JavaScript System Design Interview Questions

Use cluster to scale a server across cores as multiple processes sharing one port (isolation, simple scaling). Use worker_threads for CPU-bound tasks within a process when you need shared memory or low-latency messaging.

Watch heap growth with diagnostics, remove lingering listeners/timers, and bound caches (LRU/TinyLFU). Track GC pauses and allocation rates; use heap snapshots in staging, and alert on upward trends.

Use Promises/async/await for clear request–response flows. Use Streams for large/continuous data with backpressure (file uploads, ETL). Mix them: wrap streams in promises for lifecycle control.

Define idempotent handlers, per-job retries with jitter/backoff, dead-letter queues, and visibility timeouts. Expose metrics (latency, retries, DLQ size) and use priority lanes for urgent work.

Clients send an idempotency key per command; the server stores key→result and returns the prior result on retries. Combine with at-least-once delivery and idempotent writes to approximate exactly-once.

REST: simple, cacheable, web-friendly. GraphQL: client-driven shapes across many views. gRPC: low-latency, typed, great for internal microservices/streaming. Many stacks expose REST externally and use gRPC internally.

Use a WebSocket server for exchanging SDP offers/answers and ICE candidates, integrate STUN/TURN, auth participants, and scale rooms with Redis pub/sub or a broker. Persist minimal session state and support reconnect/resume.

Choose SQL for strong consistency, joins, and transactions; choose NoSQL for flexible schemas, high write throughput, or massive key/value workloads. Decide by access patterns, consistency needs, and operational maturity.

Version APIs with OpenAPI, validate payloads with JSON Schema (AJV/Zod), and run migrations (Prisma/Knex) with forward/backward compatibility. Enforce consumer-driven contracts in CI to prevent breaking changes.

Pick DB-per-tenant, schema-per-tenant, or row-level tenant_id. Add row-level security, per-tenant rate limits/quotas, and partitioning to reduce noisy-neighbor issues.

Assume at-least-once across networks. Achieve effective exactly-once with idempotent endpoints, dedupe keys, and outbox patterns tying DB writes to event emission.

Blue-green: flip all traffic at once for fast rollback (simple changes, low risk). Canary: shift gradually with SLO guards and feature flags (risky changes, heavy traffic, many clients).

Store per-key state (in-process or Redis+Lua), implement tryAcquire() with tokens/time, support bursts and Retry-After, and export metrics. For multi-node fairness, keep state in Redis.

Use WebSockets (Socket.IO) with rooms, Redis adapter for horizontal fanout, and persistence for message history. Add delivery receipts, rate limits, and backpressure on hot rooms.

Create base62 IDs, store target+TTL, serve via CDN/edge cache, and stream click events to analytics (queue/Kafka). Use negative caching for misses and soft-delete expired links.

Issue presigned PUT URLs from the API, upload directly to object storage, then trigger processing via webhook/queue (validation, transcoding, thumbnails). Report status via polling or push; secure with HMAC-signed callbacks.