System design · simulated, measured, compared

Learn system design by watching it break.

This system is about to lose its cache. See what happens next.

Explore Build from scratch or solve a problem →

Free runs up to 5,000 requests/sec · 6 free problems · 100 welcome credits with Google sign-in

A recorded incident · scroll to play itthe designSkip the story ↓
01 / 06 · The system

Six components.
Nothing wrong yet.

A load balancer, three APIs, a cache and Postgres. The cache answers 9 reads in 10.

02 / 06 · Traffic

2,500 requests a second.

Blue goes to the cache. Only the orange trickle reaches Postgres.

03 / 06 · The flush

A deploy flushes the cache.

Every read now hits Postgres: ten times its usual load. Its 40 connections fill and queries stack up.

04 / 06 · The crash

Postgres goes down.

Requests bounce off. p99 looks fine, because failed requests are not counted in it.

05 / 06 · One change

One change.

Add a read replica to Postgres. $447 more a month: that is the trade-off.

06 / 06 · The rerun

Same traffic. Same flush.
It holds.

Same requests, same seed. The replica takes the reads the primary could not.

BeforeAfter
Goals met0 of 22 of 2
Success, whole run69.7%100.0%
p99, OK requests869 ms37 ms
Cloud cost / month$1,591$2,038
Try this scenarioWhy it happens
Three ways in

Start where you want.

Want a product to design instead? 56 with a brief, from YouTube to an LLM API.

1

Fix a broken system

Run it, watch it fail, fix it, prove it on the same traffic.

Try the first challenge → 2

Build from scratch

An empty lab: no problem, no goals. Add components, send traffic, see what breaks.

Open an empty lab → 3

Follow a learning path

Fundamentals, interviews or AI infrastructure, in order.

See the paths →
Who it's for

Built for the design round.

The interview engineers face, rehearsed on real numbers.

Prepare for the system design round for roles like

AI engineering
AI engineerAI infrastructure engineerLLM application and agent engineerML platform engineerForward deployed engineer
Software engineering
Backend engineerSenior / staff engineerSRE and platform engineerSolutions architectEngineering manager

The skills those rounds test

Capacity estimatesCaching and hit ratesQueues and backpressureTimeouts and retriesReplication and failoverRate limiting and attacksLLM serving and quotasCost trade-offs

Agentic infrastructure you'll run into

Tool-call fan-outRunaway agent loops and step capsMCP tool serversContext, checkpoints and compactionSpend limits and approval gatesPrompt-injection defenceProvider quotas and fallbacks

System design rounds are standard at

Engineers on these teams sat a system design round on the way in. Practise the same kind of round here, on a system that really runs.

  • OpenAI
  • Anthropic
  • NVIDIA
  • Google
  • Meta
  • Microsoft
  • Amazon
  • Apple
  • Netflix
  • Uber
  • Stripe
  • Airbnb
  • Databricks
  • LinkedIn
  • Palantir
  • Scale AI
  1. 1Learn how each part behaves under load
  2. 2Practise the round: design it, break it, fix it
  3. 3Walk in with numbers you have measured
Start practising, free

Names and marks are trademarks of their owners, shown only to identify companies where this interview is standard. ScaleMySystems is not affiliated with or endorsed by any of them and does not reproduce their questions.

Why system design, now

AI writes the code.
Architecture decides if it survives.

Shipping takes an afternoon. Holding on launch day is still a design decision.

CAPACITY

Code is cheap. Capacity isn't.

Launch-day traffic finds the queue you didn't size.

AI FEATURES

Every AI feature is a distributed system

Gateway, retrieval, vector store, model provider: one request.

COST

Latency and token limits set the bill

See when caching, batching and routing pay off.

AGENTS

Agents multiply everything

One request, many calls. Small errors compound.

RELIABILITY

Providers go down too

Your fallback is what users see.

CAREER

Interviews moved with it

Expect LLM serving, RAG and agents.

Your AI coach

You design.
The AI coaches.

Silent until you ask. Then a nudge, not the answer.

