CAP Theorem with Tom the Prankster

Link: https://newsletter.systemdesigncodex.com/p/cap-theorem

1. Essential Elements: Consistency, Availability, Partition Tolerance.
2. Partition Tolerance: Must-have due to inherent unreliability in communication networks.
3. Choice between Consistency and Availability:
   • Consistency: All nodes see the same data at the same time. Requires a trade-off with availability.
   • Availability: Ensures that the system is always operational, but might compromise on having the latest data across all nodes.

The Inevitable Law Governing Software Design

Link: https://newsletter.systemdesigncodex.com/p/the-inevitable-law-governing-software-design

• Basic Principle: The structure of a software system reflects the communication structure of its creating organization.
• Example: In a company with inventory, invoicing, and shipping departments, the software will likely have separate systems for each, mirroring these divisions.
• Implications:
  • Software integration quality depends on how well these departments communicate.
  • Better communication leads to more effective and integrated software modules.
• Strategies for Addressing Conway’s Law:
  1. Acknowledge It: Recognize its impact on software design.
  2. Structure Teams Effectively: Place teams working on similar systems close to each other for better communication.
  3. Avoid Dividing by Technology: Instead of splitting teams by tech layers (front end, back end), focus on business features for smoother collaboration.
  4. Use Architectural Insights: Align team structures with desired software architecture, understanding that organizational decisions influence software design.

The Ingredients to Delicious Software

Link: https://newsletter.systemdesigncodex.com/p/the-ingredients-to-delicious-software

1. Scalability:

   • Ability to handle increased workload efficiently.
   • Important to identify the point where scaling becomes cost-ineffective.

2. Latency & Throughput:

   • Latency: Time taken to respond to a request (e.g., time to serve a cheese sandwich).
   • Throughput: Number of requests handled in a given time (e.g., serving multiple customers).

3. Availability and Consistency:

   • Availability: Ability to operate despite issues (e.g., with one cook absent).
   • Measured in 'nines' (e.g., 99.9% availability).
   • Consistency: Synchronization of information across different parts of the system (e.g., order copies being in sync).

What happens when you type a URL into your browser?

Link: https://blog.bytebytego.com/p/what-happens-when-you-type-a-url

When you type a URL into your browser:

1. URL Parsing: The browser identifies the HTTP protocol, domain, path, and resource.
2. DNS Lookup: It searches for the IP address of the domain, checking various caches.
3. TCP Connection: Establishes a connection with the server.
4. HTTP Request: Sends a request for the specific resource.
5. Server Response: The server sends back the requested content.
6. Rendering: The browser displays the webpage.

How does CDN work?

Link: https://blog.bytebytego.com/p/how-does-cdn-work

1. Domain Name Lookup:

   • Bob enters www.myshop.com in his browser.
   • The browser checks the local DNS cache for the domain.

2. DNS Resolver:

   • If not in the local cache, the browser contacts the DNS resolver (usually via the ISP).

3. Recursive Domain Resolution:

   • The DNS resolver performs recursive resolution for www.myshop.com.

4. CDN Integration:

   • Instead of pointing directly to the London server, the authoritative name server redirects to a CDN domain (www.myshop.cdn.com).

5. Load Balancer Query:

   • The DNS resolver queries the CDN load balancer domain (www.myshop.lb.com).

6. Optimal Server Selection:

   • The CDN load balancer selects the best CDN edge server based on factors like the user’s location and server load.

7. Content Delivery:

   • The browser connects to the chosen CDN edge server to load content.
   • The content includes static (e.g., images, videos) and dynamic elements.
   • If content is not on the edge server, it's fetched from higher-level CDN servers or the origin server in London.

8. CDN Network:

   • This process is part of a geographically distributed CDN network for efficient content delivery.

Time complexity

Link: https://newsletter.francofernando.com/p/time-complexity

1. Purpose: Time complexity evaluates how an algorithm's performance scales with the size of the input data.

2. Types of Complexity:

   • Worst-Case Complexity: Maximum number of steps for any input of size n. Most commonly used as it provides guarantees about the algorithm's upper limit.
   • Best-Case Complexity: Minimum number of steps for any input of size n.
   • Average-Case Complexity: Average number of steps over all possible instances of input size n.

