HLD Interview Questions — Study Guide Index
Deep-dive answers to all 50 HLD interview questions, organized by topic.
Documents
| # | Document | Topics | Questions |
|---|---|---|---|
| 01 | Architecture Overview & Core Components | System layers, WebSocket, API Gateway, auth, rate limiting, Lambda vs ECS, monolith vs microservices | Q1–Q5, Q21, Q40 |
| 02 | Intent Classification & Orchestration | Intent catalog, DistilBERT classifier, fan-out routing, adding new intents | Q6, Q11, Q16, Q39 |
| 03 | Conversation Memory & Context Management | DynamoDB schema, TTL, multi-turn context, circuit breaker, ElastiCache hot path | Q7, Q13, Q23 |
| 04 | RAG Pipeline & LLM Response Generation | Offline indexing, OpenSearch, Bedrock APIs, hallucination prevention, model adapter | Q8, Q9, Q18, Q22, Q24, Q25 |
| 05 | Recommendations, Personalization & Caching | Amazon Personalize, cold start, Redis caching, cache invalidation, feedback loop | Q10, Q16, Q29 |
| 06 | Scalability, Performance & Cost | Traffic spikes, latency optimization, cost breakdown, multi-storefront | Q19, Q26, Q27, Q35 |
| 07 | Fault Tolerance & Reliability | Circuit breaker, graceful degradation tiers, 99.95% SLA, chaos engineering | Q17, Q23, Q34, Q36 |
| 08 | Security, Safety & Guardrails | Guardrails pipeline, Bedrock Guardrails, prompt injection defense, PII, GDPR delete | Q14, Q28, Q30, Q37 |
| 09 | Analytics & Observability | Kinesis pipeline, 4-tier metrics, feedback loop, A/B testing prompts | Q15, Q20, Q33 |
| 10 | Testing & Deployment Strategy | 9-layer test strategy, golden set eval, chaos tests, LLM canary deployment | Q31, Q38 |
| 11 | Architect-Level Strategy & Business Alignment | Flywheel, build vs. buy, ROI, 3-year evolution, competitive moat, shutdown criteria | Q41–Q50 |
Questions by Difficulty
Easy (Q1–Q10)
| Q | Question | Document |
|---|---|---|
| Q1 | What is the overall architecture of MangaAssist? | 01 |
| Q2 | Why use WebSocket instead of REST for the chat interface? | 01 |
| Q3 | How does the system authenticate users? | 01 |
| Q4 | What is the role of API Gateway? | 01 |
| Q5 | How does rate limiting work? | 01 |
| Q6 | How does the intent classifier work? | 02 |
| Q7 | How does conversation memory work? | 03 |
| Q8 | How does the RAG pipeline work? | 04 |
| Q9 | How does the LLM generate a response? | 04 |
| Q10 | How do recommendations work? | 05 |
Medium (Q11–Q25)
| Q | Question | Document |
|---|---|---|
| Q11 | How does the orchestrator fan out requests? | 02 |
| Q12 | How does the system handle streaming responses? | 01 |
| Q13 | How does circuit breaker pattern prevent cascading failures? | 03 |
| Q14 | How does the guardrails pipeline work? | 08 |
| Q15 | How does the analytics pipeline work? | 09 |
| Q16 | How does caching work end-to-end? | 05 |
| Q17 | What happens if the order service goes down? | 07 |
| Q18 | How is OpenSearch populated and kept up to date? | 04 |
| Q19 | How does the system handle a 10x traffic spike? | 06 |
| Q20 | How are user feedback signals collected and used? | 09 |
| Q21 | How does the system support multiple storefronts? | 01 |
| Q22 | How does the system prevent hallucinated product recommendations? | 04 |
| Q23 | What happens if DynamoDB is unavailable? | 03 |
| Q24 | How do you choose between different LLM models? | 04 |
| Q25 | How does RAG handle product catalog updates? | 04 |
Hard (Q26–Q38)
| Q | Question | Document |
|---|---|---|
| Q26 | How do you optimize end-to-end latency? | 06 |
| Q27 | What is the cost per conversation? | 06 |
| Q28 | How do you protect against prompt injection? | 08 |
| Q29 | How do async patterns improve performance? | 05 |
| Q30 | How does the system handle PII? | 08 |
| Q31 | How do you roll out a new LLM version safely? | 10 |
| Q33 | How does A/B testing work for prompts? | 09 |
| Q34 | How do you achieve 99.95% SLA? | 07 |
| Q35 | How does the architecture scale from 100K to 10M conversations/day? | 06 |
| Q36 | How does chaos engineering work? | 07 |
| Q37 | How do you implement GDPR right-to-be-forgotten? | 08 |
| Q38 | What is the end-to-end testing strategy before launch? | 10 |
| Q39 | How do you add a new intent without breaking existing ones? | 02 |
Architect Level (Q40–Q50)
| Q | Question | Document |
|---|---|---|
| Q40 | How would you design this differently if starting over today? | 01 |
| Q41 | What flywheel effects does this create? | 11 |
| Q42 | Chatbot vs. improved search — which is more valuable? | 11 |
| Q43 | What are the three biggest risks? | 11 |
| Q44 | How do you measure ROI? | 11 |
| Q45 | How does the architecture evolve over 3 years? | 11 |
| Q46 | How do you defend against a manga-specialized competitor? | 11 |
| Q47 | What organizational challenges did this face? | 11 |
| Q48 | Which 3 intents would you launch with first? | 11 |
| Q49 | Build vs. buy — which components? | 11 |
| Q50 | When would you shut this project down? | 11 |
Quick Reference — Key Services Map
| AWS Service | Used for | Document |
|---|---|---|
| Amazon Bedrock (Claude 3.5) | LLM response generation | 04 |
| Amazon SageMaker | Intent classifier hosting (DistilBERT) | 02 |
| Amazon DynamoDB | Conversation history, session state | 03 |
| Amazon OpenSearch Serverless | RAG vector index | 04 |
| Amazon ElastiCache (Redis) | Hot path caching | 05 |
| Amazon Personalize | Product recommendations | 05 |
| Amazon Kinesis | Analytics event streaming | 09 |
| Amazon Redshift | Long-term analytics | 09 |
| Amazon ECS Fargate | Core orchestration service | 01 |
| AWS Lambda | Event processors, lightweight handlers | 01 |
| AWS API Gateway | WebSocket management | 01 |
| AWS Step Functions | Complex multi-step workflows | 02 |
| Amazon Connect | Human agent escalation | 07 |
| AWS Fault Injection Simulator | Chaos testing | 10 |