Interview Playbook System design interview strategy: level-by-level rubric, 45-minute timelines for HLD and LLD, communication phrases, common mistakes, a 60-question bank from beginner to advanced with pointers, a design prompt index, and 4-week and 7-day study plans.
The other pages teach the material.
This page is about performing under interview conditions: what is graded at each level, how to spend 45 minutes, what to say, which questions to practice, and how to schedule your preparation.
What Is Graded at Each Level The 45-Minute HLD Timeline The LLD and Machine Coding Timeline What to Say Common Mistakes Question Bank Design Prompt Index Study Plans Final Checklist
The same prompt is used across levels, and the bar rises.
Level Expected behavior Junior Understand the problem, propose a workable design with basic components (LB, app tier, DB, cache), explain choices when asked, accept guidance Mid (SDE 2) Drive the conversation, gather requirements, estimate scale, choose data stores with reasons, identify bottlenecks and fix them, discuss basic failure cases Senior Own the whole design. Present alternatives and pick one with trade-offs. Deep dive two components at implementation level. Cover failure modes, consistency, observability, deployment, cost. Staff and above Handle ambiguity and organizational constraints, define the problem before solving it, phase the roadmap (build now, evolve later), migration and rollout, cross-team impact, build vs buy, long-term operability, risk
Signals that raise your rating at any level:
You ask sharp clarifying questions and the answers change your design.
Numbers drive decisions (you say "so we need sharding" because of an estimate).
You name what you give up with each choice.
You spot failure modes before the interviewer does.
You can go deeper on any box on request.
Phase Do Avoid Requirements 3 to 5 functional, 3 to 5 non-functional with numbers, list what you exclude Starting to draw immediately Estimates QPS, storage, bandwidth, peak factor, one design decision from them Ten minutes of arithmetic API and data Endpoints, main entities, keys, indexes, read and write paths Vague "a database" boxes High level One diagram, trace one write and one read end to end Twenty components with no flow Deep dive The interviewer's choice, or your riskiest component: alternatives, algorithms, failure behavior Skimming everything equally Failures and scale SPOFs, hot spots, what if 10x traffic, what if a region fails, monitoring Ending without discussing failure Wrap up Recap, what you would build next, what you would monitor Trailing off
Use the seven-step framework as the backbone.
For a 60 to 90 minute coding round, see the detailed plan in LLD Fundamentals .
Minutes (of 60) Activity 0 to 8 Clarify, write scope, list entities and use cases 8 to 15 Class diagram, entry-point API, extension points 15 to 45 Code the core flow end to end, then the rest 45 to 55 Demo main, edge cases, fix bugs 55 to 60 Discuss extensions and concurrency
Priorities: working and clean beats clever and unfinished.
Say aloud where you would add a pattern, even if you do not have time to code it.
Situation Say Starting "Before I design, I want to confirm scope and scale." Making an assumption "I will assume 100M DAU and a 100 to 1 read to write ratio, tell me if that is off." Choosing a technology "I need X property, so I will use Y. The trade-off is Z." Comparing options "There are two approaches, A and B. I would pick A because..., B wins if..." Handling a bottleneck "The database write rate is the first thing to break, so..." Admitting uncertainty "I have not run this at that scale. My reasoning is..., and I would validate it with a load test." Interviewer pushes back "Good point, that changes X. If we need Y, I would switch to..." Running low on time "I will go deep on the write path and sketch the rest."
Make trade-offs explicit with these pairs:
Pair Typical sentence Consistency and availability "For the feed I choose availability, a slightly stale view is fine. For payments I choose consistency." Latency and cost "A cache tier costs money, and it takes p99 from 80 ms to 10 ms." Read cost and write cost "Precomputing the feed makes writes heavy and reads cheap." Simplicity and scalability "One Postgres primary is enough until roughly this size, so I start there." Sync and async "Emailing does not need to block the request, so it goes through a queue." Build and buy "Auth is commodity, I would use a managed identity provider." Accuracy and speed "Approximate counts with HyperLogLog are fine for analytics."
State the problem the component solves.
Give two or three options.
Pick one with reasons tied to requirements.
Say how it fails and what you do about it.
Say how you would observe it.
Mistake Why it hurts Fix No clarification You solve the wrong problem Spend 5 minutes on scope and numbers Buzzword architecture (Kafka, Redis, Kubernetes everywhere) No reasons, no fit Justify each component by a requirement Over-engineering Multi-region microservices for 1,000 users Right-size and state the growth path Ignoring numbers No basis for decisions Two or three key estimates Silent thinking Interviewer cannot help or grade Think aloud, structure as you go One happy path No failure handling Walk each component: slow, down, wrong Single point of failure left in Immediate red flag Name and remove SPOFs Hand-waving the data model Data is the design Tables or documents, keys, indexes, access patterns Not listening to hints Interviewer is steering you Treat hints as instructions Defensive on feedback Signals poor collaboration Adjust the design and explain the effect Skipping security and observability at senior level Missing half the job Add auth, rate limits, metrics, alerts, rollout LLD: coding before modeling Tangled classes Entities and interfaces first LLD: patterns for show Extra indirection Only at stated variation points
Sixty questions grouped by level.