3. Big Oh Notation: Simplifies the expression of an algorithm's worst-case complexity by focusing on growth rates rather than precise step counts.

4. Common Complexity Classes:

   • Constant - O(1): Time is independent of input size (e.g., adding two numbers).
   • Logarithmic - O(log n): Each step cuts the problem size in half (e.g., binary search).
   • Linear - O(n): Time grows linearly with input size (e.g., finding max in an array).
   • Superlinear - O(n log n): Combines linear and logarithmic growth (e.g., Quicksort, Mergesort).
   • Quadratic - O(n^2): Time grows with the square of input size (e.g., insertion sort).
   • Cubic - O(n^3): Involves triple nested loops (e.g., certain dynamic programming algorithms).
   • Exponential - O(c^n): Time doubles with each addition to input size (e.g., enumerating subsets).
   • Factorial - O(n!): Time grows with the factorial of input size (e.g., generating permutations).

8 Reasons Why WhatsApp Was Able to Support 50 Billion Messages a Day With Only 32 Engineers

Link: https://newsletter.systemdesign.one/p/whatsapp-engineering

1. Single Responsibility Principle:

   • Focus on core feature: Messaging.
   • Avoided feature creep and unnecessary functionalities.
   • Prioritized reliability above all.

2. Technology Stack:

   • Chose Erlang for server functionalities due to its scalability and support for hot-loading.
   • Erlang's efficient threading and context-switching mechanisms contributed to performance.

3. Utilizing Existing Solutions:

   • Leveraged open-source solutions like Ejabberd, an Erlang-based messaging server.
   • Customized existing solutions to fit specific needs.
   • Integrated third-party services for functionalities like push notifications.

4. Cross-Cutting Concerns:

   • Emphasized aspects like monitoring and alerting for service health.
   • Implemented Continuous Integration and Continuous Delivery for software development.

5. Scalability Strategies:

   • Adopted diagonal scaling, combining horizontal and vertical scaling methods.
   • Ran servers on FreeBSD, optimized for handling millions of connections.
   • Overprovisioned servers for handling traffic spikes and potential failures.

6. Continuous Improvement (Flywheel Effect):

   • Regularly measured performance metrics to identify and eliminate bottlenecks.
   • Maintained a cycle of continuous feedback and improvement.

7. Focus on Quality:

   • Conducted load testing to identify and address single points of failure.
   • Used simulated production traffic for realistic testing.

8. Small Team Size:

   • Kept the engineering team small (32 engineers) to maintain efficiency and reduce communication overhead.

This Is How Quora Shards MySQL to Handle 13+ Terabytes

Link: https://newsletter.systemdesign.one/p/mysql-sharding

1. Vertical Sharding:

   • Implementation: Separating tables into different servers (leader-follower model).
   • Purpose: Enhances write scalability.
   • Challenges: Replication lag, transactional limitations, and potential performance issues for large tables.

2. Horizontal Sharding:

   • Reasons for Adoption: Addressing challenges with large tables such as schema changes and error risks.
   • Approach: Splitting a logical table into multiple physical tables.

3. Key Decisions in Horizontal Sharding:

   • Build vs. Buy: Opted to build their own sharding solution, reusing vertical sharding logic.
   • Shard Level: Focused on sharding at the table level due to extensive use of secondary indexes.
   • Sharding Method: Chose range-based partitioning, favoring common range queries.
   • Metadata Management: Stored shard metadata in Apache Zookeeper.
   • Database API: Modified to handle sharding columns and keys, enhancing security against SQL injections.
   • Sharding Column Selection: Based on latency sensitivity and query per second (QPS) considerations.
   • Cross-Shard Indexes: Used to optimize non-sharding column queries, though with potential performance and consistency trade-offs.
   • Number of Shards: Maintained a lower count to reduce latency in non-sharding column queries.

