Gmail

Overview

Gmail is the world’s largest email service with 1.8B+ users, processing and storing hundreds of billions of emails. Its architecture must handle high-throughput email delivery (receiving and sending), massive full-text search across years of email history, real-time spam and phishing detection that blocks 99.9%+ of threats, and smart features like auto-categorization, Smart Reply, and Smart Compose. Gmail pioneered the “search, don’t sort” approach to email with its label-based organization and conversation threading. Built on Google’s Bigtable and Spanner infrastructure, Gmail provides strong consistency and global availability.

Scale

Requirements

High-Level Architecture

Loading diagram...

Key Components

Data Model

Loading diagram...

Spam Detection Pipeline

Gmail blocks ~15 billion spam emails per day with a 99.9%+ detection rate — one of the most effective spam filters ever built.

Multi-layer approach:

1. Connection-level: IP reputation, SPF/DKIM/DMARC validation, rate limiting. Rejects obvious spam before the email body is even received.
2. Content analysis: ML classifiers analyze email headers, body text, URLs, and attachments. Models are trained on billions of user-reported spam signals.
3. URL scanning: Every URL in every email is checked against Google’s Safe Browsing database. Suspicious URLs are sandboxed — clicked links go through a Google redirect that re-checks the destination in real time.
4. Attachment scanning: Attachments are scanned for malware using multiple antivirus engines and sandboxed execution. Encrypted attachments and password-protected ZIPs receive extra scrutiny.
5. Behavioral signals: New sender patterns, bulk sending characteristics, and engagement rates (do users delete without reading?) feed into the spam model.

Adversarial training: Spammers constantly evolve tactics. Gmail’s models are retrained daily on fresh data, and the system can deploy emergency rules within minutes of detecting a new spam campaign.

Loading diagram...

Bigtable Email Storage

Gmail stores email content in Bigtable, Google’s distributed wide-column store, which provides the performance characteristics email needs.

Data model: Each user’s emails are stored in a single Bigtable row range, sorted by a key combining user_id and reverse timestamp. This means the most recent emails are physically stored together, making inbox loading a single sequential read.

Column families: Email data is split into column families — headers (small, frequently read), body (larger, read on demand), and attachments (large blobs stored separately with references). This allows loading the inbox (headers only) without reading full email bodies.

Compaction and storage: Bigtable’s LSM-tree storage engine handles Gmail’s write-heavy workload (billions of new emails daily) efficiently. Background compaction merges sorted runs and reclaims space from deleted emails.

Metadata in Spanner: While email content lives in Bigtable, structured metadata (labels, thread relationships, search index pointers) lives in Spanner for its strong consistency and SQL query capabilities. A label change needs to be instantly visible across all devices — Spanner’s external consistency guarantees this.

Smart Reply and Smart Compose

Gmail’s Smart Reply suggests three short response options, while Smart Compose autocompletes sentences as users type. Both are on-device and server-side ML features.

Smart Reply: A sequence-to-sequence model reads the incoming email and generates three diverse response suggestions (e.g., “Thanks!”, “Sounds good, I’ll be there.”, “Let me check and get back to you.”). The model is trained on anonymized response patterns — not on reading users’ actual emails. A diversity mechanism ensures the three suggestions cover different intents.

Smart Compose: As the user types a reply, a language model predicts the next few words and displays them as gray text. The model conditions on the email being replied to (for context) and the user’s typing so far. Pressing Tab accepts the suggestion.

On-device inference: Both features run a lightweight model on-device for privacy (the email content doesn’t leave the device for these features). A larger server-side model provides higher-quality suggestions when available.

Personalization: The models adapt to individual writing style over time — learning common phrases, greetings, and sign-offs that a specific user tends to write.

Architectural Tradeoffs

# Learn From a Gmail-like System

You are an expert software architect and teacher. I want to learn from a Gmail-inspired system design (The world's largest email service with AI-powered organization and search) without pretending this is the exact private production architecture.

Use the following reference as a learning model. Keep assumptions visible, avoid unsupported scale claims, and explain the engineering forces behind each design choice.

---

## System Design Reference
import { Mermaid, Scale, ComponentList, ComponentCard, Requirements, Tradeoff } from '@/components/mdx';

## Overview

Gmail is the world's largest email service with 1.8B+ users, processing and storing hundreds of billions of emails. Its architecture must handle high-throughput email delivery (receiving and sending), massive full-text search across years of email history, real-time spam and phishing detection that blocks 99.9%+ of threats, and smart features like auto-categorization, Smart Reply, and Smart Compose. Gmail pioneered the "search, don't sort" approach to email with its label-based organization and conversation threading. Built on Google's Bigtable and Spanner infrastructure, Gmail provides strong consistency and global availability.

## Scale

<Scale items={{
  "Users": "1.8B+",
  "Emails sent/received per day": "300B+",
  "Spam blocked per day": "~15B",
  "Storage per free account": "15 GB",
}} />

## Requirements

<Requirements
  functional={[
    'Send and receive emails with attachments',
    'Conversation threading and label-based organization',
    'Full-text search across all emails',
    'Spam, phishing, and malware detection',
    'Auto-categorization (Primary, Social, Promotions, Updates)',
    'Smart Reply and Smart Compose suggestions',
    'Calendar and Drive integration',
  ]}
  nonFunctional={[
    'Email delivery latency <5 seconds for 99th percentile',
    'Spam detection accuracy >99.9%',
    'Search results in <500ms across years of email',
    'High availab...

[Full prompt copied to clipboard]