The "Strong answer includes" column is a checklist, not a script.
# Question Strong answer includes Read 1 Vertical or horizontal scaling? Vertical first for simplicity, horizontal for redundancy and limits, stateless tier Fundamentals 2 What is a load balancer, and which algorithms exist? L4 vs L7, round robin, least connections, consistent hash, health checks Building Blocks 3 Cache-aside vs write-through? Read and write paths, staleness, write cost, delete on write Building Blocks 4 SQL or NoSQL? Access patterns, transactions, scale, schema flexibility, start with SQL Building Blocks 5 What is an index and what does it cost? B-tree, faster reads, slower writes, space, leftmost prefix Building Blocks 6 What does a CDN do? Edge caching, latency, offload, cache keys, invalidation by versioned URLs Building Blocks 7 Replication vs sharding? Copies for availability and reads, partitions for write and size scale Building Blocks 8 Explain CAP. Partition forces C or A, PACELC else latency vs consistency, examples Fundamentals 9 Latency vs throughput? Definitions, batching trade-off, percentiles Fundamentals 10 What is idempotency and why does it matter? Safe retries, idempotency keys, PUT vs POST Building Blocks 11 REST, GraphQL or gRPC? Public vs client-driven vs internal, trade-offs Building Blocks 12 Why use a message queue? Decoupling, spikes, retries, async work, at-least-once Building Blocks 13 Stateless vs stateful services? Any instance serves any request, sessions in shared store, state moves to data tier Fundamentals 14 What does DNS TTL control? Caching duration, failover speed vs load Building Blocks 15 Estimate QPS for 10M DAU with 20 requests each. 200M per day, about 2,300 per second average, peak 3x Fundamentals
# Question Strong answer includes Read 16 How do you invalidate and evict a cache? TTL, delete on write, LRU or LFU, CDC-driven invalidation Building Blocks 17 How do you handle a hot key? Local cache, replicate, key splitting, shard-aware limits Building Blocks 18 Why consistent hashing? Minimal remapping on resize, virtual nodes, examples Building Blocks 19 Fan-out on write vs read for a feed? Write amplification, celebrities, hybrid Large-Scale Designs 20 Can you get exactly-once delivery? At-least-once plus idempotent consumers, Kafka transactions Building Blocks 21 How do you pick a shard key? Cardinality, even load, query alignment, hot spots Building Blocks 22 How do you provide read-your-writes with replicas? Route to primary after write, log position, sticky replica Fundamentals 23 WebSocket, SSE or polling? Direction, cost, reconnect, fit by use case Building Blocks 24 Compare rate limiting algorithms. Token bucket, fixed window edge burst, sliding window, Redis atomicity Classic Designs 25 How do Kafka partitions, groups and ordering work? Ordering per partition, parallelism bounded by partitions, offsets Building Blocks 26 Offset or cursor pagination? Depth cost, shifting data, keyset, stability Building Blocks 27 What is a cache stampede and how do you stop it? Single-flight, jittered TTL, stale-while-revalidate, locks Building Blocks 28 Design an idempotent payment API. Idempotency key stored atomically, state machine, unknown outcome, reconciliation Large-Scale Designs 29 Keep a search index in sync with the DB. CDC or outbox, idempotent updates, rebuild path Building Blocks 30 Blue-green or canary? Cost, rollback speed, metric-gated rollout, DB compatibility Reliability and Operations
# Question Strong answer includes Read 31 Explain Raft leader election and commit. Terms, majority votes, log replication, up-to-date log rule Distributed Systems 32 2PC or saga? Blocking coordinator, compensations, isolation loss, idempotency Distributed Systems 33 What is split brain and how do you prevent it? Epochs, majority, fencing tokens, STONITH Distributed Systems 34 Isolation levels and write skew? Anomaly table, snapshot isolation, serializable, fixes Distributed Systems 35 Does W + R > N guarantee linearizability? No: sloppy quorums, concurrent writes, clocks, read repair Distributed Systems 36 How do you order events across machines? Not wall clocks, Lamport and vector clocks, HLC, per-partition sequence Distributed Systems 37 Design multi-region active-active data. Home region, multi-leader conflicts, CRDTs, global consensus cost, RPO and RTO Distributed Systems 38 Design for 99.99% availability. Error budget math, redundancy, no shared fate, deploy safety, tested failover Reliability and Operations 39 Why do retries cause outages? Amplification, metastable state, budgets, backoff with jitter, shedding Distributed Systems 40 What is a cell-based architecture? Independent cells, router, blast