Google Backend Engineer Interview Questions

I would start by clarifying whether we need custom aliases, expiration, analytics, and abuse protection, because each of those changes the design. My first version would store a mapping from short code to long URL in a durable database and use a fast read path for redirects. To scale to billions of links, I would shard the storage, cache hot redirects, and make the short-code generation collision-safe. The reason I would separate reads from writes is that redirect traffic is usually much heavier than link creation traffic. I would also think about rate limiting, deduplication, and how to recover if one shard goes down.

I would first ask what is being cached, because session data, computed results, and database rows all have different consistency needs. My default design would use sharding to spread keys across nodes, with replication if availability matters a lot. I would keep the cache lookup path very fast and choose an eviction policy based on access patterns, usually LRU or LFU. The trade-off is that stronger consistency needs more coordination, while a looser cache is easier to scale and faster to read. I would also cover hot keys, invalidation, and what the system should do when a cache node fails.

I would first look at retries, timeouts, and whether the consumer is at-least-once rather than exactly-once. In most systems, duplicates come from a retry path or from reprocessing after a failure, so I would trace the event ID through the pipeline. To fix it, I would make the operation idempotent and store a deduplication key or processed-event record. The reason I choose that direction is that retries are unavoidable in distributed systems, so the system has to tolerate duplicates safely. I would also think about how long deduplication records should be kept and how that affects storage.

I would put a rate limiter at the edge, usually in the API gateway or service middleware, so the service can reject or slow traffic before it overloads the backend. My first choice would be a token bucket because it allows bursts while still enforcing long-term limits. I would keep the limiter state in a fast shared store if multiple servers need the same policy. The trade-off is between strict fairness and user experience, and I would usually favor a design that handles short spikes gracefully. I would also talk about what happens if the limiter store is down and whether the system should fail open or fail closed.

I would use Dijkstra's algorithm. I would keep a priority queue of the next node to explore and repeatedly relax outgoing edges from the node with the smallest current distance. The reason I choose Dijkstra is that it gives the correct shortest path efficiently when all weights are non-negative. The time complexity is O((V + E) log V) with a binary heap implementation. If the interviewer adds negative weights, I would say Dijkstra is no longer safe and I would need a different algorithm.