Making Your Database Highly Available

Link: https://newsletter.systemdesigncodex.com/p/making-your-database-highly-available

Redundancy

• Purpose: Ensure continuous database operation even if one server fails.
• Not Backup: Unlike backups, redundancy involves running multiple active database instances.
• Cost of Outage: Can be significantly high, averaging $7,900 per minute.
• Redundancy Patterns:
  • Active-Passive: One active server handles requests while others stand by.
  • Active-Active: Multiple servers handle requests simultaneously.
  • Multi-Active: An extension of Active-Active with more complex setups.

Isolation

• Goal: Minimize disaster impact by physically separating database components.
• Degrees of Separation:
  • Server: Different servers in the same data center.
  • Rack: Separate racks within a data center.
  • Data-Center: Multiple data centers.
  • Availability Zone: Distinct zones within a cloud provider's network.
  • Region: Geographically dispersed locations.

How Rate Limiting Works?

Link: https://newsletter.systemdesigncodex.com/p/how-rate-limiting-works

How Rate Limiting Works:

1. Concept: Limits the number of requests sent to a server.
2. Implementation: A rate limiter is used to control traffic to servers or APIs.

Key Concepts

1. Limit: Maximum number of requests allowed in a set time frame (e.g., 600 requests per day).
2. Window: The duration for the limit, varying from seconds to days.
3. Identifier: A unique attribute (like User ID or IP address) to identify request senders.

Designing a Rate Limiter

• Process:
  1. Count Requests: Track the number of requests from a user or IP.
  2. Limit Exceeded: If count exceeds the limit, block or restrict further requests.
• Considerations:
  • Storage of request counters.
  • Rate limiting rules.
  • Response strategy for blocked requests.
  • Rule change implementation.
  • Maintaining application performance.

System Components

• Rate Limiter Component: Checks incoming requests against the rules and stored data (number of requests made).
• Rules Engine: Defines the rate limiting rules.
• Cache: Stores rate-limiting data for high throughput and low latency.
• Response Handling:
  • Allow request if within limit.
  • Block request if over limit, typically with HTTP status code 429.

Improvements

• Silent Drop: Fool attackers by silently dropping excess requests.
• Cached Rules: Enhance performance with a cache for the rules engine and background updates for rule changes.

Caching

Link: https://newsletter.francofernando.com/p/caching

Concept of Caching

• Purpose: Speeds up data access by storing data temporarily in a fast-access hardware or software layer.
• Cache Hit: Data is found in the cache.
• Cache Miss: Data is not in the cache and must be fetched from its original location.

Caching in Distributed Systems

• Levels: Hardware, OS, front-end, web apps, databases, etc.
• Roles:
  • Reducing latency.
  • Saving network requests.
  • Storing results of resource-intensive operations.
  • Avoiding repetitive operations.

Types of Caching

• Application Caching: Integrated into app code, checks cache before database access. Examples: Memcached, Redis.
• Database Caching: Built into databases, requires no code changes, optimizes data retrieval.

Considerations and Challenges

• Cache Miss Rate: High miss rates can add more latency.
• Stale Data: Ensuring cache data is up-to-date and relevant.

Caching Strategies

1. Cache Aside (Lazy Loading):

   • Direct read from cache. If miss, read from DB and update cache.
   • Advantages: Good for read-heavy workloads. Cache only stores necessary data.
   • Disadvantages: Can serve stale data. Initial cache misses.

2. Read Through:

   • Interact only with cache. Cache manages data fetching from DB.
   • Simplifies app code but complicates cache implementation.

3. Write Through:

   • Writes data to cache and DB simultaneously.
   • Ensures data consistency. Higher write latency.

4. Write Back (Asynchronous Writing):

   • Writes data to cache, then asynchronously to DB.
   • Lower write latency. Good for write-heavy workloads.

5. Write Around:

   • Writes directly to DB, cache only stores read data.
   • Good for infrequently read data. Higher read latency for new data.

Choosing a Cache Strategy

• Depends on data access patterns.
• Cache-Around: Good for general-purpose, read-intensive applications.
• Write-Heavy Workloads: Write-back approaches are beneficial.
• Infrequent Reads: Write-around strategy.

