System Design Crash Course: From CPU Cache to Load Balancers and Sharding
Key points from freeCodeCamp's system design course: availability and SLAs, CAP, networking, API design, caching, CDNs, proxies, load balancers and database scaling.
By Tech Team Strategy · Nov 17, 2026 · 5 min read

This hour-long freeCodeCamp course by Hayk Simonyan covers the system design concepts you need for an interview, where, as he puts it, nobody wants to see you write code, only how you "glue an entire system together." It runs from the parts inside one computer to production architecture, design requirements, networking, protocols, APIs, caching, proxies, load balancers and databases. Here are the main points from each section.
This article is our own summary, with short quotes, from the video's automatic transcript. Watch the full video above for everything that was said.
The short version
- Good design means scalability, maintainability and efficiency, and planning for failure. At its heart are moving, storing and transforming data.
- Every decision is a trade-off. The CAP theorem says a distributed system can achieve only two of consistency, availability and partition tolerance at once.
- Measure what matters: availability (99.9% allows about 8.76 hours of downtime a year; 99.999% about 5 minutes), SLOs and SLAs, throughput and latency.
- Get the design right early. Unlike code, a system is very hard to redesign later.
- Caching, CDNs, proxies and load balancers are the main tools for speed and reliability; sharding, replication, indexing and query optimization for databases.
The words worth stealing from this conversation, what they mean, and who said them.
Elegant words
- graceful degradation
- The ability of a system to maintain limited functionality even when portions of it have failed.
- Hayk Simonyan: “Resilience: redundancy, fault tolerance and graceful degradation.”
- idempotent
- An operation that produces the same result no matter how many times it is executed.
- Hayk Simonyan: “GET requests that are idempotent and never change data, versioning for backward compatibility, rate limiting against DDoS attacks, and CORS settings.”
- single point of failure
- A part of a system that, if it fails, will stop the entire system from working.
- Hayk Simonyan: “A load balancer is itself a single point of failure, so run redundant pairs with failover.”
- trade-off
- A situational decision that involves losing one quality or aspect of something in return for gaining another.
- Hayk Simonyan: “Every decision is a trade-off.”
Tech vocabulary
- ACID-compliant
- Databases that guarantee Atomicity, Consistency, Isolation, and Durability for reliable transaction processing.
- Hayk Simonyan: “Relational databases (PostgreSQL, MySQL, SQLite) are ACID-compliant.”
- CDNs
- Content Delivery Networks, geographically distributed groups of servers that work together to provide fast delivery of internet content.
- Hayk Simonyan: “CDNs cut latency, add availability and offer security such as DDoS protection.”
- CI/CD
- Continuous Integration and Continuous Deployment, a method to frequently deliver apps to customers by introducing automation.
- Hayk Simonyan: “CI/CD moves code from repository to production through tests with no manual steps.”
- consistent hashing
- A special kind of hashing such that when a hash table is resized, only a few keys need to be remapped.
- Hayk Simonyan: “Load-balancing algorithms: round robin, least connections, least response time, IP hashing, weighted variants, geographic routing and consistent hashing on a hash ring.”
- gRPC
- A high-performance, open-source universal Remote Procedure Call framework developed by Google.
- Hayk Simonyan: “gRPC runs on HTTP/2 with protocol buffers, efficient for microservices but less readable.”
- latency
- The time delay between a cause and its effect, or the time it takes for a data packet to travel.
- Hayk Simonyan: “Throughput versus latency (time for one request).”
- load balancers
- Devices or software that distribute network or application traffic across a cluster of servers.
- Hayk Simonyan: “Load balancers and reverse proxies like Nginx spread traffic; storage runs on separate servers.”
- rate limiting
- A strategy for limiting network traffic to prevent users from exhausting system resources.
- Hayk Simonyan: “versioning for backward compatibility, rate limiting against DDoS attacks, and CORS settings.”
- reverse proxies
- Servers that sit in front of web servers and forward client requests to those web servers.
