System Design
Learn System Design
Introduction to System Design
How to Learn System Design?
Key Characteristics of Distributed Systems
Scalability
Availability
Latency and Performance
Concurrency and Coordination
Monitoring and Observability
Resilience and Error Handling
Fault Tolerance vs. High Availability
Flashcards Review
Chapter Assessment
Network Essentials
HTTP vs. HTTPS
TCP vs. UDP
HTTP: 1.0 vs. 1.1 vs 2.0 vs. 3.0
URL vs. URI vs. URN
What Happens When You Type a URL into the Browser
Flashcards Review
Chapter Assessment
Long-Polling vs. WebSockets vs. Server-Sent Events
Introduction to Real-Time Communication
What is Long-Polling?
What is WebSocket?
What are Server-Sent Events?
Difference Between Long-Polling, WebSockets, and Server-Sent Events
Flashcards Review
Chapter Assessment
Domain Name System (DNS)
Introduction to DNS
DNS Resolution Process
DNS Load Balancing and High Availability
Flashcards Review
Chapter Assessment
Proxies
What is a Proxy Server?
Uses of Proxies
VPN vs. Proxy Server
Flashcards Review
Chapter Assessment
Load Balancing
Introduction to Load Balancing
Load Balancing Algorithms
Uses of Load Balancing
Load Balancer Types
Stateless vs. Stateful Load Balancing
High Availability and Fault Tolerance
Scalability and Performance
Challenges of Load Balancers
Flashcards Review
Chapter Assessment
API Gateway
Introduction to API Gateway
Usage of API gateway
Advantages and disadvantages of using API gateway
Flashcards Review
Chapter Assessment
API Design
What Is an API?
What Are REST APIs?
Resources, Not Actions
HTTP Methods and Their Semantics
URL Design
Request and Response Shapes
Status Codes and Error Design
Pagination from the Consumer's View
Idempotency Keys
Versioning and Backward Compatibility
Concurrency and Conditional Requests
REST vs gRPC vs GraphQL
Flashcards Review
Chapter Assessment
Rate Limiting and Throttling
What Is Rate Limiting
Rate Limiting Algorithms
Distributed Rate Limiting
Rate Limiting in Practice
Flashcards Review
Chapter Assessment
Caching
Introduction to Caching
Why is Caching Important?
Types of Caching
Cache Replacement Policies
Cache Invalidation
Cache Read Strategies
Cache Coherence and Consistency Models
Caching Challenges
Cache Performance Metrics
Flashcards Review
Chapter Assessment
CDN
What is CDN?
Origin Server vs. Edge Server
CDN Architecture
Push CDN vs. Pull CDN
Flashcards Review
Chapter Assessment
Data Partitioning
Introduction to Data Partitioning
Partitioning Methods
Data Sharding Techniques
Benefits of Data Partitioning
Common Problems Associated with Data Partitioning
Flashcards Review
Chapter Assessment
Redundancy and Replication
What is Redundancy?
What is Replication?
Replication Methods
Data Backup vs. Disaster Recovery
Flashcards Review
Chapter Assessment
CAP & PACELC Theorems
Introduction to CAP Theorem
Components of CAP Theorem
Trade-offs in CAP Theorem
Examples of CAP Theorem in Practice
Beyond CAP Theorem
System Design Trade-offs in Interviews
Flashcards Review
Chapter Assessment
Databases (SQL vs. NoSQL)
Introduction to Databases
SQL Databases
NoSQL Databases
SQL vs. NoSQL
ACID vs BASE Properties
Real-World Examples and Case Studies
SQL Normalization and Denormalization
In-Memory Database vs. On-Disk Database
Data Replication vs. Data Mirroring
Database Federation
Flashcards Review
Chapter Assessment
Indexes
What are Indexes?
