Follow me for system design & book-writing tips.

Alex Xu is a software engineer and author. His book System Design Interview - An Insider’s Guide is one of Amazon's best-selling books in the “Distributed Systems & Computing” category.  As of December 2020, his book is being translated into 7 different languages (Traditional Chinese, Russian, Japanese, Spanish, Polish, Simplified Chinese, and Korean).

--------------
Newsletter: https://blog.bytebytego.com
Twitter: https://twitter.com/alexxubyte
YouTube: https://www.youtube.com/c/ByteByteGo

Alex Xu

MCP vs RAG vs AI Agents

Everyone is talking about MCP, RAG, and AI Agents. Most people are still mixing them up. They’re not competing ideas. They solve very different problems at different layers of the stack.

MCP (Model Context Protocol) is about how LLMs use tools. Think of it as a standard interface between an LLM and external systems. Databases, file systems, GitHub, Slack, internal APIs. 

Instead of every app inventing its own glue code, MCP defines a consistent way for models to discover tools, invoke them, and get structured results back. MCP doesn’t decide what to do. It standardizes how tools are exposed.

RAG (Retrieval-Augmented Generation) is about what the model knows at runtime. The model stays frozen. No retraining. When a user asks a question, a retriever fetches relevant documents (PDFs, code, vector DBs), and those are injected into the prompt.

RAG is great for:
- Internal knowledge bases
- Fresh or private data
- Reducing hallucinations

But RAG doesn’t take actions. It only improves answers.

AI Agents are about doing things. An agent observes, reasons, decides, acts, and repeats. It can call tools, write code, browse the internet, store memory, delegate tasks, and operate with different levels of autonomy.

--
Subscribe to our weekly newsletter to get a Free System Design PDF (368 pages): https://lnkd.in/gauQcE45

#systemdesign #coding #interviewtips 
.

Author of 4 Bestselling Books | Co-Founder of ByteByteGo

No alternative text description for this image

RabbitMQ vs Kafka vs Pulsar

RabbitMQ, Kafka, and Pulsar all move messages, but they solve very different problems under the hood.
This diagram looks simple, but it hides three very different mental models for building distributed systems.

RabbitMQ is a classic message broker. Producers publish to exchanges, exchanges route messages to queues, and consumers compete to process them. 

Messages are pushed, acknowledged, and then gone. It’s great for task distribution, request handling, and workflows where “do this once” really matters.

Kafka flips the model. It’s not a queue, it’s a distributed log. Producers append events to partitions. Data stays there based on retention, not consumption. Consumers pull data using offsets and can replay everything. 

This is why Kafka works so well for event streaming, analytics, and pipelines where multiple teams need the same data at different times.

Pulsar tries to combine both worlds. Brokers handle serving traffic, while BookKeeper stores data in a durable ledger. Consumers track position with cursors instead of offsets. 

This separation lets Pulsar scale storage and compute independently and support both streaming and queue-like patterns.

Choosing between them isn’t about “which is faster” or “which is popular.” It’s about how you want data to flow, how long it should live, and how many times it needs to be read.

--
Subscribe to our weekly newsletter to get a Free System Design PDF (368 pages): https://lnkd.in/gauQcE45

#systemdesign #coding #interviewtips 
.

Al Topics to Learn before Taking Al/ML Interviews (Part 1)

AI interviews often test fundamentals, not tools. This visual splits them into two buckets that show up repeatedly: Traditional AI and Modern AI.

Traditional AI focuses on fundamental ML topics, mostly from before neural networks became dominant.

Modern AI focuses on neural network foundations and newer concepts like transformers, RAG, and post-training.

Interviewers generally expect you to know both. They expect you to explain how they work, when they break, and the trade-offs.

Use this as a checklist, and make sure you can explain each topic clearly before your next AI interview.

Part 2 in the comments.

Over to you: which topic here do you find hardest to explain under interview pressure? What else is missing?

--
Subscribe to our weekly newsletter to get a Free System Design PDF (368 pages): https://lnkd.in/gauQcE45

#systemdesign #coding #interviewtips 
.

12 Architectural Concepts Developers Should Know

1 - Load Balancing: Distributed incoming traffic across multiple servers to ensure no single node is overwhelmed.
2 - Caching: Stores frequently accessed data in memory to reduce latency.
3 - Content Delivery Network (CDN): Stores static assets across geographically distributed edge servers so users download content from the nearest location.
4 - Message Queue: Decouples components by letting producers enqueue messages that consumers process asynchronously.
5 - Publish-Subscribe: Enables multiple consumers to receive messages from a topic.
6 - API Gateway: Acts as a single entry point for client requests, handling routing, authentication, rate limiting, and protocol translation.
7 - Circuit Breaker: Monitors downstream service calls and stops attempts when failures exceed a threshold.
8 - Service Discovery: Automatically tracks available service instances so components can locate and communicate with each other dynamically.
9 - Sharding: Splits large datasets across multiple nodes based on a specific shard key.
10 - Rate Limiting: Controls the number of requests a client can make in a given time window to protect services from overload.
11 - Consistent Hashing: Distributes data across nodes in a way that minimizes reorganization when nodes join or leave.
12 - Auto Scaling: Automatically adds or removes compute resources based on defined metrics.

