Google Backend Engineer Interview Questions

I would first check whether the issue is in the producer, the transport layer, or the ingestion pipeline. Late or out-of-order logs often come from batching, retries, clock skew, or asynchronous delivery, so I would inspect timestamps and delivery guarantees end to end. To improve the system, I would keep logging asynchronous but add sequence metadata or event IDs where ordering matters. The reason I would do that is that logging should stay cheap for the application, but still be reliable enough to debug real issues. I would also think about buffering, backpressure, and what to do when the downstream store is temporarily unavailable.

I would first identify whether the slowdown is from prefix lookup, ranking, or serving traffic. For a hot query, I would cache the result aggressively and make sure the prefix index can handle repeated reads without rebuilding the answer every time. I would also consider local aggregation so the system does less repeated work at the backend. The reason I choose caching is that autocomplete traffic is often concentrated on a small number of prefixes. I would also mention cache invalidation and how to keep results fresh enough without hurting latency.

I would separate metadata from file contents. The file bytes would go into object storage, while a database would manage names, owners, permissions, and lookup metadata. For fast reads, I would add caching and a CDN if files are frequently downloaded from many locations. The trade-off is between simplicity and performance, but separating metadata and content usually makes the system easier to scale. I would also talk about resumable uploads, access control, and failure handling for partial writes.

I would first clarify delivery priority, user preferences, and whether the same event should go to multiple channels. My design would have an event intake layer, a routing layer, and separate workers for each delivery channel. I would make the delivery asynchronous so a slow email provider does not block push or in-app delivery. The reason I would do it this way is that each channel has different reliability and latency behavior. I would also think about retries, deduplication, and user opt-out rules.

I would first look at the balancing policy, health checks, and whether the traffic is sticky to a single session or key. If one instance is hot, it might be because of a bad hash, uneven weights, or a long-lived sticky session. I would also check whether the supposedly healthy instances are actually slower and being avoided. The reason I start there is that load balancing problems usually come from routing policy or health perception, not just raw traffic volume. I would also mention connection draining and failover behavior.