Solve today's problemTry a mock interview
  • ?
    Hints, one at a timePoints at the red card, never at the fix.
  • !
    Explains what brokeAnswered from the run’s own numbers.
  • ✓
    Grades like an interviewerEvery test case, hidden ones too. Optimal or over-built?
ScaleMySystems AIOne server on launch day
0/2 goals Success 90.3% · p99 580 ms
give me a hint
Hint 1 of 3: look at which card turns red first.
scale the API to 4 replicas
2/2 goals Success 100% · p99 32 ms
Submit
Accepted: 2 stars. Close to optimal: 4 replicas, 2 would carry this.
What's inside

A flight simulator for architecture

Any design, any scale, in your browser.

ENGINE

Per-request simulation

Queues fill, retries amplify, nodes crash.

VOICE

Talk to it

Say “add a cache”. It builds, reruns and reads the result back.

SECURITY

Attack your own system

DDoS, bots, credential stuffing: see what a WAF stops.

COST

Estimated cloud cost

Every fix shows its price.

What you'll practise

From a single cache to an agent harness

352 problems, 138 components, four paths: fundamentals, interviews, AI infrastructure, Kubernetes.

Traffic and compute 30

Load balancerAPI gatewayCDNApp serviceServerlessService meshCircuit breakerGraphQLgRPCAutoscaling

Data and messaging 23

CacheSQL databaseKey-value storeMessage queueEvent streamObject storageSearch indexOutbox / CDCStream processorWarehouse

AI pipeline 33

Semantic cacheEmbedding modelVector DBLLM endpointAI gatewayGuardrailsModel routerHybrid searchPrefill / decode poolsKV cache tierSpeculative decodingMulti-LoRAGraphRAG

Agents 22

Agent orchestratorMCP toolsCode sandboxHuman approvalAgent memoryA2A agentsSpend governorAgent harnessKV-aware routerTraffic mirrorToken brokerSubagentsComputer useCheckpoints

Security 15

DDoS protectionWAFBot managementmTLS proxyZero-trust proxySIEMFraud detectionEgress proxy

Design a product 56

YouTubeNetflixBookMyShowAmazon S3WhatsAppUberStripe paymentsGoogle DocsDropboxTwitter timelineZoomStock exchangeWeb crawlerChatGPTRAG searchCoding agent

Problems 243

One server on launch dayCold cache after a deploySlow database rescuePayment webhooksURL shortenerEvent pipelineRAG assistantEdge defenceAgent runawayYouTube end to endLLM gateway + fallbackMCP agents
Pricing

Practise free. Pay for the expensive parts.

Unlock a problem once; everything inside it is free. 6 problems are free, and new accounts get 100 welcome credits.

What a problem costs

“One server on launch day” and 5 other starter problemsfree
Unlock a Medium problem, once15 credits
Everything inside it: runs, changes, components, hints, Submit, resultsfree
One question to the AI (hints and “what broke?” are free)3 credits
18 of the 100 welcome credits a new account gets18 credits
FREE
$0
  • 6 starter problems, free
  • 100 welcome credits when you sign in with Google
  • Everything inside a problem you open is free
160 CREDITS
$5
  • About 10 Medium problems
  • Or about 26 components on a blank canvas
  • AI coach included: about 53 AI answers
Buy →
400 CREDITS
$10
  • About 26 Medium problems
  • Or about 66 components on a blank canvas
  • AI coach included: about 133 AI answers
Buy →
1,100 CREDITS
$25
  • 10% bonus credits
  • About 73 Medium problems
  • Or about 183 components on a blank canvas
  • AI coach included: about 366 AI answers
Buy →
2,400 CREDITS
$50
  • 20% bonus credits
  • About 160 Medium problems
  • Or about 400 components on a blank canvas
  • AI coach included: about 800 AI answers
Buy →
Easy problem 10Medium problem 15Hard problem 25Solved design 10AI answer 3AI review 20

Top-up calculator → · Full price list Prices in US dollars; rupees in India.

Start in one click

Break your first design today

Build from scratch Solve a problem GContinue with Google

6 free problems and 100 welcome credits when you sign in with Google.