Over to you: Which architectural concept will you add to the list?

--
Subscribe to our weekly newsletter to get a Free System Design PDF (368 pages): https://lnkd.in/gauQcE45

#systemdesign #coding #interviewtips 
.

undefined undefined

bytebytego

9 Key AI Concepts Explained in 7 minutes (YouTube video link at the end).

- Tokenization
- Text Decoding
- Prompt Engineering
- Multi Step AI Agents
- RAGs
- RLHF
- VAE
- Diffusion Models
- LoRA

--
Watch the whole YouTube video here: https://lnkd.in/gyrnFSrF

#systemdesign #coding #interviewtips
.

10 Types of API Testing

1 - Smoke Testing: Quickly checks if the core functionalities of the API are working as expected.
2 - Functional Testing: Validates that the API behaves as specified in the functional specifications.
3 - Integration Testing: Ensures different modules or services interact correctly through the APIs.
4 - Regression Testing: Verifies that new changes don’t break existing API functionalities.
5 - Load Testing: Measures how the API performs under expected load conditions.
6 - Stress Testing: Tests the API behavior under extreme or peak load conditions.
7 - Security Testing: Assesses the API for vulnerabilities and potential external threats.
8 - UI Testing: Checks the interaction between the user interface and the API.
9 - Fuzz Testing: Sends random or invalid data to the API to detect unexpected crashes or errors.
10 - Reliability Testing - Evaluates the API’s consistency and stability over an extended period.

Over to you: Will you add any other API testing type to the list?

--
Subscribe to our weekly newsletter to get a Free System Design PDF (368 pages): https://lnkd.in/gauQcE45

#systemdesign #coding #interviewtips 
.

11 Ways To Use AI To Increase Your Productivity

AI is changing how we work. People who use AI get more done in less time. You do not need to code. You need to know which tool to use and when.

For example, instead of reading long technical blogs, you can upload them to Google’s NotebookLM and ask it to summarize the key points.

Or you can use Otter.ai to turn meeting transcripts into action items, decisions, and highlights.

Here is a list of 19 tools that can speed up your daily workflow across different areas. Save this for the next time you feel stuck getting started.

Over to you: What’s the underrated AI tool that others might not know about?

--
Subscribe to our weekly newsletter to get a Free System Design PDF (368 pages): https://lnkd.in/gauQcE45

#systemdesign #coding #interviewtips 
.

Agent Skills, Clearly Explained.

Why do we need Agent Skills? Long prompts hurt agent performance. Instead of one massive prompt, agents keep a small catalog of skills, reusable playbooks with clear instructions, loaded only when needed.

Here is what the Agent Skills workflow looks like:

1. User Query: A user submits a request like “Analyze data & draft report”.
2. Build Prompt + Skills Index: The agent runtime combines the query with Skills metadata, a lightweight list of available skills and their short descriptions.
3. Reason & Select Skill: The LLM processes the prompt, thinks, and decides: "I want Skill X."
4. Load Skill into Context: The agent runtime receives the specific skill request from the LLM. Then, it loads SKILL. md and adds it into the LLM's active context.
5. Final Output: The LLM follows SKILL. md, runs scripts, and generates the final report.

By dynamically loading skills only when needed, Agent Skills keep context small and the LLM’s behavior consistent.

Over to you: What skills would you find most useful in agents?

--
Subscribe to our weekly newsletter to get a Free System Design PDF (368 pages): https://lnkd.in/g7M3YdKr

#systemdesign #coding #interviewtips 
.

GPT-5 is not one model. 

It is a unified system with multiple models, safeguards, and a real-time router.
This post and diagram are based on our understanding of the GPT 5 system card.
When you send a query, the mode determines which model to use and how much work the system does.

Instant mode sends the query directly to a fast, non-reasoning model named GPT-5-main. It optimizes for latency and is used for simple or low-risk tasks like short explanations or rewrites.

Thinking mode uses a reasoning model named GPT-5-thinking that runs multiple internal steps before producing the final answer. This improves correctness on complex tasks like math or planning.

Auto mode adds a real-time router. A lightweight classifier looks at the query and decides whether to use GPT-5-main or GPT-5-thinking when deeper reasoning is needed.

Pro mode does not use a different model. It uses GPT-5-thinking but samples multiple reasoning attempts and selects the best one using a reward model.

Across all modes, safeguards run in parallel at various stages. A fast topic classifier determines whether the topic is high-risk, followed by a reasoning monitor that applies stricter checks to ensure unsafe responses are blocked.

Over to you: What's your favorite AI chat bot?

--
Subscribe to our weekly newsletter to get a Free System Design PDF (368 pages): https://lnkd.in/g7M3YdKr

#systemdesign #coding #interviewtips 
.

Luiz Mosciaro

How does gRPC work?

RPC (Remote Procedure Call) is called “remote” because it enables communications between remote services when services are deployed to different servers under microservice architecture. From the user’s point of view, it acts like a local function call.