Eviction Policies

• Manage Limited Cache Space:
  • FIFO: First in, first out.
  • LIFO: Last in, first out.
  • LRU: Least recently used.
  • MRU: Most recently used.
  • LFU: Least frequently used.
  • RR: Random replacement.

Database Replication Under the Hood

Link: https://newsletter.systemdesigncodex.com/p/database-replication-under-the-hood

Statement-based Replication

• How It Works: The leader logs every SQL write statement (INSERT, UPDATE, DELETE) and forwards these statements to follower nodes.
• Advantages:
  • Efficient in network bandwidth, only SQL statements are transferred.
  • Portable across different database versions.
  • Simpler to implement.
• Limitations:
  • Non-deterministic functions (e.g., NOW(), UUID()) yield different values on replicas.
  • Transactions involving auto-incrementing columns must be executed in the same order.
  • Potential unforeseen effects due to triggers or stored procedures.

Shipping the Write-Ahead Log (WAL)

• Concept: The WAL, an append-only sequence of all writes, is shared with follower nodes.
• Usage: Common in databases like PostgreSQL.
• Advantage: Creates an exact replica of the leader’s data structures.
• Disadvantage: Tightly coupled to the storage engine, making it less flexible with database version changes and hindering zero-downtime upgrades.

Row-Based Replication

• Functionality: Uses a logical log showing writes in a row format.
• Operation Details:
  • Inserts log new values for all columns.
  • Deletes log identifiers for deleted rows.
  • Updates log identifiers and new values for modified columns.
• Advantage: Decouples from the storage engine, allowing backward compatibility and version flexibility between leader and follower databases.

Choosing Replication Methods

• The choice depends on the specific requirements of the system, such as:
  • Network efficiency.
  • Consistency requirements.
  • Database version compatibility.
• Statement-based Replication: Best for simple, less concurrent environments.
• WAL Shipping: Suitable for systems where exact replica and data integrity are critical.
• Row-Based Replication: Ideal for environments requiring flexibility and compatibility across different database versions.

Consistent Hashing

Link: https://newsletter.francofernando.com/p/consistent-hashing

Caching Servers

• Use Case: Store frequently accessed data in fast, in-memory caches.
• Hashing Role: Ensures identical requests are sent to the same server by hashing request attributes (IP, username, etc.).
• Challenge: Maintaining effective caching when servers are added or removed.

Data Partitioning

• Purpose: Distribute data across multiple database servers.
• Hashing Function: Data keys are hashed to determine the server where data will be stored.
• Limitation: Similar to caching, adding or removing servers complicates data distribution.

The Hashing Problem

• Goal: Map keys (data identifiers or workload requests) to servers efficiently.
• Desired Properties:
  • Balancing: Equal distribution of keys among servers.
  • Scalability: Easily adding or removing servers with minimal reconfiguration.
  • Lookup Speed: Quickly finding the server for a given key.

Naïve Hashing Approach

• Method: Number servers, use hash(key) % N to assign keys to servers.
• Drawback: Not scalable. Changing server count (N) requires remapping all keys.

Consistent Hashing

• Concept: Treat hash values as a circular space. Map keys and servers onto this circle.
• Operation: Assign each key to the nearest server on the circle in a clockwise direction.
• Advantages:
  • Only a fraction of keys need remapping when adding/removing servers.
  • Better scalability.
• Issue: Does not guarantee even key distribution (balancing).

Virtual Nodes Solution

• Strategy: Introduce replicas or virtual nodes for each server on the hash circle.
• Benefits:
  • Better balancing due to smaller ranges and more uniform key distribution.
  • Faster rebalancing when servers are added or removed.
  • Support for server fault tolerance and heterogeneity.
• Implementation: Assign more virtual nodes to more powerful servers for load balancing.

Why Replication Lag Occurs in Databases

Link: https://newsletter.systemdesigncodex.com/p/why-replication-lag-occurs-in-databases

• Concept: Replication Lag is the delay between a write operation on the leader node and its replication on follower nodes in a database system.

