Design a URL Shortener (TinyURL)

1. What is a URL Shortener?

A URL Shortener is a service that creates shorter aliases for long URLs. When a user clicks these short aliases, they are redirected to the original long URL. This service is a staple of system design interviews because it touches on all the fundamentals: hashing, databases, caching, and concurrency.

Real-World Examples:

  • TinyURL: The classic example.
  • Bit.ly: Focuses on analytics and enterprise features.
  • t.co: Twitter’s internal shortener to save character count.

[!TIP] Try it yourself: Imagine you are sending an SMS. Would you rather type https://www.google.com/search?q=system+design+interview+guide... or https://tiny.url/x7z9? The latter is cleaner, cheaper (SMS cost), and easier to type.

2. Requirements & Goals

Functional Requirements

  1. Shorten: Given a long URL, generate a unique short alias (e.g., http://tiny.url/j8s9).
  2. Redirect: Given a short alias, redirect the user to the original long URL.
  3. Custom Alias (Optional): Users should be able to specify a custom alias (e.g., tiny.url/my-blog).
  4. Expiry: Links should expire after a default timespan (e.g., 5 years) to save space.

Non-Functional Requirements

  1. Availability: The system must be highly available. If the service is down, all URL redirections fail (Global Impact).
  2. Latency: URL redirection happens in real-time. Minimizing latency (< 100ms) is critical.
  3. Read-Heavy: There will be many more redirections (reads) than new link creations (writes). A 100:1 ratio is common.

Extended Requirements

  1. Analytics: Track click counts, location, and user agent.
  2. REST API: Allow other services to generate links programmatically.

3. Capacity Estimation

Traffic Estimates

Let’s assume 100 Million new URLs are generated per month.

  • Write QPS: 100,000,000 / (30 × 24 × 3600) &approx; 40 writes/sec
  • Read QPS (100:1 ratio): 40 × 100 = 4,000 reads/sec
  • Peak Traffic: Assume 5x multiplier for spikes. Write: 200 QPS, Read: 20k QPS.

Storage Estimates

Let’s assume we store the data for 5 years.

  • Total URLs: 100M × 12 months × 5 years = 6 Billion records
  • Average URL size: 500 bytes.
  • Total Storage: 6 Billion × 500 bytes = 3 TB

Bandwidth Estimates

  • Ingress (Write): 40 writes/sec × 500 bytes &approx; 20 KB/s (Negligible).
  • Egress (Read): 4,000 reads/sec × 500 bytes &approx; 2 MB/s.
    • Note: This is just for the payload. If we include 302 redirects, the browser handles the payload of the target site, but our server just sends headers.

[!WARNING] While 3TB fits on a single modern hard drive, we distribute it across multiple nodes for Throughput (IOPS) and Reliability (Replication), not just capacity.

4. System APIs

We will use REST-style APIs for simplicity and compatibility.

Create Short URL

POST /api/v1/shorten

Request Body:

{
  "longUrl": "https://www.very-long-website.com/article/123",
  "customAlias": "my-article",
  "expireDate": "2025-12-31"
}

Response:

  • 200 OK: Success.
  • 400 Bad Request: Invalid URL or Custom Alias already exists.
  • 429 Too Many Requests: Rate limit exceeded.
{
  "shortUrl": "https://tiny.url/my-article",
  "error": null
}

Redirect

GET /{short_alias}

Response:

  • Status: 301 or 302 (See below).
  • Location: https://www.very-long-website.com/article/123
  • Cache-Control: public, max-age=3600 (If 301).

5. Database Design

We need to store 6 Billion records. The data is simple: ShortKey -> LongURL.

SQL Schema (MySQL/PostgreSQL)

CREATE TABLE urls (
    id BIGINT AUTO_INCREMENT PRIMARY KEY, -- 8 bytes, used for Base62
    short_key VARCHAR(7) NOT NULL,        -- 7 bytes (Indexed)
    long_url VARCHAR(2048) NOT NULL,      -- 2 KB max
    user_id BIGINT,                       -- Foreign Key
    created_at TIMESTAMP DEFAULT NOW(),
    expiration TIMESTAMP
);
CREATE INDEX idx_short_key ON urls(short_key);

NoSQL Schema (DynamoDB/Cassandra)

Since we don’t need complex joins, NoSQL is a great fit for horizontal scaling.