The diagram below illustrates the overall data flow for gRPC.

Step 1: A REST call is made from the client. The request body is usually in JSON format.

Steps 2 - 4: The order service (gRPC client) receives the REST call, transforms it, and makes an RPC call to the payment service. gPRC encodes the client stub into a binary format and sends it to the low-level transport layer.

Step 5: gRPC sends the packets over the network via HTTP2. Because of binary encoding and network optimizations, gRPC is said to be 5X faster than JSON.

Steps 6 - 8: The payment service (gRPC server) receives the packets from the network, decodes them, and invokes the server application.

Steps 9 - 11: The result is returned from the server application, and gets encoded and sent to the transport layer.

Steps 12 - 14: The order service receives the packets, decodes them, and sends the result to the client application.

Over to you: Have you used gPRC in your project? What are some of its limitations?

--
Subscribe to our weekly newsletter to get a Free System Design PDF (368 pages): https://lnkd.in/g7M3YdKr

#systemdesign #coding #interviewtips 
.

We started a new YouTube channel dedicated to AI fundamentals, infrastructure and models.

So far, we’ve published four videos.

We plan to publish 1 video every week.

Since this channel is new, we’re still shaping its direction and would genuinely love your input. 

If there are topics you’d like to see covered, please share them with us. Thank you!

--
Subscribe here: https://lnkd.in/eGFwfd2d

#systemdesign #coding #ai
.

A nice cheat sheet of different monitoring infrastructure in cloud services.

This cheat sheet offers a concise yet comprehensive comparison of key monitoring elements across the three major cloud providers and open-source / 3rd party tools.

Let's delve into the essential monitoring aspects covered:
- Data Collection: Gather information from diverse sources to enhance decision-making.
- Data Storage: Safely store and manage data for future analysis and reference.
- Data Analysis: Extract valuable insights from data to drive informed actions.
- Alerting: Receive real-time notifications about critical events or anomalies.
- Visualization: Present data in a visually comprehensible format for better understanding.
- Reporting and Compliance: Generate reports and ensure adherence to regulatory standards.
- Automation: Streamline processes and tasks through automated workflows.
- Integration: Seamlessly connect and exchange data between different systems or tools.
- Feedback Loops: Continuously refine strategies based on feedback and performance analysis.

Over to you: How do you prioritize and leverage these essential monitoring aspects in your domain to achieve better outcomes and efficiency?

--
Subscribe to our weekly newsletter to get a Free System Design PDF (368 pages): https://lnkd.in/gauQcE45

#systemdesign #coding #interviewtips 
.

🚀 Last Day to Enroll: Become an AI Engineer | By building, not just watching | Cohort 4!

After the amazing response to our first two cohorts, with nearly 1,000 people joining, we are excited to share that we’re opening the next round of Become an AI Engineer.

This is not your typical AI course focused only on tools and frameworks. The mission is simple: help engineers build strong foundations and practical end to end skills to grow confidently into AI engineering roles.

What makes this cohort stand out:
- Learn by building: You will create real world AI applications instead of just watching videos
- Clear and structured path: A thoughtfully designed curriculum that takes you from core concepts to advanced topics, step by step
- Live feedback and mentorship: Get hands on guidance from instructors and learn alongside peers
- Strong community: Learning is faster and more motivating when you are not doing it alone

We care deeply about skill building, not passive learning or surface level theory. The goal is for every participant to leave with the confidence and ability to build real AI systems.

Today is the last day to enroll before it starts.

--
Check it out here: https://lnkd.in/gZEmb6rh

#AI #AIEngineer #MachineLearning 
.

Deploying or upgrading services is risky. In this post, we explore risk mitigation strategies.

The diagram below illustrates the common ones.

Multi-Service Deployment
In this model, we deploy new changes to multiple services simultaneously. This approach is easy to implement. But since all the services are upgraded at the same time, it is hard to manage and test dependencies. It’s also hard to rollback safely.

Blue-Green Deployment
With blue-green deployment, we have two identical environments: one is staging (blue) and the other is production (green). The staging environment is one version ahead of production. Once testing is done in the staging environment, user traffic is switched to the staging environment, and the staging becomes the production. This deployment strategy is simple to perform rollback, but having two identical production quality environments could be expensive.

Canary Deployment
A canary deployment upgrades services gradually, each time to a subset of users. It is cheaper than blue-green deployment and easy to perform rollback. However, since there is no staging environment, we have to test on production. This process is more complicated because we need to monitor the canary while gradually migrating more and more users away from the old version.

A/B Test
In the A/B test, different versions of services run in production simultaneously. Each version runs an “experiment” for a subset of users. A/B test is a cheap method to test new features in production. We need to control the deployment process in case some features are pushed to users by accident.

Over to you - Which deployment strategy have you used? Did you witness any deployment-related outages in production and why did they happen?

--
Subscribe to our weekly newsletter to get a Free System Design PDF (368 pages): https://lnkd.in/g7M3YdKr

#systemdesign #coding #interviewtips 
.

taplio