A Summary and Speculation on Google’s Overall Architecture

Based on publicly available sources and the author’s experience, this article presents a summary and speculation of Google’s overall architecture.

Products

Google’s services can be grouped into six major categories: various search services (web, image, video), advertising systems (AdWords, AdSense), productivity tools (Gmail, Google Apps), geographic products (Maps, Google Earth, Google Sky), video streaming (YouTube), and the PaaS platform Google App Engine.

Design Principles

Google’s design philosophy can be distilled into six key principles:

Scale, Scale, Scale – massive scalability drives the development of frameworks like MapReduce and platforms like Google App Engine.

Fault tolerance – distributed systems must handle frequent hardware and software failures, as illustrated by the high failure rate in large X86 clusters.

Low latency – minimizing response time is critical for user experience, prompting the deployment of local data centers.

Cheap hardware and software – Google builds its own stack (MapReduce, BigTable, GFS) on inexpensive X86 servers running open‑source Linux.

Prefer moving computation over moving data – processing data where it resides reduces network costs.

Service‑oriented architecture – loosely coupled services (hundreds to thousands) enable rapid development, testing, and scaling.

Speculative Overall Architecture

Google’s workloads fall into three categories:

Local interaction – services that run close to the user to reduce latency (e.g., web search).

Content delivery – large‑scale storage, generation, and management of data (e.g., indexing, YouTube videos, Gmail storage) using Google’s custom distributed stack.

Critical business – enterprise‑grade systems such as advertising platforms that require high SLA.

Google’s data centers can be divided into two types:

Giant Data Centers

These host over 100,000 servers, are located near power plants, and focus on cost‑effective high‑throughput content delivery, using custom hardware and software. Example: a facility in Oregon consuming 103 MW.

Medium‑Sized Data Centers

These contain thousands to tens of thousands of servers, are placed close to users, and prioritize low latency and high availability, often using standard hardware (e.g., Dell servers, MySQL databases). Example: the former Google China data center in Beijing.

The table below compares the two types:

Aspect

Giant Data Center

Medium‑Sized Data Center

Workload

Content delivery

Local interaction / Critical business

Location

Near power plant

Near users

Design focus

High throughput, low cost

Low latency, high availability

Server customization

High

Low

SLA

Normal

High

Server count

>100k

>1k

Number of centers

Less than 10

Dozens

PUE estimate

1.2

1.5

Conclusion

When a typical user accesses Google services, the request is routed based on IP or ISP to the nearest local data center; if that center cannot satisfy the request, it is forwarded to a remote content‑delivery center. Advertising requests are sent directly to specialized critical‑business data centers.

All observations are based on public information and personal speculation and do not reflect Google’s actual internal operations.