• Leader-based Replication Setup:

  • Writes are processed by a single node (leader).
  • Read queries can be served by any replica (follower).
  • Common in systems with more reads than writes.

• Asynchronous vs. Synchronous Replication:

  • Synchronous: All replicas must confirm write operations, causing potential unavailability if a replica is down.
  • Asynchronous: Allows distribution of reads across followers, but can lead to outdated reads if a follower lags.

• How Replication Lag Occurs:

  1. User A updates data on the leader node.
  2. Leader sends replication data to followers.
  3. User B reads from a follower (replica 2) before it's updated, receiving outdated information.
  4. Replica 2 eventually gets updated.

• Implications:

  • Lag duration varies from fractions of a second to minutes.
  • Causes temporary data inconsistencies (eventual consistency).
  • Large lags can significantly impact application performance.

• Challenge: Managing replication lag to minimize data inconsistencies and ensure efficient operation.

Problems Caused by Database Replication

Link: https://newsletter.systemdesigncodex.com/p/problems-caused-by-db-replication

1. Vanishing Updates

   • Scenario: User updates data on the leader node, but a subsequent read request to a lagging replica shows outdated data.
   • Problem: User experiences frustration as their updates appear to vanish.
   • Solution: Implement read-after-write consistency. Methods include:
     • Reading user-modified data from the leader.
     • Tracking recent writes with timestamps.
     • Monitoring and limiting queries on lagging replicas.

2. Going Backward in Time

   • Issue: User sees an update (e.g., a new comment) and then it disappears upon refreshing, due to a lagging replica.
   • User Experience: Confusion and inconsistency.
   • Solution: Ensure Monotonic Reads.
     • Users always read from the same replica.
     • Use hashing based on User ID for replica selection.

3. Violation of Causality

   • Problem: In sharded databases, replication lag causes sequence disorder in communication (e.g., a reply appears before the original message).
   • Result: Appears as if cause and effect are reversed.
   • Solution: Provide consistent prefix reads.
     • Ensures writes are read in the order they were made.

How Request Coalescing Works

Link: https://newsletter.systemdesigncodex.com/p/how-request-coalescing-works

Concept: Request Coalescing is a technique for optimizing database queries by reducing redundant requests for the same data.

Application: Successfully used by Discord to manage trillions of messages efficiently.

Functionality:

1. Setup: Involves intermediary data services between the API layer and the database.
2. Process:
   • When the first request is made, a worker task is initiated in the data service.
   • Subsequent requests for the same data subscribe to this existing task.
   • The worker task queries the database once and returns the result to all subscribers simultaneously.

Differences from Caching:

1. Request Initiation: In request coalescing, only the first request triggers a database query. Subsequent ones wait for its result. In caching, all requests would hit the cache.
2. Use with Caching: Request coalescing can complement caching by reducing the number of hits to the cache.

Internal Working (Based on Discord's Implementation):

• Each worker task maintains a local state with requests and a list of requesters.
• Responses are propagated to all waiting requesters upon arrival.

Applicability:

• Request Coalescing is particularly useful for systems with high concurrency and redundant requests.
• The necessity of this technique depends on the scale and specific challenges of the system.

How to Migrate a MySQL Database

Link: https://newsletter.systemdesign.one/p/how-to-migrate-a-mysql-database

Context: Tumblr's MySQL database, spanning 21 terabytes and 60+ billion rows across 200+ servers, necessitated a migration strategy that minimizes user impact.

Challenges:

• Maintaining high availability and scalability.
• Minimizing downtime and user impact during migration.

Strategies Used:

1. CQRS Pattern (Command and Query Responsibility Segregation):

   • Separated read and write operations for the database.
   • Ensured continuous read availability during migration.

2. Leader-Follower Replication:

   • Leader in a remote data center handled read-write operations.
   • Local data center had followers for handling read requests.
   • Used persistent connections to reduce latency issues.

3. Database Proxy (ProxySQL):

   • Positioned in the local data center.
   • Maintained persistent connections to the remote leader.
   • Enabled connection pooling, improving performance and reducing disconnections.