How a B-Tree Index Works
Types of Indexes
B-Tree vs. LSM Tree
Indexes in Distributed Systems
Flashcards Review
Chapter Assessment
Bloom Filters
Introduction to Bloom Filters
Benefits & Limitations of Bloom Filters
Variants and Extensions of Bloom Filters
Applications of Bloom Filters
Flashcards Review
Chapter Assessment
Quorum
Why Quorum?
What is Quorum?
Flashcards Review
Chapter Assessment
Leader and Follower
What is Leader and Follower Pattern?
Flashcards Review
Chapter Assessment
Heartbeat
What is Heartbeat?
Flashcards Review
Chapter Assessment
Checksum
What is Checksum?
Uses of Checksum
Flashcards Review
Chapter Assessment
Distributed Messaging System
Introduction to Messaging System
Introduction to Kafka
Messaging patterns
Popular Messaging Queue Systems
RabbitMQ vs. Kafka vs. ActiveMQ
Scalability and Performance
Flashcards Review
Chapter Assessment
Distributed File Systems
What is a Distributed File System?
Architecture of a Distributed File System
Key Components of a DFS
Flashcards Review
Chapter Assessment
Security
What is Security and Privacy?
What is Authentication?
What is Authorization?
Authentication vs. Authorization
OAuth vs. JWT for Authentication
What is Encryption?
What are DDoS Attacks?
Flashcards Review
Chapter Assessment
Misc Concepts
Batch Processing vs. Stream Processing
XML vs. JSON
Synchronous vs. Asynchronous Communication
Push vs. Pull Notification Systems
Microservices vs. Serverless Architecture
Message Queues vs. Service Bus
Stateful vs. Stateless Architecture
Event-Driven vs. Polling Architecture
Flashcards Review
Chapter Assessment
Quiz - System Design Fundamentals
Quiz
System Design Trade-offs
Importance of Discussing Trade-offs
Strong vs Eventual Consistency
Latency vs Throughput
ACID vs BASE Properties in Databases
Read-Through vs Write-Through Cache
Batch Processing vs Stream Processing
Load Balancer vs. API Gateway
API Gateway vs Direct Service Exposure
Proxy vs. Reverse Proxy
API Gateway vs. Reverse Proxy
SQL vs. NoSQL
Primary-Replica vs Peer-to-Peer Replication
Data Compression vs Data Deduplication
Server-Side Caching vs Client-Side Caching
REST vs RPC
Polling vs. Long-Polling vs. WebSockets vs. Webhooks
CDN Usage vs Direct Server Serving
Serverless Architecture vs Traditional Server-based
Stateful vs Stateless Architecture
Hybrid Cloud Storage vs All-Cloud Storage
Token Bucket vs Leaky Bucket
Read Heavy vs Write Heavy System
Quiz
How to Approach a System Design Interview
System Design Interviews - A step by step guide
Functional vs. Non-functional Requirements
What are Back-of-the-Envelope Estimations?
Things to Avoid During System Design Interview
System Design Master Template
System Design Master Template
Quiz
Designing Pastebin
Designing Pastebin
Quiz - Designing Pastebin
Designing Instagram
Designing Instagram
Quiz - Designing Instagram
Designing Dropbox
Designing Dropbox
Quiz - Designing Dropbox
Designing Facebook Messenger
Designing Facebook Messenger
Quiz - Designing Facebook Messenger
Designing Twitter
Designing Twitter
Quiz - Designing Twitter
Designing Youtube or Netflix
Designing Youtube or Netflix
Quiz - Designing Youtube
Designing Typeahead Suggestion
Designing Typeahead Suggestion
Quiz - Designing Typeahead Suggestion
Designing an API Rate Limiter
Designing an API Rate Limiter
Quiz - Designing an API Rate Limiter
Designing Twitter Search
Designing Twitter Search
Quiz - Designing Twitter Search
Designing a Web Crawler
Designing a Web Crawler
Quiz - Designing a Web Crawler
Designing Facebook’s Newsfeed
Designing Facebook’s Newsfeed
Quiz - Designing Facebook’s Newsfeed
Designing Yelp or Nearby Friends
Designing Yelp or Nearby Friends
Quiz - Designing Yelp or Nearby Friends
Designing Uber backend
Designing Uber backend
Quiz - Designing Uber backend
Designing Ticketmaster
Designing Ticketmaster
Quiz - Designing Ticketmaster
Dynamo: How to design a key value store?
