Design an Email Service (like Gmail)

Email Send Flow

1. User clicks "Send" → POST /api/messages/send
2. Validate: recipients exist, attachment size < 25MB, rate limit check
3. Store email body + attachments to Blob Store (S3)
4. Store email metadata to Bigtable/Cassandra
5. Enqueue to send queue (Kafka topic: outgoing-emails)
6. SMTP Sender Worker picks up from queue:
   a. DNS MX lookup: recipient domain → find receiving mail server
   b. Open TLS connection to receiving server (STARTTLS)
   c. Authenticate: sign with DKIM key for sender's domain
   d. Transmit email via SMTP protocol
   e. Receiving server ACKs → mark as delivered
   f. If rejected → generate bounce (DSN) → deliver to sender's inbox
7. If temporary failure: retry with exponential backoff (1min → 72h)
   After 72h of retries → permanent failure → bounce to sender

Optimization for internal emails (sender@gmail → recipient@gmail):
   Skip SMTP entirely → directly store in recipient's mailbox → 100ms delivery

Incoming Email Flow (SMTP Receive)

1. External sender's MTA connects to our SMTP gateway (MX record)
2. SMTP handshake: EHLO, MAIL FROM, RCPT TO
3. Before accepting DATA:
   a. SPF check: is sender's IP authorized for their domain?
   b. Rate limiting: too many emails from this IP? → 421 temporary reject
   c. Recipient exists? → 550 user unknown if not
4. Accept DATA → email content streamed
5. DKIM verification: check cryptographic signature
6. DMARC evaluation: combine SPF + DKIM → pass/fail policy
7. Spam classification: ML model scores email (0-1)
   Score > 0.7 → spam folder; 0.3-0.7 → show warning; < 0.3 → inbox
8. Virus scan: check attachments for malware (ClamAV)
9. Store body to Blob Store, metadata to Bigtable
10. Index in Elasticsearch for search
11. Push notification to recipient (if enabled)
12. Return 250 OK to sender's MTA

Spam Filtering Architecture

Layer 1: Connection-level filters (< 1ms per connection)
  - IP reputation: is this IP in known spammer lists? (Spamhaus ZEN)
  - Volume limits: > 1000 emails from this IP in last hour → rate limit
  - DNS checks: does IP reverse-resolve to a legitimate hostname?
  Catches: ~70% of spam before content inspection

Layer 2: Content analysis (10-50ms per email)
  - SPF/DKIM/DMARC: cryptographic proof of sender legitimacy
  - Spam rules: regex patterns, known spam phrases, HTML structure
  - URL analysis: links in email → check against phishing databases
  - Attachment scanning: ClamAV for malware
  Catches: additional ~20% of spam

Layer 3: ML classification (100-300ms per email)
  - NLP model: BERT fine-tuned on spam/ham corpus
  - Features: text, sender reputation, social graph, historical interaction
  - Score > 0.85 → spam folder; 0.5-0.85 → warning; < 0.5 → inbox
  Catches: additional ~9% of spam

Result: < 0.1% false positive rate (legitimate email marked as spam)

Push vs Pull (IMAP IDLE vs Polling)

IMAP IDLE: server holds connection open, pushes EXISTS notification. Google Sync: FCM (Android) / APNs (iOS) for battery-efficient push. Web client: WebSocket for real-time inbox updates.

SPF, DKIM, DMARC

SPF: DNS record listing authorized IPs for a domain. DKIM: Cryptographic signature in email header. DMARC: Policy telling receivers what to do if SPF/DKIM fails (reject/quarantine).