- Hayk Simonyan: “Load balancers and reverse proxies like Nginx spread traffic; storage runs on separate servers.”
- sharding
- A database partitioning strategy that separates very large databases into smaller, faster, more easily managed parts.
- Hayk Simonyan: “Scale vertically or horizontally with sharding and replication.”
- SLAs
- Service Level Agreements, commitment contracts between a service provider and a client.
- Hayk Simonyan: “An SLO might say 99.9% of requests answer within 300 milliseconds; an SLA is a contract, and missing it can mean refunds.”
- throughput
- The amount of material or items passing through a system or process, such as data over a network.
- Hayk Simonyan: “Throughput versus latency. Batching raises throughput but can add latency.”
- WebRTC
- A free, open-source project providing web browsers and mobile applications with real-time communication via simple APIs.
- Hayk Simonyan: “WebRTC for browser video calls; MQTT for low-power IoT devices.”
- WebSockets
- A communications protocol providing full-duplex communication channels over a single TCP connection.
- Hayk Simonyan: “WebSockets give two-way real-time updates for chat, sports scores and stock feeds.”
Trends and buzzwords
- GraphQL
- A query language for APIs and a runtime for fulfilling those queries with existing data.
- Hayk Simonyan: “GraphQL returns exactly what's asked but complex queries can strain the server.”
- NoSQL
- A non-relational database design that does not require a fixed schema and is easy to scale.
- Hayk Simonyan: “NoSQL is schema-less and suits unstructured data and scale.”
Inside one computer
Data lives on disk (non-volatile; an SSD reads around 500 to 3,500 MB per second, a hard drive about 80 to 160), in RAM (volatile, often above 5,000 MB per second) and in the CPU's L1, L2 and L3 caches (a few nanoseconds). The CPU runs compiled machine code; the motherboard connects everything.
A production app
- CI/CD (Jenkins, GitHub Actions) moves code from repository to production through tests with no manual steps.
- Load balancers and reverse proxies like Nginx spread traffic; storage runs on separate servers.
- Logging and monitoring (PM2 on the back end, Sentry on the front end) feed alerts, often into a Slack channel.
- Debugging: find the issue in the logs, reproduce it in staging ("never debug directly in the production environment"), then ship a hotfix before a permanent fix.
Design requirements
- CAP theorem (Eric Brewer): consistency (every node has the same data), availability (always responsive) and partition tolerance (works despite network splits). A bank favors consistency and partition tolerance, even if some transactions are slower.
- Availability and SLAs: an SLO might say 99.9% of requests answer within 300 milliseconds; an SLA is a contract, and missing it can mean refunds.
- Resilience: redundancy, fault tolerance and graceful degradation.
- Throughput (requests per second, queries per second, bytes per second) versus latency (time for one request). Batching raises throughput but can add latency.
Networking
IPv4 addresses are 32-bit (about 4 billion); IPv6 is 128-bit. Packets carry IP headers. TCP is reliable, with sequence numbers and a three-way handshake; UDP is faster but unreliable, fine for video calls and live streaming. DNS is the internet's phone book, overseen by ICANN, with A records (IPv4) and AAAA records (IPv6). Firewalls guard networks; ports identify services (80 for HTTP, 22 for SSH).
Application protocols
HTTP is stateless request-response with status codes (200s success, 300s redirect, 400s client error, 500s server error) and methods (GET, POST, PUT, PATCH, DELETE). WebSockets give two-way real-time updates for chat, sports scores and stock feeds. Also covered: SMTP, IMAP and POP3 for email; FTP and SSH; WebRTC for browser video calls; MQTT for low-power IoT devices; AMQP for enterprise messaging (as in RabbitMQ); and RPC, calling a remote function as if it were local.
API design
Using a Shopify-style e-commerce API, he maps CRUD to endpoints (POST, GET, PUT/PATCH, DELETE on /api/products). REST is stateless and widely consumable but can over- or under-fetch. GraphQL returns exactly what's asked but complex queries can strain the server. gRPC runs on HTTP/2 with protocol buffers, efficient for microservices but less readable. Best practices: nested endpoints for relationships (/users/{id}/orders), pagination and filters (limit, offset, date ranges), GET requests that are idempotent and never change data, versioning (v1, v2) for backward compatibility, rate limiting against DDoS attacks, and CORS settings.