Dynamo: Introduction
High-Level Architecture
Data Partitioning
Replication
Vector Clocks and Conflicting Data
The Life of Dynamo’s put() & get() Operations
Anti-entropy Through Merkle Trees
Gossip Protocol
Dynamo Characteristics and Criticism
Summary: Dynamo
Quiz: Dynamo
Mock Interview: Dynamo
Designing YouTube Likes Counter (medium)
YouTube Likes Counter
Quiz
Cassandra: How to Design a Wide-column NoSQL Database?
Cassandra: Introduction
High-level Architecture
Replication
Cassandra Consistency Levels
Gossiper
Anatomy of Cassandra's Write Operation
Anatomy of Cassandra's Read Operation
Compaction
Tombstones
Summary: Cassandra
Quiz: Cassandra
Mock Interview: Cassandra
Kafka: How to Design a Distributed Messaging System?
Messaging Systems: Introduction
Kafka: Introduction
High-level Architecture
Kafka: Deep Dive
Consumer Groups
Kafka Workflow
Role of ZooKeeper
Controller Broker
Kafka Delivery Semantics
Kafka Characteristics
Summary: Kafka
Quiz: Kafka
Mock Interview: Kafka
Chubby: How to Design a Distributed Locking Service?
Chubby: Introduction
High-level Architecture
Design Rationale
How Chubby Works
File, Directories, and Handles
Locks, Sequencers, and Lock-delays
Sessions and Events
Master Election and Chubby Events
Caching
Database
Scaling Chubby
Summary: Chubby
Quiz: Chubby
Mock Interview: Chubby
HDFS: How to Design File Storage System?
Hadoop Distributed File System: Introduction
High-level Architecture
Deep Dive
Anatomy of a Read Operation
Anatomy of a Write Operation
Data Integrity & Caching
Fault Tolerance
HDFS High Availability (HA)
HDFS Characteristics
Summary: HDFS
Quiz: HDFS
Mock Interview: HDFS
GFS: How to Design a Distributed File System Storage?
Google File System: Introduction
High-level Architecture
Single Master and Large Chunk Size
Metadata
Master Operations
Anatomy of a Read Operation
Anatomy of a Write Operation
Anatomy of an Append Operation
GFS Consistency Model and Snapshotting
Fault Tolerance, High Availability, and Data Integrity
Garbage Collection
Criticism on GFS
Summary: GFS
Quiz: GFS
Mock Interview: GFS
BigTable: How to Design a Wide Column Storage System?
BigTable: Introduction
BigTable Data Model
System APIs
Partitioning and High-level Architecture
SSTable
GFS and Chubby
Bigtable Components
Working with Tablets
The Life of BigTable's Read & Write Operations
Fault Tolerance and Compaction
BigTable Refinements
BigTable Characteristics
Summary: BigTable
Quiz: BigTable
Mock Interview: BigTable
Designing Reddit (medium)
Design Reddit
Quiz
Designing Notification Service (medium)
Designing a Notification System
Quiz
Design Google Calendar (medium)
Design Google calendar (Medium)
Quiz
Design a Recommendation System (medium)
Design a Recommendation System for Netflix
Quiz
Designing Gmail (medium)
Design Gmail
Quiz
Designing Google News (medium)
Design Google News, a Global News Aggregator System (Medium)
Quiz
Designing Unique ID Generator (medium)
Design Unique ID Generator (Easy)
Quiz
Designing Code Judging System (medium)
Design Code Judging System like LeetCode (Medium)
Quiz
Designing Payment System (hard)
Design Payment System
Quiz
Designing Flash Sale System (hard)
Design a Flash Sale for an E-commerce Site (Hard)
Quiz
Designing Reminder Alert System (hard)
Design a Reminder Alert System
Quiz
System Design Patterns
Introduction: System Design Patterns
1. Bloom Filters
2. Consistent Hashing
3. Quorum
4. Leader and Follower
5. Write-ahead Log
6. Segmented Log
7. High-Water Mark
8. Lease
9. Heartbeat
10. Gossip Protocol
11. Phi Accrual Failure Detection
12. Split Brain
13. Fencing
14. Checksum
15. Vector Clocks
16. CAP Theorem
17. PACELC Theorem
18. Hinted Handoff
19. Read Repair
20. Merkle Trees
Quiz
Designing a URL Shortening Service like TinyURL
url shortening
distributed systems
api design
caching
+3
This case study follows the seven-step interview method. Use the System Design Master Template while you work through it.
