that runs the end-to-end machine learning lifecycle from data preparation to training, evaluation, and deployment. Vertex AI manages training, evaluation, and serving, while Cloud Run provides serverless execution for event-driven steps. Together they enable scheduled retraining, event-driven retraining on new data or drift, and safe promotion from development to production.
Automated pipeline execution means Vertex AI Pipelines runs start on a schedule or in response to an event instead of a person starting them manually. A scheduled run handles recurring work such as retraining on a fixed interval, while a triggered run responds to new data, code changes, or monitoring signals. Event-driven orchestration connects systems directly to triggers, where BigQuery ingesting new data can trigger training and Cloud Logging can publish a message when Vertex AI Model Monitoring emits skew and drift alerts.
Cloud Build is the continuous integration and delivery engine that builds, tests, and packages ML code into versioned containers and pipeline definitions. Cloud Build imports source from Artifact Registry, GitHub, or Bitbucket, then executes a build defined in a cloudbuild.yaml file as a series of build steps where each step runs in a Docker container. Cloud Build captures build-time values through substitutions such as $COMMIT_SHA, $REPO_NAME, $BRANCH_NAME, and $BUILD_ID so an image traces back to the source change.
The Cloud Build workflow from source commit to pipeline deployment converts tested code into versioned artifacts:
/workspace directory.$COMMIT_SHA.component.yaml file to point to the newly built images.pipeline.json definition file.pipeline.json to Artifact Registry for versioning and reuse.Exam tip: Cloud Build triggers respond to code commits and produce versioned images and pipeline.json, while scheduled Vertex AI Pipelines runs and data or drift events trigger retraining on new data.
Modular pipeline architecture means building the ML lifecycle as small, reusable steps that run in dependency order. A pipeline component is a self-contained set of code that performs one step such as preprocessing or training. A pipeline job is an execution instance of a pipeline definition, represented by the PipelineJob resource as tasks connected by input-output dependencies. The orchestrator builds a directed acyclic graph (DAG) so a training step runs only after its transformed dataset arrives.
| Option | Version requirement | Choose when |
|---|---|---|
| Kubeflow Pipelines (KFP) | SDK v2.0 or later | For general use cases and custom frameworks such as PyTorch, scikit-learn, or XGBoost |
| TensorFlow Extended (TFX) | v0.30.0 or later | When the workflow uses TensorFlow to process terabytes of structured or text data |
Assembly chains components from ingestion to deployment, where data flows from Cloud Storage into preprocessing, then training data flows into training that outputs a model artifact, then the model flows into Vertex AI Model Registry, then the registered name flows into a deploy component that assigns the model to an endpoint. Vertex ML Metadata automatically tracks which dataset, parameters, code, and component version produced which model and metrics. Execution governance uses dedicated IAM service accounts with least privilege, dedicated projects or VPCs with firewalls, Shielded VMs or Confidential Computing, and Private Service Connect for private communication to other Google Cloud services.
Vertex AI Feature Store is a managed service for centralized feature management and secure feature serving. A feature is a measurable attribute of an entity used to train a model or make a prediction. BigQuery serves as the offline store, so teams do not copy feature data into a separate offline store, and the same BigQuery source can feed both training and online inference while retaining historical data for retraining.
Data flows in order from BigQuery to online inference:
FeatureOnlineStore instance, choosing optimized online serving for ultra-low latency or Bigtable online serving for high data volumes.Online serving retrieves the latest values for a small subset of entities for real-time inference, while offline serving exports high volumes of current and historical data for training. When Feature Store is used as a vector database for RAG Engine, embedding vectors stored in BigQuery flow into the online store and RAG Engine queries them with low latency. Choosing training storage separates ML data from operational systems, where BigQuery holds tabular data as materialized tables and Cloud Storage holds unstructured data in large sharded formats of at least 100 MB with 100 to 10,000 shards, using TFRecord for TensorFlow and Avro for other frameworks.