Caching and CDNs
A site hosted in Finland may load in about 100 milliseconds in Europe but 3 to 5 seconds in Mexico. Caching keeps copies closer: browser caching (controlled by the Cache-Control header), server caching (in memory, such as Redis), database caching and CDNs. Write strategies are write-around, write-through (consistent but slower) and write-back (fast but risks loss on a crash). Eviction policies are LRU, FIFO and LFU. Pull CDNs fetch from the origin on first request; push CDNs suit large, rarely updated files. CDNs cut latency, add availability and offer security such as DDoS protection; dynamic and real-time content still hits the origin.
Proxies and load balancers
A forward proxy sits in front of clients, hiding their IP, controlling and monitoring access, caching, and anonymizing browsing. A reverse proxy sits in front of servers: load balancers, CDNs, web application firewalls, SSL offloading. Load-balancing algorithms: round robin, least connections, least response time, IP hashing (session persistence), weighted variants, geographic routing and consistent hashing on a hash ring. Health checks stop traffic to failed servers. Options include F5 BIG-IP and Citrix hardware, HAProxy and Nginx software, and AWS, Azure and Google Cloud services. A load balancer is itself a single point of failure, so run redundant pairs with failover, monitor it, auto-scale and self-heal, or use DNS failover.
Databases
Relational databases (PostgreSQL, MySQL, SQLite) are ACID-compliant: atomicity, consistency, isolation, durability. NoSQL (MongoDB, Cassandra, Redis, Neo4j) is schema-less and suits unstructured data and scale. In-memory databases (Redis, Memcached) are for caching and sessions. Scale vertically (bigger machine, limited) or horizontally with sharding (range-based, directory-based or geographic) and replication (master-slave or master-master). To go faster: caching, indexing frequently queried columns, and query optimization, such as minimizing joins and using EXPLAIN plans.
What it means for your business
- Write down availability targets. The gap between 99.9% and 99.999% is hours of downtime versus minutes, and very different costs.
- Automate deployments and alerts. CI/CD plus alerts to a team channel shortens outages.
- Version your APIs and rate-limit them to protect customers and your servers.
- Cache and use a CDN for customers far from your servers.
- Never run a single load balancer or database where downtime costs money.
Video details
- Title: System Design Concepts Course and Interview Prep
- Instructor: Hayk Simonyan
- Published by: freeCodeCamp.org on YouTube
- Watch: youtube.com/watch?v=F2FmTdLtb_4
- Note: figures and quotes come from the video's automatic transcript and are lightly cleaned. Check the video for exact wording.
Want to put this to work?
We help owners turn AI trends into a plan, a product and a launch. Book a free call or see our services.
Full transcript
Read the whole interview, word for word, with a timestamp on every answer. Free with a sign-in.
Sign in to read itOriginally published by Tech Team Strategy
What did you think of this story?
+10 tokensTap to vote ↓
Get the briefing
The week's essential stories, free in your inbox.
More

How LangChain Interviews Forward Deployed Engineers: Case Project, Behavioral and Tips
How LangChain hires forward deployed engineers: the case-project demo, behavioral interview, time management, keeping up with AI and diverse backgrounds.
Nov 17, 2026 · 4 min

Rippling FDE Mock Interview: How to Push Back, Solve End to End and Work Across Teams
How a Rippling forward deployed engineer answers behavioral interview questions on pushing back on clients, end-to-end problem solving and cross-team work.
Nov 17, 2026 · 4 min

Database Sharding Explained: Techniques, Trade-offs and When to Use It
Database sharding basics: geo, range and hash-based sharding, manual vs automatic sharding, hotspots, and the pros and cons before you split your database.
Nov 17, 2026 · 3 min
Comments
Sign in to commentJoin the conversation and earn 10 tokens.