eBay System Design Interview Questions

Use a token bucket algorithm backed by a distributed in-memory store. Apply limits per API key, IP, and user account. Rate limiting should happen at the API gateway, with local counters for speed and occasional sync to a central store for accuracy. Provide clients with clear rate-limit headers so they can self-throttle.

Adopt defense-in-depth: encrypted transport (TLS), encrypted data at rest, fine-grained RBAC, and service-to-service authentication via mTLS. Monitor for anomalies using SIEM tools. Apply security patches via rolling updates, and segment services by network boundaries. Implement rate-limiting and WAF rules to mitigate abuse.

Use a write-optimized store (like Cassandra) to handle high-frequency votes. Partition by poll ID to avoid hotspots. Maintain a real-time counter in Redis for instant result display, and periodically sync to a persistent store for durability. Implement vote validation to prevent duplicates or fraud.

Adopt an event-driven model: business events (order shipped, price drop) are published to a message bus, consumed by a notification service, and delivered via email, SMS, or push. Store user preferences for channels and quiet hours. Ensure delivery confirmation and retries for reliability.

Use leader-based coordination for each auction to maintain strict ordering of bids. Store auction state in-memory (e.g., Redis) with replication to prevent loss. Apply optimistic concurrency to reject late or invalid bids. Push real-time bid updates to clients via WebSockets. On auction close, commit final results to a durable store.

Precompute item similarity graphs offline using collaborative filtering or embedding models. Store results in a low-latency key-value store (e.g., Redis, RocksDB). Partition by tenant ID and product ID. Serve requests directly from this store to meet sub-100 ms latency, refreshing recommendations periodically.

Stream transaction and refund events into a processing engine (Flink/Kafka Streams). Apply statistical or ML models to detect deviations from baseline refund rates. Store anomalies in a dashboard-friendly store like Elasticsearch and alert via Slack/email. Include drill-down capabilities for root-cause analysis.

Use a row-store (e.g., Postgres) for transactional chat messages with indexes on conversation ID and timestamp. Stream inserts to a column-store (e.g., ClickHouse) for analytics. The row-store handles live chat and retrieval, while the column-store supports aggregated reports without impacting real-time performance.

Split into metadata service (file structure, permissions) and blob storage service (file chunks). Store chunks in object storage, with deduplication via content hashes. Use delta-sync for efficient updates and support version history. Implement end-to-end encryption for private files. For collaboration, use file-locking or CRDT-based merge strategies.

Break into catalog, search, cart, order management, payment, and fulfillment services. Use event-driven architecture so updates (e.g., inventory changes) propagate quickly. Ensure high availability via multi-region deployment, and integrate fraud detection early in the purchase flow. Caching and pagination are critical for high-volume product queries.