Migration Process:

1. Preparation:
   • Stored metadata of leaders, followers, and proxies in each data center.
2. Migration Execution:
   • Shifted the database leader from Data Center A to B.
   • Automated tools redirected followers and proxies to the new leader.
3. Outcome:
   • Followers continued serving read requests.
   • Write requests were briefly halted or buffered, resulting in minimal user impact.

Consideration for Further Improvement:

• Leader-Leader Replication: Could enhance write availability but poses a risk of data conflicts.
• Reason for Non-Use: Potential conflicts might be why Tumblr opted against this approach.

Durability

Link: https://newsletter.francofernando.com/p/durability

Core Objective: Durability, ensuring data is not lost despite failures like power outages, system crashes, or hardware issues.

Single-Node Database Persistence

1. Durability Method: Data is written to nonvolatile storage (hard drive, SSD).
2. Transaction Processing:
   • Log Writing: Data first written to a log file before making actual data updates.
   • Update Execution: After log entry, the database updates the actual data.
   • Role of Log: Enables reprocessing of transactions to restore consistent state post-failure.
   • Efficiency: Log writing is fast due to its append-only nature, minimizing seek time.

Distributed Database Persistence

1. Complexity: Higher due to the need for coordination across multiple servers.
2. Two-Phase Commit Protocol:
   • Coordinator Role: A designated server coordinates the commit process.
   • Process:
     • Coordinator sends commit instruction to all participant servers.
     • Waits for acknowledgments from all participants.
     • Finalizes the transaction with a commit or rollback based on responses.

Redis

Link: https://newsletter.francofernando.com/p/redis

Redis Overview:

• Redis stands for REmote DIctionary Server.
• It's an open-source, key-value database store.
• Functions as a data structure server, supporting various data structures like Strings, Lists, Sets, Hashes, Sorted Sets, and HyperLogLogs.

History:

• Created by Salvatore Sanfilippo in the late 2000s.
• Developed to address scaling issues with MySQL in real-time analytics.
• Gained popularity and wide adoption due to its efficiency and flexibility.

Operations and Data Types:

• Basic operations include GET and SET.
• Supports diverse data structures, each with specific use cases and operations.

Redis Architectures:

1. Single Instance: Simplest form, running on the same or a separate server.
2. Replicated Instances: Primary instance replicated across secondary instances for parallel read requests and backup.
3. Sentinel: Manages high availability, monitoring, and failure handling.
4. Cluster: Distributes data across multiple machines using sharding.

Data Persistency:

• Offers two methods:
  • RDB (Redis Database Backup): Snapshot-based backups.
  • AOF (Append Only File): Logs every change for more recent backups.
• Choice between RDB and AOF depends on the need for speed vs. data recency.

Single-thread Model:

• Utilizes a single-threaded model for operations, avoiding multi-threading overhead.
• Performance typically limited by memory and network, not CPU.

Use Cases:

• Database: As a primary key-value store.
• Cache: For storing frequent queries or caching API requests.
• Pub/Sub: For scalable and fast messaging systems.

Salt and Pepper

Link: https://newsletter.francofernando.com/p/salt-and-pepper

1. Hashing

• Method: Converts plain text passwords into a random string of characters.
• Process: User's password is hashed and compared with the stored hash during login.
• Common Algorithms: MD5, SHA family. However, these are vulnerable to rainbow table attacks.

2. Salting

• Purpose: Enhances hashing by defending against pre-computation attacks like rainbow tables.
• Implementation:
  • Generate a unique salt for each password.
  • Combine salt with the password and hash the result.
  • Store the salt in plain text and the hashed password in the database.
• Validation Process:
  1. Retrieve the salt from the database.
  2. Combine entered password with salt and hash.
  3. Compare with stored hash for validation.
• Uniqueness: Ensures each stored hash is unique, even for identical passwords.

3. Peppering

• Function: Adds an extra layer of security to salting.
• Mechanism:
  • Add a pepper value to the password before hashing.
  • The pepper is not stored in the database.
• Login Process:
  • Attempt combinations of password and pepper until a match is found.
• Benefit: Significantly increases the effort required for brute force attacks.