Imagine that you want to share this address:
https://www.designgurus.io/course/grokking-the-system-design-interview
The address works, but it is long. It takes space in messages, printed material, and social media posts.
A URL shortening service gives the address a smaller alias:
https://tinyurl.com/vzet59pa
When someone opens the short link, the service sends them to the original address.
The service has two basic jobs:
- Turn a long URL into a short link.
- Redirect a short link to its original URL.
The first job creates data. The second job reads that data.
These jobs look simple, but they run at very different scales. Our design must handle that difference.
Short links do more than save space. Every click first reaches the shortening service.
This lets the service count clicks. Link owners can measure campaigns, compare referrers, and learn where their audience is located.
Analytics is therefore a real requirement. It will affect the redirect response and the final architecture.
Short links can also hide affiliate URLs. They may send different devices to different pages.
Try creating a link on TinyURL before you continue. Notice the choices the product gives you.
Try it yourself
Design the system before reading the solution. Use this brief so your design starts with the same requirements.
Core requirements
- Given a long URL, return a short link.
- Following a short link redirects the user to the original URL.
- Users can choose a custom alias.
- Links expire after a default period. Users can also choose an expiry time.
- Link owners can view click statistics.
Extended requirement
A link can be private. Only permitted users may open it.
Write down four things: a load estimate, an API, a data model, and a first architecture diagram.
Do not worry about matching the solution. A different choice is useful when you can explain its trade-offs.
Sketch it here
Designing a URL Shortener (video)
Now compare your first design with the video. Look for decisions that follow from a specific requirement.
Step 1: Clarify the Requirements
💡 A design is only correct for a specific set of requirements. Confirm the scope before drawing the architecture.
Suppose the interviewer says, "Design TinyURL." Do not start with servers or databases.
First, ask what the service must do. Then ask which qualities matter most.
Functional requirements describe what users can do:
- Given a long URL, the service returns a short link.
- Following a short link redirects the user to the original URL.
- Users can choose a custom alias.
- Links expire after a default period. Users can also choose an expiry time.
- Link owners can view click statistics.
Analytics belongs in the main list. It changes two later decisions.
First, we must observe every click. Second, click processing must not slow down redirects.
Extended requirement
A link can be private. Only permitted users may open it.
Non-functional requirements describe how the service should behave:
- High availability. Existing links should keep working when parts of the system fail.
- Low redirect latency. A redirect adds an extra network request, so it must be fast.
- Hard-to-guess links. Random guesses should rarely find another user's link.
We will not design billing, paid plans, link previews, or destination scanning. These features are outside the agreed scope.
The requirements now give us tests for every later choice. For example, a predictable key would fail the third non-functional requirement.
Step 2: Estimate the Scale
💡 Estimate only the numbers that may change the design. Round the inputs so the arithmetic stays easy to check.
We need two starting assumptions:
- How many new links are created each month?
- How many redirects happen for each new link?
Assume 500 million new links per month and 100 redirects for every write.