{
  "pk": "short_key:x7z9",  // Partition Key
  "long_url": "https://...",
  "created_at": 1672531200
}

Trade-off: SQL provides ACID guarantees (useful for uniqueness checks). NoSQL provides easier scaling. For this problem, both work well, but SQL is often preferred for the AUTO_INCREMENT feature which aids Base62 encoding.

6. High-Level Design

The architecture separates the Write Path (Shortening) from the Read Path (Redirecting) to optimize for the 100:1 read ratio.

System Architecture: URL Shortener
Write Path (Green) vs Read Path (Blue)
Shorten (Write)
Redirect (Read)
Client
LB Layer
Application Services
Data Persistence
Users
Load Balancer
Shortener Svc
• Generates ID
• Writes to DB
Redirect Svc
• Checks Cache
• Returns 302
KGS
Key Gen Service
Redis (Keys)
Pre-gen keys
Main DB
Mapping Table

Shard 1 Shard 2
Redis Cluster
LRU Cache for Hot URLs (20%)
Kafka: clicks_topic
Worker
Aggregates Data
POST /shorten Get ID INSERT GET /abc1 Cache Hit? Cache Miss Async Log

7. Component Design: Generating Keys

This is the core problem. How do we generate a short, unique string like x7z9?

Approach A: Hashing (MD5/SHA256)

Hash the Long URL: MD5(long_url).

  • Result is 128-bit string (32 hex chars).
  • Truncate to 7 characters.
  • Problem: Collisions. Two different URLs might hash to the same first 7 characters.
  • Resolution: If collision, append a salt (e.g., current time) and re-hash. This requires checking the DB for every write, which is slow (Bloom Filters can help, but it’s still probabilistic).

Approach B: Base64 Encoding

Base64 uses A-Z, a-z, 0-9, +, /.

  • Problem: The characters + and / are reserved in URLs. They might be interpreted as spaces or directory separators.
  • Resolution: Use “Base64URL” variant (- and _ instead). However, Base62 is standard for this use case.

Approach C: Base62 Conversion (The Winner)

We use a Counter (Database ID or Distributed ID) and convert it to Base62. This is a Bijective (one-to-one) mapping, guaranteeing no collisions.

  • Base62 Characters: 0-9 (10) + a-z (26) + A-Z (26) = 62.
  • Why 7 characters? 627 &approx; 3.5 Trillion combinations. Enough for 100 years.

Algorithm:

  1. Get Unique ID: 123456789
  2. Convert to Base62: 8M0kX
  3. Short URL: tiny.url/8M0kX

Key Generation Service (KGS)

To scale, we don’t want the App Server to hit the Database for a new ID every time.

  • KGS: A dedicated service (see “KGS” in the diagram) that pre-generates keys (e.g., abc1, abc2) and stores them in a Redis Set (labeled “Redis (Keys)”).
  • Concurrency: App servers POP a key from Redis. This is atomic and fast.
  • Two Tables:
    1. unused_keys: Keys ready to be assigned.
    2. used_keys: Keys already assigned.
  • Trade-off: If the KGS crashes, we lose the pre-generated keys in memory. This is acceptable as we have trillions of keys available.

8. Data Partitioning & Sharding

To handle 6 Billion records and 4000 reads/sec (per node), we must shard the database.

Strategy 1: Range-Based Sharding

  • Method: Partition by the first character of the Short Key.
    • Shard A: Keys starting with A-C
    • Shard B: Keys starting with D-F
  • Pros: Simple to implement.
  • Cons: Unbalanced Load. If many URLs start with ‘A’, Shard A becomes a hotspot.

Strategy 2: Hash-Based Sharding (Consistent Hashing)

  • Method: ShardID = hash(short_key) % NumShards.
  • Pros: Even Distribution. Traffic is spread uniformly across all nodes.
  • Cons: Re-sharding (adding nodes) is complex, requiring data migration. We use Consistent Hashing to minimize movement.

Verdict: Use Hash-Based Sharding on the short_key.

9. Reliability, Caching, & Load Balancing

301 vs 302 Redirects

  • 301 Moved Permanently: Browser caches the redirect. Subsequent requests bypass our server. Fast, but we lose analytics.
  • 302 Found: Browser hits our server every time. Slower, but allows accurate tracking.
  • Choice: Use 302 if analytics are required.

Caching Strategy

  • Cache-Aside: App checks Redis -> Miss -> DB -> Write to Redis.
  • Eviction: Use LRU (Least Recently Used).
  • Pareto Principle: 20% of URLs generate 80% of traffic. Caching just the top 20% fits easily in memory.