Key Takeaways:

• Combining Techniques: Using both salting and peppering provides robust protection.
• Importance of Uniqueness: Unique salts and peppers make each hash distinct.
• Updating Practices: Continuously update and improve password storage methods to counteract new hacking techniques.

Moving from Monolithic to Microservices

Link: https://newsletter.systemdesigncodex.com/p/from-monolithic-to-microservices

1. Modular Monolith Approach

• Concept: Incorporates modular design within a monolithic architecture.
• Characteristics:
  • Loosely-coupled modules.
  • Well-defined boundaries.
  • Explicit dependencies.
• Structure: Application divided into independent modules.
• Deployment: Still maintains single application deployment.
• Advantages:
  • Streamlines development and maintenance.
  • Offers microservices-like benefits without associated complexities.

2. Evolution to Vertical Slice Architecture

• Design Shift: From horizontal layers to vertical slices of business functionality.
• Benefits:
  • Scoped changes to specific business areas.
  • Easier feature addition and modification.
• Microservices Potential: Vertical modules can gradually evolve into independent microservices.
• Learning Opportunity: Provides insights into domain and functional splits.

Key Takeaway

• Balance: No inherent superiority of microservices over monoliths or vice versa.
• Evolutionary Approach: Adapt the architecture to evolving application needs.
• Pragmatism: Choose the architecture that best suits the project's requirements and context.

The Secret Trick to High-Availability

Link: https://newsletter.systemdesigncodex.com/p/the-secret-trick-to-high-availability

Strategies for Static Stability

1. Active-Active High Availability:

   • Implementation: Distribute traffic across instances in multiple Availability Zones (AZs).
   • Example: If two instances are needed, create three (50% over-provisioning).
   • Benefit: Maintains full capacity even if an entire AZ fails.

2. Active-Passive High Availability:

   • Use Case: For stateful services like databases.
   • Setup: Primary instance in one AZ and a standby in another.
   • Function: Standby becomes primary if the original primary AZ goes down.

Criticism and Justification

• Criticism: Viewed as resource wasteful due to over-provisioning.
• Justification:
  • Essential for mission-critical applications where downtime is unacceptable.
  • Used by major cloud services like AWS (EC2, S3, RDS) to prevent outages.

Key Takeaway

• Outages as a Norm: Disruptions are inevitable; planning for them is crucial.
• Risk Management: Over-provisioning is a strategic choice to mitigate downtime risks.
• Context-Dependent: The level of static stability required varies based on the system's criticality.

Static stability, while resource-intensive, is a fundamental approach for ensuring continuous operation in high-stake environments where reliability and uptime are non-negotiable.

4 Types of NoSQL Databases

Link: https://newsletter.systemdesigncodex.com/p/4-types-of-nosql-databases

1. Document Databases

• Examples: MongoDB, Couchbase, RavenDB.
• Data Storage: In the form of JSON, BSON, or XML documents.
• Advantages: Align closely with domain-level data objects in applications.
• Use Case: Ideal for projects requiring a structure close to application data.

2. Key-Value Store

• Examples: Redis, etcd, DynamoDB.
• Structure: Data stored as key-value pairs.
• Simplicity: Resembles a two-column table (key and value).
• Use Cases: Caching, shopping carts, user profiles.

3. Column-Oriented Database

• Examples: Apache Cassandra, Apache HBase.
• Storage Method: Data stored in columns rather than rows.
• Advantages: Efficient for analytics and aggregations on specific columns.
• Considerations: Not strongly consistent; write operations can be complex.

4. Graph Databases

• Examples: Neo4j, Amazon Neptune.
• Concept: Focuses on relationships between data elements (nodes and links).
• Strengths: Eliminates the need for multiple table joins as in SQL databases.
• Use Cases: Knowledge graphs, social networks, map-like applications.

Decision Guide:

• Document DBs: Versatile, suitable for most applications traditionally using SQL.
• Key-Value Stores: For applications requiring fast read/write access to data items.
• Column-Oriented: Analytics and operations on large datasets.
• Graph Databases: Applications where relationships are central to the data model.