Every estimate below follows from these two numbers.
Request load
A day has about 100,000 seconds. This rounded value makes the request-rate calculation simple.
500M links / 30 days = about 17M writes per day
17M / 100,000 seconds = about 170 writes per second
Rounded write rate = 200 writes per second
200 writes x 100 redirects = 20,000 reads per second
20,000 reads x 3 peak factor = 60,000 peak reads per second
The service is read-heavy. Redirects happen about one hundred times more often than link creation.
Storage
Assume that we keep each link for five years. Also assume that one stored record uses about 500 bytes.
500M x 12 months x 5 years = 30 billion links
30 billion x 500 bytes = about 15 TB
Fifteen terabytes will not fit on one database machine. We will need to divide the data across machines.
Bandwidth
Each request moves a small amount of text.
200 writes x 500 bytes = about 100 KB per second in
20,000 reads x 500 bytes = about 10 MB per second out
Ten megabytes per second is modest. Bandwidth does not require a special design here.
Cache size
Popular links receive most of the traffic. Assume that 20 percent of requested links produce 80 percent of redirects.
20,000 x 100,000 seconds = about 2 billion reads per day
20% x 2 billion x 500 bytes = about 200 GB
This is an upper bound. Many requests open the same popular links, so the real hot data set will be smaller.
What the estimates tell us
Three results affect the architecture:
- Thirty billion rows require partitioning. One machine cannot store all the data.
- The 100-to-1 ratio makes caching valuable. A cache can remove most reads from the database.
- Bandwidth is not the limit. The system moves URLs and small records, not large files.
The purpose of estimation is not precision. It is to discover which parts of the system need special treatment.
Step 3: Define the API
💡 The API turns requirements into concrete operations. Missing fields or responses often reveal a missing design decision.
We need four main operations: create a link, follow it, read its statistics, and delete it.
Create a short link
POST /urls
Authorization: Bearer <token> optional for public links
Idempotency-Key: <unique value>
body: {
original_url,
custom_alias?,
expires_at?,
visibility?
}
201 Created: {
short_url,
short_key,
expires_at,
visibility
}
An idempotency key identifies one creation request. Retrying the same request must not create a second short link.
The client does not send a trusted user_id. The gateway validates the token and derives the user identity from it.
If a custom alias already exists, return 409 Conflict. If the input is invalid, return 400 Bad Request.
Follow a short link
GET /{short_key}
302 Found: Location: <original_url>
401 Unauthorized: sign-in is required for this private link
403 Forbidden: the signed-in user cannot open this link
404 Not Found: the key does not exist
410 Gone: the link has expired
Use 302 Found rather than 301 Moved Permanently. Browsers may cache a 301 and stop sending later clicks to our service.
A 302 keeps every click visible. This preserves the analytics requirement from Step 1.
Read click statistics
GET /urls/{short_key}/analytics?from=<time>&to=<time>
Authorization: Bearer <token>
200 OK: {
total_clicks,
clicks_by_day,
top_referrers,
top_countries
}
Only the link owner may read these statistics. The counts may lag by a short time because analytics can be eventually consistent.
Delete a link
DELETE /urls/{short_key}
Authorization: Bearer <token>
204 No Content
Only the owner may delete a link. A later redirect to that key must not return the old destination.
Limits
Rate-limit creation per account. Rate-limit anonymous creation and redirects per IP address.
These limits stop one caller from consuming the key pool. They also make large-scale key guessing harder.
The API now covers every functional requirement. We can use it to decide what data the system must store.
Step 4: Define the Data Model
💡 Start with the reads and writes. They tell you which fields and indexes the system needs.
The main lookup is simple. A redirect receives a short key and needs the original URL.
Each URL record stores:
short_key: the primary key used by redirects.original_url: the destination returned in the redirect.user_id: the owner of the link.created_at: when the link was created.expires_at: when the link stops working.visibility: whether the link is public or private.
