This presentation outlines the journey of a fictional e-commerce shoe company, Gshu, from a simple on-premise garage setup to a globally scalable, data-driven enterprise powered by Google Cloud. The speakers, Pacy and Michael, Google software engineers, detail the evolution of their infrastructure.
Initially, Gshu deployed a basic web application using self-assembled servers and databases. As traffic grew, they scaled horizontally with load balancing and memory caching, but faced limitations with on-premise management. The core of the presentation focuses on migrating this infrastructure to Google Cloud, introducing managed services for networking (VPC), compute (Cloud Run, GKE, Compute Engine), load balancing, diverse database options (Cloud SQL, Spanner, Firestore, Bigtable), caching (Cloud Memory Store), and robust DNS (Cloud DNS).
The second half of the talk delves into leveraging Google Cloud for advanced data processing and AI. This includes data ingestion (Cloud Storage, Pub/Sub), processing (Dataflow, Dataproc, Data Prep), warehousing and analytics (BigQuery with ML capabilities), and sophisticated machine learning through Vertex AI. Vertex AI's offerings cover pre-trained models (Gemini, Vision AI, etc.), AutoML, custom model development, MLOps tools (Pipelines, Registry, Monitoring), and advanced hardware (TPU), enabling the creation of intelligent, predictive applications.
Building Gshu: An E-commerce Journey from Garage to Google Cloud
0:00 Introduction to Gshu: Google engineers Pacy and Michael present a hypothetical e-commerce shoe company, "Gshu," tracing its development from a startup idea to a cloud-powered platform.
0:34 Initial On-Premise Setup: The first iteration involves a garage-based setup with custom-built web and application servers, relational databases (for user data, transactions), and non-relational databases (for articles, product catalogs). Gemini code assistant is mentioned for frontend development.
3:02 Scaling Challenges and Solutions (On-Premise): As sales grow, the system experiences traffic overload. Solutions include vertical scaling (more powerful hardware) and horizontal scaling (adding more machines). Load balancers are introduced to distribute traffic across multiple web and application servers, and a memory caching layer is added for relational databases to improve query performance.
5:35 Migrating to Google Cloud - Core Infrastructure:
5:42 Virtual Private Cloud (VPC): Provides managed networking for all cloud resources.
5:53 Compute Options: Discusses Cloud Run and App Engine for serverless containerized applications, Google Kubernetes Engine (GKE) for flexible container orchestration, and Google Compute Engine (GCE) for maximum control over virtual machines.
7:36 Cloud Load Balancing: A fully distributed, software-defined system offering global reach with anycast IP for low latency and high scalability (millions of queries/second), including internal load balancing for backend services.
8:20 Database Choices: Cloud SQL (managed relational for MySQL, PostgreSQL, SQL Server), Cloud Spanner (massively scalable relational with strong consistency), Firestore (serverless document database for real-time queries, mobile/IoT), and Bigtable (wide-column NoSQL for heavy reads/writes, low latency, ideal for time-series/recommendations).
9:56 Cloud Memory Store: A fully managed in-memory data storage service (Redis, Memcached) for caching frequently accessed data, ensuring low latency and high performance.
10:44 Cloud DNS: Google's high-volume authoritative DNS serving, offering 100% SLA and low latency through a global network of anycast naming servers.
11:11 Advancing with Data Processing and AI: The presentation shifts to leveraging data for business insights and intelligent applications on Google Cloud.
11:58 Data Ingestion: Cloud Storage is used for scalable batch (offline) data like logs, while Pub/Sub acts as a serverless messaging service for real-time streaming data from web/app servers.
12:40 Data Processing: Options include Dataflow (serverless for batch/streaming, Apache Beam-based), Dataproc (managed Hadoop/Spark clusters), and Data Prep (no-code UI for data transformation).
13:56 Data Warehousing & Analytics: BigQuery is highlighted as a serverless data warehouse that scales to petabytes, supports standard SQL, and offers built-in machine learning capabilities (BigQuery ML) for predictions.
15:23 Advanced Machine Learning with Vertex AI: Google's all-in-one AI platform offers a spectrum of ML solutions:
15:51 Pre-trained Models: Includes Gemini (multimodal), Vision AI (logo/face/text detection), Speech-to-Text/Text-to-Speech, and Natural Language AI.
17:16 AutoML: Automatically selects the best model architecture from user datasets, reducing the need for ML expertise.
17:49 Custom Training & Deployment: Provides a full environment for data scientists and ML engineers to build and deploy custom models using frameworks like PyTorch and TensorFlow.
18:19 MLOps Tools & Features: Covers Model Garden (an "app store" for AI models), Agent Builder (for creating chatbots/agents with little code), Vertex AI Pipelines (automating ML workflows), Model Registry (version control for models), Model Monitoring (ensuring production performance), and the use of TPUs for cost-efficient, scalable model deployment.
21:02 Q&A Session: The speakers address audience questions regarding validation of pre-trained models, version control for models (Model Registry), and future plans for IP usage management in GCP.