AI Hypercomputer is an integrated supercomputing system that combines performance-optimized infrastructure, open software, and flexible consumption models. Performance-optimized infrastructure supplies accelerators, networking, and storage, open software supplies optimized frameworks and orchestration, and consumption options supply flexible provisioning. Because the three layers are designed together, scheduling overhead drops and accelerators spend more time on productive computation measured as goodput.
| Compute option | Accelerators | Orchestration | Scaling |
|---|---|---|---|
| Cloud Run | NVIDIA RTX PRO 6000 Blackwell and L4, no TPUs | Serverless, Google-managed OS | Automatic based on requests or CPU consumption |
| Google Kubernetes Engine (GKE) | All available NVIDIA GPUs and all available Google TPUs | Kubernetes, batch jobs through Kueue | Automatic pod and node scaling |
| Cluster Director | NVIDIA GB300 Ultra, GB200, B200, H200, and H100, no TPUs | Slurm | Automatic based on CPU consumption |
General GPUs are managed as independent units, while clustered GPUs are managed as a single tightly coupled system with synchronized maintenance. Cloud TPUs are designed for large matrix operations, include on-chip high-bandwidth memory (HBM), and connect in groups called slices that scale with little to no code changes. Reservation-bound capacity provides densely placed VMs allocated close together, while Spot VMs with GPUs provide lower prices but persist only for the life of the VM and are less likely allocated close together. Topology Aware Scheduling (TAS) with Kueue places workers to minimize communication overhead, which matters for large language model training where gradient aggregation requires specific rank order.
Continue with this study guide: Leveraging Vertex AI for end-to-end ML workflows
Prebuilt AI APIs (opens in a new tab) match input type and job to the right managed model without training from scratch. Analytical APIs describe existing images, video, audio, or text, while generative models create new content or reason across modalities. BigQuery can call several of these APIs directly with remote models for bulk transcription and text understanding.
Cloud Vision API is a pretrained API that analyzes a single still image in one request and returns structured annotations. An application sends image bytes or a Cloud Storage reference with one or more feature types, and annotations flow back from the API to the application. Because each request is independent, the API has no sense of motion or change over time. Behavior is controlled by explicit feature types such as OBJECT_LOCALIZATION, TEXT_DETECTION, DOCUMENT_TEXT_DETECTION, LABEL_DETECTION, FACE_DETECTION, and SAFE_SEARCH_DETECTION.
Video Intelligence API is a pretrained API that analyzes video as a sequence of frames over time. Data flows from a video file in Cloud Storage or a local upload to an annotation operation, and time-stamped segments flow back to the application. The API annotates labels across a clip, detects persons and faces as they move, detects text in frames, flags explicit content, detects shot changes, and transcribes speech. Imagen is a text-to-image model on Vertex AI that produces or edits images from a text prompt, including inpainting where the model creates the masked area.
| API or Model | Input | Primary action |
|---|---|---|
| Cloud Vision API | Single image or document page | Detects objects, text, faces, labels, and properties in that frame |
| Video Intelligence API | Video file or stream with audio and frames | Tracks labels, persons, faces, text, shots, and speech over time |
| Imagen and Gemini | Text prompt, optionally plus image, video, or audio | Generates new images, edits regions, or answers across modalities |
Exam tip: Use TEXT_DETECTION for sparse text in a photo, but use DOCUMENT_TEXT_DETECTION for dense documents, PDF/TIFF files, or handwriting.
Conversational interfaces turn audio into text, understand the text, and turn a response back into audio. Speech-to-Text transcribes audio into written text, so a voicemail recording flows to the API and text flows back. Text-to-Speech does the opposite direction, where text flows to the API and synthesized audio flows back for playback. Cloud Natural Language API understands written text such as classification and sentiment, where text flows from a table or application to the API and insights flow back.
Dialogflow orchestrates multi-turn conversations, where Dialogflow ES fits standard linear dialogs and Dialogflow CX fits complex branching with state handling and scaling. An automated voice agent connects these steps in a fixed order:
Supporting services surround this loop, where Contact Center as a Service queues and routes interactions, Agent Assist helps human agents after handoff, and Cloud Translation API translates text across languages. In BigQuery, transcription uses a remote model over the Speech-to-Text endpoint with ML.TRANSCRIBE, and text understanding uses ML.UNDERSTAND_TEXT against standard tables.
Vertex AI Search is a prebuilt search solution for building a discovery engine or Retrieval Augmented Generation (RAG) application. Data flows from the user query to Vertex AI Search, and matching documents flow back from the corpus, so teams do not have to build ranking from scratch. Grounding connects Gemini output to enterprise data in a Vertex AI Search data store, where Vertex AI Search supplies retrieved documents and Gemini generates an answer using that context.
