❓ System Design Concepts - Q&A¶

Pairs with System Design Guide, components/, and Distributed Systems. Format and tag legend: Overview.

Q1. Horizontal vs vertical scaling? Vertical scaling adds power to one machine (simple, but hits a hardware ceiling and is a single point of failure). Horizontal scaling adds more machines (near-unlimited, needs load balancing and handles failure better). Distributed systems favor horizontal. Follow-ups: What makes a service easy to scale horizontally? (statelessness.) Easy · MAANG

Q2. Why do stateless services scale better? Any instance can serve any request, so you can add/remove instances freely behind a load balancer and lose one without losing session data. State is pushed to shared stores (DB, cache, session store) instead of living in the app server. Follow-ups: Where does session state go then? (Redis/DB, or a signed token.) Med · MAANG

Q3. How do you do a back-of-the-envelope estimate? State assumptions (users, actions/user/day), derive QPS = actions/day ÷ ~86,400, apply a ×2-3 peak factor, then size storage (bytes×items×retention×replication) and bandwidth (QPS×payload). Order-of-magnitude is the goal. See Latency Numbers & Estimation. Follow-ups: 1M writes/day is roughly how many per second? (~12/s before peak.) Med · MAANG

Q4. Explain the CAP theorem and how it guides design. During a network partition you must choose consistency or availability. A payment system leans CP (reject rather than serve wrong data); a social feed leans AP (stay up, tolerate staleness). PACELC adds a latency-vs-consistency trade-off even without partitions. Follow-ups: Classify a system you know as CP or AP. Hard · MAANG

Q5. L4 vs L7 load balancing? L4 balances on TCP/UDP (IP + port) - fast, protocol-agnostic. L7 balances on application data (HTTP path, headers, cookies) - enables routing, TLS termination, and content-based rules at higher cost. See Load Balancer. Follow-ups: When do you need L7? (path-based routing, sticky sessions, header rules.) Med · MAANG

Q6. What load-balancing algorithms exist? Round robin, weighted round robin, least connections, IP/consistent hash (session affinity), and latency-based. Choice depends on whether backends are uniform and whether you need stickiness. Follow-ups: Why might least-connections beat round robin? (uneven request cost.) Med · MAANG

Q7. Why use a message queue? To decouple producers from consumers, absorb spikes (buffer), enable async processing, and provide retries/backpressure. It trades immediate consistency for resilience and throughput. See Message Queues. Follow-ups: Queue vs pub/sub? (point-to-point work vs broadcast to many subscribers.) Med · Amazon

Q8. At-least-once vs exactly-once in a queue? Most queues give at-least-once (retries can duplicate), so consumers must be idempotent. Exactly-once is hard end-to-end; it's usually at-least-once delivery + idempotent processing/dedup keys. At-most-once drops on failure. Follow-ups: How do you make a consumer idempotent? (dedup key, upsert, idempotency table.) Hard · Amazon

Q9. How does Kafka scale and preserve ordering? A topic is split into partitions distributed across brokers; consumers in a group each own some partitions (parallelism). Ordering is guaranteed only within a partition, so you key messages that must stay ordered onto the same partition. Follow-ups: What limits your consumer parallelism? (partition count.) Hard · Amazon

Q10. What is consistent hashing and where is it used? Placing nodes and keys on a hash ring so adding/removing a node only remaps ~K/N keys instead of nearly all. Virtual nodes even out load. Used by caches, DynamoDB, Cassandra. See Hashing. Follow-ups: What problem does it solve vs hash mod N? Med · Amazon

Q11. Sharding vs replication - when do you need each? Shard when data or write volume exceeds one node (splits different data). Replicate for availability and read scaling (copies the same data). They combine: shard for scale, replicate each shard for resilience. See Database Sharding. Follow-ups: Why can't replicas scale writes? Med · MAANG

Q12. How do you generate unique IDs at scale? Options: UUIDv4 (random, no coordination, but not sortable and scatters index writes), Snowflake (64-bit timestamp+machine+sequence, time-sortable, needs worker IDs), or DB ticket/range servers. Snowflake is the common answer. Follow-ups: Why are time-sortable IDs nice for a B-tree index? (sequential inserts, good locality.) Hard · Amazon

Q13. Fan-out on write vs fan-out on read (news feed)? Fan-out on write pushes each post into followers' feeds at write time (fast reads, expensive for celebrities). Fan-out on read builds the feed at read time (cheap writes, slower reads). Real systems use a hybrid - precompute for most, read-time merge for high-follower accounts. Follow-ups: Why is a celebrity with 100M followers a problem for fan-out on write? Hard · Meta

Q14. What's a CDN and when does it help? Edge caching of content close to users to cut latency and offload the origin. Great for static assets and cacheable responses; also absorbs spikes and adds DDoS protection. Less useful for highly personalized/dynamic data. Follow-ups: How does content get invalidated at the edge? (TTL, purge/invalidation API.) Med · MAANG

Q15. What is a rate limiter and which algorithm would you pick? Caps requests per client per window to prevent abuse/overload. Token bucket allows bursts, leaky bucket smooths, sliding-window counter balances accuracy and memory. Store counters in Redis for a distributed limit; enforce at the gateway and/or per service. Follow-ups: How do you rate-limit across many app servers? (shared Redis counter, atomic incr.) Med · MAANG

Q16. How do you design for high availability? Remove single points of failure via redundancy (multiple instances/AZs/regions), health checks + failover, load balancing, replication, and graceful degradation. Measure with availability "nines" (99.99% ≈ 52 min downtime/year). Follow-ups: What's the difference between failover and load balancing? Med · MAANG

Q17. What is a circuit breaker? A resilience pattern: after repeated failures to a dependency, "open" the circuit and fail fast (skip calling it) for a cooldown, then probe with a trial request before closing again. Prevents cascading failures and gives the dependency time to recover. See Microservices. Follow-ups: How does it differ from a retry? (retry hammers; breaker stops hammering.) Hard · Amazon

Q18. Strong vs eventual consistency - when each? Strong: every read sees the latest write (needed for balances, inventory, uniqueness). Eventual: replicas converge over time (fine for feeds, counts, catalogs) and buys availability + low latency. Choose per data type, not globally. Follow-ups: How do you make eventual consistency acceptable? (read-your-writes, UI cues.) Med · MAANG

Q19. What is idempotency and why does it matter in APIs? An idempotent operation has the same effect whether applied once or many times, so clients can safely retry after a timeout without double-charging or duplicating. Implement with an idempotency key the server dedups on. Follow-ups: Which HTTP methods are idempotent by spec? (GET, PUT, DELETE.) Med · Amazon

Q20. How do you approach a system design interview? Clarify functional + non-functional requirements, estimate scale, sketch a high-level design, then deep-dive the bottlenecks and trade-offs - talking aloud and starting simple before scaling. See the framework in System Design Guide. Follow-ups: What's the most common mistake? (jumping to a solution before clarifying scope.) Med · MAANG

Q21. How would you scale a read-heavy system? Add caching (CDN + Redis), read replicas, and denormalized/precomputed views; push static content to a CDN. Identify the read/write ratio first - read-heavy points to caching + replicas, write-heavy to sharding + queues. Follow-ups: What consistency cost do read replicas introduce? (replication lag.) Med · MAANG

Q22. What is a single point of failure and how do you find one? Any component whose failure takes down the system (one DB, one LB, one region). Find them by tracing the request path and asking "what if this dies?" Remove via redundancy and failover. Follow-ups: Is a single load balancer a SPOF, and how do you fix it? (yes - redundant LBs + DNS/anycast.) Med · Amazon