Private links also need permission data. Custom aliases use the same short_key field as generated keys.
Click events do not belong in this record. Updating a click counter on every redirect would make one popular row a write bottleneck.
Choose the database from the access pattern
The database must store billions of small records. Most requests read one record by its primary key.
A distributed key-value or wide-column database fits this pattern. DynamoDB and Cassandra are common examples.
These databases divide records across many machines. They also support fast primary-key lookups without joins.
The user_id field does not require a relational database. The redirect path never joins the URL record with a user table.
The service needs one secondary access pattern: list the links owned by one user. An index on user_id can support that query.
Store click events separately. Analytics needs large scans and grouped counts, while redirects need single-key lookups.
One database does not need to serve both patterns.
Step 5: Draw the High-Level Design
Start with the two user actions. One action creates a link, and the other follows it.
These actions form separate paths through the system.
The write path
The write path handles about 200 requests per second.
- A request enters through the API gateway.
- The shortening service obtains an unused short key.
- The service inserts one URL record.
- The service returns the complete short URL.
The write rate is modest. The hard problem is giving every write a unique key.
The read path
The read path handles about 20,000 requests per second.
- A request enters through the API gateway.
- The redirect service looks for the key in the cache.
- On a cache miss, it reads the URL database.
- It returns a redirect to the original URL.
Reads outnumber writes by one hundred to one. The cache and read replicas therefore belong on the read path.
The two paths may begin as one application. We can deploy them separately when their scaling and availability needs become different.
Step 6: Go Deep
Choose the key length
Each generated key uses characters from a fixed alphabet. A larger alphabet gives more keys at the same length.
Suppose we use the 62 characters in A-Z, a-z, and 0-9.
Six characters provide about 56.8 billion possible keys. Our estimates require 30 billion stored links.
30 billion / 56.8 billion = about 53 percent
More than half of the six-character keyspace would be occupied. A random guess would often find a real link.
That result fails the hard-to-guess requirement from Step 1.
Seven characters provide about 3.5 trillion keys.
30 billion / 3.5 trillion = about 0.9 percent
Only about one key in 117 would resolve to a stored link. Rate limiting makes repeated guessing much harder.
Seven characters is therefore the shortest acceptable length for this design.
Choose the alphabet
Use base62 for generated keys. It contains letters and digits only.
Standard base64 also contains + and /. These characters have special meanings in URLs and often require encoding.
The slash can create another path segment. Some systems also interpret a plus as a space.
Base62 avoids both problems. URL-safe base64 with - and _ is another valid choice.
Choose how to generate keys
We now need seven-character keys that are unique and hard to predict.
There are two common approaches. We can hash each URL, or we can generate keys before requests need them.
Option 1: Hash the URL
Run a hash function like SHA-256 over the long URL. Encode the result and keep the first seven characters.
This approach has two problems.
First, the same URL always produces the same key. Two owners may need different expiry times, permissions, or analytics.
Adding the owner ID or a sequence number to the input solves this first problem.
Second, truncating the hash creates collisions. Different inputs can produce the same seven-character result.
Every write must therefore follow this loop:
- Generate a candidate key.
- Check whether the key already exists.
- Generate another key after a collision.
- Repeat until the key is free.
This loop runs while the user waits. It also becomes slower as more keys are used.
Option 2: Pre-generate keys
A Key Generation Service, or KGS, creates random keys before requests arrive.
The service checks each key once and places every unused key in a pool. A shortening server takes a verified key from that pool.
The uniqueness check still exists. It now runs in the background instead of inside a user request.
This is the main reason to use KGS.
Size the key pool
The pool only needs to stay ahead of new requests. It does not need to contain every possible key.
At 17 million new links per day, one billion keys provide about two months of supply.
1 billion keys x 7 bytes = about 7 GB of raw key data
Indexes and metadata will increase the stored size. The total is still small for a background service.