Reliability

  • Replication: Use Master-Slave replication for the DB. If Master fails, promote Slave.
  • Geo-DNS: Use a CDN or Geo-DNS to route users to the nearest data center.

10. Interactive Decision Visualizer: Hash vs Base62

Explore why we prefer Base62 over simple Hashing for this specific problem.

**The Gold Standard**: Convert a unique Integer ID (from Database or Snowflake) into a Base62 string.
Bijective Zero Collisions

ID: 123456789
→
Base62 Algo
Divide by 62
→
Key: 8M0kX

11. System Walkthrough (Dry Run)

Let’s trace a request through the system.

1. Request User sends POST /shorten to the Load Balancer.

{
  "longUrl": "https://www.google.com"
}

2. Key Generation The Shortener Service contacts Redis (KGS) to get a unique key.

  • Redis Command: SPOP unused_keys 1
  • Result: "8M0kX"

3. Persistence The service writes to the Primary DB.

INSERT INTO urls (short_key, long_url) VALUES ('8M0kX', 'https://www.google.com');

4. Response Service returns the short URL.

{
  "shortUrl": "https://tiny.url/8M0kX"
}

1. Request User visits https://tiny.url/8M0kX. Request hits Redirect Service.

2. Cache Check Service checks Redis Cache.

  • Redis Command: GET url:8M0kX
  • Result (Hit): "https://www.google.com"
  • Result (Miss): nil -> Read from DB -> Write to Redis.

3. Analytics (Async) Service pushes an event to Kafka.

{
  "topic": "clicks_topic",
  "key": "8M0kX",
  "payload": { "ts": 1672531200, "ua": "Mozilla/5.0", "ip": "1.2.3.4" }
}

4. Redirect Service returns HTTP 302.

  • Header: Location: https://www.google.com

12. Interview Gauntlet

Be ready for these follow-up questions:

  1. What if two users shorten the same long URL?
    • Ans: You can return the same short key (deduplication) to save space, or generate a new one. Deduplication requires checking the DB first (slow), so most systems just generate a new key.
  2. How do you prevent users from guessing IDs?
    • Ans: Base62 is sequential. Add a random salt before encoding, or shuffle the alphabet.
  3. What if the Redis KGS runs out of keys?
    • Ans: A background worker monitors the size of the unused_keys set. When it drops below 10k, it fetches another batch from the KGS database.
  4. How do you handle DB sharding?
    • Ans: Shard by short_key (hash-based). This ensures evenly distributed traffic.
  5. How do you implement custom aliases?
    • Ans: Check if the alias exists in the DB. If not, insert it. This requires a unique constraint on the short_key column.
  6. What happens if the cache is full?
    • Ans: Use an LRU (Least Recently Used) eviction policy.
  7. How do you filter malicious URLs?
    • Ans: Run the long URL against a blacklist (e.g., Google Safe Browsing API) before shortening.
  8. Can we use Zookeeper for ID generation?
    • Ans: Yes, ZK can manage ranges (e.g., Server A gets 1-1000, Server B gets 1001-2000).
  9. Why not use UUIDs?
    • Ans: UUIDs are 36 characters. Base62 of a UUID is still too long (~22 chars). We want short URLs (7 chars).
  10. How to handle heavy analytics writes?
    • Ans: Never write to DB on the read path. Use Kafka to buffer events and write in batches to a Columnar DB (ClickHouse).

13. Whiteboard Summary

1. Core Requirements

  • Shorten (Write): 40 QPS
  • Redirect (Read): 4k QPS (100:1)
  • Unique 7-char Key (Base62)
  • 301 vs 302 Redirect

2. Key Architecture

  • Write Path: KGS -> Redis -> DB
  • Read Path: Cache -> DB
  • Async: Kafka for Analytics
  • Sharding: Hash(short_key)

3. Data Models

  • URLs Table: `id`, `short_key`, `long_url`
  • Redis: `unused_keys` (Set), `url:{key}` (String)
  • Analytics: Columnar DB (ClickHouse)

4. Bottlenecks & Fixes

  • Collision: Use Pre-generated Keys (KGS).
  • Latency: Cache 20% hot URLs (Pareto).
  • Analytics Lag: Use Kafka (Async).
  • DB Space: TTL or cleanup job.