radius, shuffle sharding Distributed Systems 41 Explain the transactional outbox. Same-transaction event row, relay, at-least-once, inbox dedupe Distributed Systems 42 When would you use CRDTs? Convergent merge, offline, collaboration, counters and sets, limits Distributed Systems 43 Exactly-once in stream processing? Checkpoints, idempotent sinks, event time, watermarks, dedupe Large-Scale Designs 44 Zero-downtime schema migration? Expand, dual write, backfill, switch, contract Reliability and Operations 45 What did the October 2025 AWS outage teach? DNS race condition, hidden dependency, recovery collapse, guardrails Reliability and Operations
# Question Strong answer includes Read 46 How do LLM serving engines get throughput? Continuous batching, PagedAttention, prefix caching, chunked prefill, quantization AI System Design 47 A RAG bot gives wrong answers. Debug it. Eval set, retrieval first, chunking, hybrid, rerank, grounding AI System Design 48 Enforce document permissions in RAG. ACL metadata, filter inside retrieval, propagate deletes AI System Design 49 Defend an agent against prompt injection. Untrusted content, least privilege, confirmations, isolation, monitoring AI System Design 50 Halve an LLM bill. Measure tokens, trim prompts, prefix cache, route to smaller models, batch AI System Design 51 SSR, CSR, SSG or ISR? Chooser flow, SEO, personalization, hydration cost Frontend System Design 52 Improve LCP and INP. Image priority and CDN, split bundles, break long tasks, transitions Frontend System Design 53 Design a real-time dashboard UI. Batching, bounded buffers, downsampling, worker, backpressure Frontend System Design
# Question Strong answer includes Read 54 Make a parking lot extensible for new pricing and allocation. Strategy interfaces at the variation points, injected clock Machine Coding 1 55 Strategy or State? Who chooses, lifecycle, transitions, examples Design Patterns 56 Make an LRU cache thread safe. Lock, striping, atomic compute, library alternatives Machine Coding 1 57 Prevent double booking under concurrency. Atomic check and act, lock per show, hold with expiry, DB CAS Machine Coding 2 58 Refactor a class full of type switches. Polymorphism, Open/Closed, injected policies LLD Fundamentals 59 How do you avoid deadlock? Lock ordering, tryLock timeout, fewer locks, confinement Concurrency in LLD 60 Design elevator scheduling. SCAN with sorted stops, cost-based dispatch strategy, state machine Machine Coding 1
Practice by saying the design out loud with a timer.
Prompt Where Parking lot, LRU cache, elevator, vending machine, tic-tac-toe, logger Machine Coding 1 Movie ticket booking, Splitwise, pub/sub, rate limiter, ATM, file system Machine Coding 2 Extra practice with the same patterns Library management (Strategy fines, Observer reservations), chess (Strategy per piece, Command undo), snake and ladder (Strategy dice), ride booking (State, Strategy), food delivery (Observer, State), Twitter-lite (Observer, Facade), hotel booking (like movie booking with date ranges)
Week Learn Practice (out loud, timed) 1: Foundations Fundamentals , Building Blocks , LLD Fundamentals , Design Patterns Estimation drills (5 a day), URL shortener, rate limiter, LRU cache, parking lot 2: Classic designs Classic Designs , Machine Coding 1 Key-value store, notification system, typeahead, code elevator and vending machine 3: Depth Distributed Systems , Large-Scale Designs , Machine Coding 2 , Concurrency in LLD News feed, chat, payments, ticket booking, movie booking with concurrency, Splitwise 4: Modern and mocks AI System Design , Frontend System Design , Reliability and Operations Two full mock interviews per week (HLD and LLD), question bank review, record and critique yourself
Daily rhythm: 45 minutes reading, 45 minutes practicing a design out loud or coding an LLD problem, 15 minutes reviewing gaps.
Day Focus 1 Fundamentals sections 2, 4, 8 and the Building Blocks tables. Estimate drills.2 URL shortener, rate limiter, ID generator, key-value store. Say each in 30 minutes. 3 News feed, chat, notification system. Focus on fan-out, ordering, delivery. 4 Payments, ticket booking, distributed transactions and idempotency from Distributed Systems . 5 LLD Fundamentals , then code parking lot and LRU. Review the pattern chooser.6 One more LLD (movie booking or Splitwise) plus Concurrency in LLD answers. Skim the AI and reliability pages. 7 Mock: one HLD and one LLD with a friend. Review the question bank rows you missed. Sleep.
Before the interview:
Senior summary in one line: clarify, size it, design the simplest thing that meets the numbers, then find where it breaks and say what you would do about it.