The generator keeps refilling the pool. Since only 0.9 percent of keys are used, most random candidates are free.
Give keys to shortening servers
KGS hands out blocks instead of one key at a time.
Suppose one block contains 1,000 keys. KGS marks the whole block as used before returning it.
The shortening server stores the block in memory. It can now create 1,000 links without another KGS request.
Two servers cannot receive the same block because the state changes before the block leaves KGS.
If a server fails, its remaining keys are lost. We do not return them to the pool.
Losing a few keys is safer than coordinating every key assignment. The complete keyspace contains trillions of values.
Keep KGS available
KGS can fail, so run a standby instance. The standby takes control when the primary instance stops responding.
Shortening servers also keep local key blocks. They can continue creating links during a short KGS outage.
If every local block becomes empty, new link creation pauses. Existing redirects still work.
Handle custom aliases
A custom alias cannot come from the prepared pool. The user chooses its value.
Try to insert the alias as a unique key. Return 409 Conflict when another record already uses it.
Limit custom aliases to 16 characters. The short_key field can therefore use a maximum length of 16 instead of seven.
Follow a short link
Redirects form almost all traffic. Their path must stay short and available.
Process a redirect
The redirect service follows these steps:
- Look for the short key in the cache.
- Read the database after a cache miss.
- Check whether the record exists.
- Check its expiry time and permissions.
- Return the redirect.
- Publish a click event without waiting for analytics.
Each failed check has a different response:
- Return 404 Not Found when the key never existed.
- Return 410 Gone when the link has expired.
- Return 401 Unauthorized when a private link requires sign-in.
- Return 403 Forbidden when the signed-in user lacks permission.
Otherwise, return 302 Found with the destination in the Location header.
Choose between 301 and 302
Both responses redirect the browser. Their caching behavior is different.
A 301 says the redirect is permanent. Browsers may cache it and stop contacting our service.
This reduces traffic, but later clicks disappear from our analytics.
A 302 keeps the redirect temporary. Each click still reaches the redirect service.
Use 302 because click statistics are a requirement. Without per-click analytics, 301 would reduce more traffic.
Cache popular links
A cache stores frequently used records in memory. Memory reads are much faster than database reads.
Each cache entry contains the original URL, expiry time, and visibility data. The redirect service can complete every check from this entry.
On a cache miss, read the database. Then place the returned record in the cache.
URL records are easy to cache because their destinations never change. A delete operation removes the cache entry, and every hit checks the expiry time.
Begin with the 200 GB estimate from Step 2. Adjust the size after measuring the real cache hit rate.
Use least recently used, or LRU, eviction. LRU removes the entry that has gone unused for the longest time.
Cache missing and expired keys for a short time too. This is called negative caching.
Negative caching stops repeated requests for dead keys from reaching the database.
Partition the URL data
Thirty billion rows need many database machines. Partitioning divides the rows among those machines.
Hash the short key and map the result to a partition. Random keys will spread evenly across the partitions.
Use consistent hashing to reduce data movement when machines change. Only a small part of the key range moves after adding a machine.
Range partitioning is less useful here. Redirects never request an ordered range of short keys.
Custom aliases can also make first-character ranges uneven. Hash partitioning avoids this imbalance.
Replicate each partition
Keep at least two copies of every partition on other machines.
Replicas protect the data after a machine failure. They can also serve reads and reduce load on the primary copy.
Managed databases like DynamoDB and Cassandra perform partitioning and replication for the application.
Remove expired links
Checking the whole database for expired rows would create constant work. Use lazy cleanup instead.
- When a redirect finds an expired link, return
410 Goneand schedule its removal. - Run a low-priority background job to remove expired links that receive no traffic.
Some expired rows may remain stored for a while. The redirect service still refuses to serve them.
Never reuse an expired short key. Old emails, bookmarks, and printed links may still contain it.
Reusing that key could send an old link to a new owner's destination. The large keyspace makes reuse unnecessary.
