Google Cloud Professional Cloud Architect Data, Analytics, and AI Solution Architecture Syllabus

What is Dataproc?

A fully managed service for running open-source Apache Hadoop, Spark, Hive, Presto, and related tools on Google Cloud, with fast cluster provisioning (typically under 90 seconds).

What Dataproc pattern minimizes cost for transient workloads?

Ephemeral (job-scoped) clusters: spin up a cluster for a single job using autoscaling and Spot/preemptible workers, store data in Cloud Storage (not HDFS), then delete the cluster.

Why use Cloud Storage instead of HDFS as the data layer for Dataproc?

Decoupling storage from compute lets clusters be ephemeral and stateless, enables sharing data across jobs/clusters, and provides durable, scalable storage via the Cloud Storage connector.

When should you choose Dataproc over Dataflow?

Choose Dataproc when migrating existing Hadoop/Spark jobs or when you need the OSS ecosystem; choose Dataflow for new serverless, autoscaling unified stream/batch pipelines with no cluster management.

What is Looker?

Google Cloud's enterprise BI and data-application platform that uses a semantic modeling layer (LookML) to define metrics centrally and query data in place within the warehouse.

What is LookML?

Looker's modeling language that defines dimensions, measures, and relationships once in a central semantic layer, generating SQL on the fly so all users share consistent, governed metric definitions.

What is the difference between Looker and Looker Studio?

Looker is an enterprise, governed BI platform with a centralized LookML semantic model; Looker Studio (formerly Data Studio) is a free, self-service tool for ad hoc dashboards and reports.

What is Dataplex?

An intelligent data fabric / governance service that unifies, catalogs, organizes, and governs distributed data across data lakes, warehouses, and marts with centralized metadata, quality, and security.

How does Dataplex organize data assets logically?

Using lakes (logical domains), zones within lakes (raw vs. curated), and assets that map to physical resources like Cloud Storage buckets and BigQuery datasets.

What governance capabilities does Dataplex provide?

Unified data catalog and metadata discovery, data quality and profiling, data lineage, and centralized access policies across distributed storage.

What are the key dimensions for choosing the right database on Google Cloud?

Data model (relational vs. NoSQL vs. key-value), consistency needs, scale/throughput, latency, query patterns (transactional vs. analytical), and global vs. regional reach.

What is Cloud Spanner?

A fully managed, horizontally scalable relational database providing strong (external) consistency, SQL support, and synchronous replication across regions—combining relational structure with global scale.

What technology gives Spanner globally consistent transactions?

TrueTime, an API using atomic clocks and GPS to provide globally synchronized time with bounded uncertainty, enabling external consistency for distributed transactions.

When should you choose Spanner over Cloud SQL?

Choose Spanner when you need relational/SQL with horizontal scale beyond a single node, very high availability (up to 99.999%), and strong consistency across regions; choose Cloud SQL for traditional single-region MySQL/PostgreSQL/SQL Server workloads.

What is Cloud Bigtable?

A fully managed, petabyte-scale, wide-column NoSQL database for high-throughput, low-latency workloads, ideal for time-series, IoT, financial, and analytical data with massive read/write volumes.

Why is Bigtable row key design critical?

Bigtable is a sorted key-value store; poorly designed keys (e.g., monotonically increasing timestamps) cause hotspotting on a single node. Good keys distribute load evenly and support efficient range scans.

When choose Bigtable vs. BigQuery?

Bigtable for high-throughput, low-latency operational reads/writes and single-row lookups by key; BigQuery for ad hoc analytical SQL queries and aggregations over large datasets.

What is Firestore and its primary use case?

A serverless, fully managed NoSQL document database with real-time synchronization and offline support, ideal for mobile/web app backends, user profiles, and semi-structured documents.

What is Memorystore?

A fully managed in-memory data store service for Redis and Memcached, used for sub-millisecond caching, session storage, and leaderboards to offload backend databases.

Compare Firestore and Bigtable for workload selection.

Firestore: document model, strong consistency, real-time sync, moderate scale for app backends. Bigtable: wide-column, very high write throughput and massive scale for time-series/analytics—no real-time client sync.

What is Vertex AI?

Google Cloud's unified machine learning platform that brings data engineering, model training (AutoML and custom), deployment, MLOps, and generative AI (Model Garden, Gemini) into a single environment.

What is the difference between AutoML and custom training in Vertex AI?

AutoML lets you train high-quality models with minimal code on your data (Google handles architecture/tuning); custom training gives full control using your own code, frameworks, and containers.

What are Google's pre-trained ML APIs and give examples.

Ready-to-use APIs requiring no training: Vision API (image analysis), Natural Language API (entity/sentiment), Speech-to-Text and Text-to-Speech, Translation API, and Video Intelligence API.

When should you use a pre-trained API versus AutoML versus custom training?

Pre-trained API for common tasks with no custom data; AutoML when you need a custom model on your labeled data with little ML expertise; custom training when you need full control over architecture and code.

What generative AI services does Vertex AI offer?

Access to foundation models (Gemini, Imagen, etc.) via Model Garden, plus tools for prompt design, grounding, fine-tuning, embeddings, and RAG to build generative AI applications.

Which accelerators does Google Cloud offer for ML workloads, and when use each?

GPUs (general-purpose for training/inference across many frameworks) and TPUs (custom ASICs optimized for large-scale matrix math, best for large TensorFlow/JAX deep-learning training).

What architectural factors matter most when designing for large ML training workloads?

Choice of accelerator (GPU/TPU), distributed training strategy, high-throughput data pipeline (e.g., from Cloud Storage), use of Spot VMs for cost, and managed Vertex AI training to avoid infra management.

What is MLOps?

The practice of applying DevOps principles to machine learning: automating and managing the end-to-end ML lifecycle—data prep, training, deployment, monitoring, and retraining—for reliable, reproducible models in production.

What is model/data drift and how is it handled in MLOps?

Drift is degradation of model performance when production data diverges from training data (data drift) or relationships change (concept drift). It's handled via Vertex AI Model Monitoring with alerts that trigger retraining pipelines.

What Vertex AI components support the MLOps lifecycle?

Vertex AI Pipelines (orchestration), Feature Store (managed features), Model Registry (versioning), Experiments (tracking), and Model Monitoring (drift/skew detection).

What is the difference between training-serving skew and data drift?

Training-serving skew is a discrepancy between how data is processed during training versus serving (often a pipeline bug present from day one); data drift is a gradual change in the live input distribution over time after deployment.