Gemini Enterprise is an intranet search, AI assistant, and agentic platform that connects Google and third-party data sources. Data flows from connected sources into dedicated data stores, then through a Gemini Enterprise app to the user, and the app only serves content that user is allowed to see. Permissions-aware access respects existing permissions through single sign-on and user-level access, with data encrypted by default with customer-managed keys and integration with VPC Service Controls. Gemini Notebook Enterprise confines chat answers strictly to hand-selected notebook sources, so responses stay grounded in selected documents. Choose Notebook Enterprise when answers must stay confined to known sources, and choose Gemini Enterprise when the starting point is unknown across enterprise systems.
A conversational data agent answers natural-language questions from databases, where requests flow from the user to the agent, then from the agent to the database as generated SQL, and back as a summary or visualization. An agent starts as draft for personal testing, then moves to published when made available to others. Grounding with authored context such as schema descriptions keeps answers inside enterprise controls, while grounding with Google Search sends requests outside customer controls and does not support CMEK, Data Residency, or VPC Service Controls.
Vertex AI Model Garden is the central library for discovering, testing, customizing, and deploying Google and partner models. Discovery happens on the Model Garden page, where each listing has a model card with details before testing or deployment. Task matching selects the smallest fit: Codey for code generation, Imagen for image generation, text-embedding models for search vectors, and Gemini as the flagship model for text and multimodal prompts.
| Category | Description | Deployment approach |
|---|---|---|
| Foundation models | Pretrained multitask large models | Tune or customize for a task, then deploy to an endpoint |
| Fine-tunable models | Models that support custom training | Fine-tune with a notebook or pipeline, register the new version, then deploy |
| Task-specific solutions | Prebuilt models for a defined task | Deploy as ready-to-use models, many customizable on your own data |
Fine-tuning continues training of a pretrained foundation model on domain-specific data, where an encoder converts input into an intermediate representation and a decoder converts it into output using self-attention. Choose fine-tuning when a pretrained model matches the domain, AutoML when data matches a supported format and latency above 100ms is acceptable, and custom training when the problem does not match BigQuery ML or AutoML. Deployment moves a procured asset from Cloud Marketplace to a serving Vertex AI Endpoint through the registry:
gs://bucket/directory, and a serving container image URI.n1-standard-2.Only after deployment completes can applications send inference requests and receive predictions. Deployment is long-running and can take more than five minutes. Models flagged as unsafe by scans are blocked from deployment, while models flagged as suspicious can still be deployed but require thorough review.
Continue with this study guide: Configuring prebuilt solutions or APIs with Agent Platform
$COMMIT_SHA and a compiled pipeline.json, while Vertex AI Feature Store uses BigQuery as the offline store and Cloud Run provides serverless GPUs without TPUs compared to GKE with all GPUs and TPUs.Cloud VPN uses encrypted IPsec tunnels over the internet with lower bandwidth and is chosen for low cost or experimental use, while Dedicated Interconnect and Partner Interconnect provide private circuits that do not traverse the internet. Dedicated Interconnect is a direct connection to Google with 10, 100, or 400-Gbps circuits up to 8 per connection and VLANs from 50 Mbps to 50 Gbps that requires colocation equipment, while Partner Interconnect is a dedicated path through a supported provider with VLANs from 50 Mbps to 50 Gbps and no colocation equipment.
Virtual Private Cloud (VPC) Network Peering exchanges routes for VM-to-VM reachability but keeps separate firewall and Identity and Access Management controls, requires non-overlapping IPs, and is not transitive. Shared VPC shares subnets in one host-project VPC with central network control in one organization, while Private Service Connect (PSC) connects a consumer VPC privately to a published producer service without merging networks.
Spot VMs use excess Compute Engine capacity at discounts of up to 91 percent with only best-effort availability and no service level agreement, and they are stopped or deleted at any time with loss of local data such as Local SSD content. They suit stateless fault-tolerant work such as batch processing, data analytics, and media encoding, while standard VMs run until stopped or deleted with live migration and automatic restart for critical production, databases, and stateful servers.
Regional Persistent Disk replicates every write synchronously across two zones in the same region so a backup VM in the surviving zone can force-attach the disk and continue with near-zero Recovery Point Objective (RPO). Snapshots are incremental point-in-time copies of Persistent Disk volumes that can be taken from running instances and copied across regions for disaster recovery and protection against data loss.
Last updated