Record analytics outside the redirect path
Click processing must not increase redirect latency. Use an asynchronous event pipeline.
The redirect service returns 302 Found first. It then publishes a click event to a queue like Kafka.
Each event may contain:
- the short key,
- the event time,
- the referrer,
- the user's country,
- and the user agent.
Analytics consumers read events in batches. They write grouped counts into a store designed for analytics queries.
The event stream is manageable:
20,000 events per second x 200 bytes = about 4 MB per second
The queue separates redirects from analytics. A slow analytics consumer creates a backlog, but redirects continue working.
Click counts become eventually consistent. They may appear seconds or minutes after the real click.
This delay is acceptable for reports. Fast redirects are more important than immediate counters.
Control abuse
Abuse means using the service beyond its allowed limits.
Rate-limit link creation per account and per IP address. This stops one script from filling the database or exhausting prepared keys.
Rate-limit redirects per IP address too. This makes repeated guessing slower.
The API gateway applies these limits before requests reach the services.
Check access to private links
Access control decides whether one user may open a specific link.
Store the visibility flag and permission data with the URL record. The redirect service can then check access without calling another service.
Keep abuse controls and access control separate. They solve different problems.
Step 7: Find Bottlenecks and Failure Points
A complete design also explains how it behaves when a part becomes slow or unavailable.
Cache failure
A large cache failure sends many requests to the database. This sudden load can overwhelm the database.
Partition and replicate the cache. Restore traffic gradually while failed cache nodes recover.
Database failure
A failed database machine should not remove a partition. Promote a healthy replica and rebuild the missing copy.
Redirects may read from replicas. New writes may pause briefly while the database selects a new primary.
KGS failure
The standby KGS replaces a failed primary. Shortening servers continue using their local blocks during the change.
If every block becomes empty, only new link creation stops. Redirects do not depend on KGS.
Analytics backlog
A slow consumer increases the number of events waiting in the queue. Click reports become older, but redirects remain fast.
Add consumers when the backlog grows. The queue keeps the events until consumers process them.
Keep the redirect path independent
The redirect path depends only on the cache and URL database.
KGS, analytics, cleanup, and destination screening can all fail without stopping existing links.
This separation is the most important availability property in the design.
Putting It Together
Read the final diagram as three flows.
Create flow
- The API gateway checks authentication and rate limits.
- The shortening service takes a key from its local block.
- The service writes the URL record to the database.
Redirect flow
- The redirect service checks the cache.
- It reads the database only after a cache miss.
- It checks expiry and access rules.
- It returns
302 Found.
Background flow
The redirect service publishes click events to the queue. Analytics consumers process them later.
KGS refills key blocks, and the cleanup job removes expired rows. Neither task sits inside the redirect request.
Where AI Fits in This Design
The core URL shortener does not need AI. Key generation, caching, and redirects follow fixed rules.
AI can help with destination screening, which is outside the core scope. Short links hide their final destinations and can be used for phishing.
A background worker can send each destination to a classifier. A classifier is a model that labels content, such as safe or suspicious.
Run this check after link creation. Do not place the model inside the redirect path.
The worker stores the result as a flag on the URL record. The redirect service already reads flags for expiry and visibility.
Model errors need a safe response. A false positive marks a safe page as suspicious.
Show a warning page or request human review after a positive result. Do not silently delete the link.
💡 In the interview: spend most of your time on key generation and redirects. Explain why six characters fail the stated security requirement. Then explain why KGS moves collision checks out of user requests. If asked about redirects, connect
302directly to click analytics.
Key takeaway: this system has about one hundred reads for every write. Use seven-character base62 keys because six characters are too easy to guess at this scale. Pre-generate keys and give them to servers in blocks. Cache popular URL records before a partitioned and replicated database. Return
302so every click reaches the service. Send click events through a queue, so analytics never delays a redirect. Keep background services outside the redirect path, so existing links remain available during other failures.
Discussion
On This Page