Amazon Backend Engineer Interview Questions

I would use an idempotency key and associate it with the operation result or current state. When the same key arrives again, the service would return the earlier result instead of doing the work twice. The reason I choose this is that retries are expected in distributed systems, and correctness should not depend on perfect network delivery. I would also think about expiration for stored keys and how to handle cases where the first request is still running. This pattern is especially important for checkout, payments, and other state-changing actions.

I would start by clarifying whether the balancer needs session stickiness, health checks, or geographic routing. A basic version could use round robin, but in practice I would also consider least connections or weighted routing depending on the traffic pattern. The reason I keep the balancer stateless when possible is that it is much easier to scale and recover. I would discuss failover, connection draining, and avoiding unhealthy instances. If the interviewer wants more detail, I would also mention how to handle long-lived connections.

I would use Dijkstra's algorithm. I would maintain a priority queue of the next node to visit and relax the edges from each node as I pull it from the queue. The reason I choose Dijkstra is that it is the standard shortest-path solution when edge weights are non-negative. The time complexity is O((V + E) log V) with a binary heap. If the interviewer introduces negative edges, I would say I need a different algorithm because Dijkstra would no longer be safe.

I would use a durable streaming layer so producers can write quickly without waiting for downstream processing. The events would be partitioned by a stable key so consumers can process them in parallel. The reason I choose streaming ingestion is that it scales far better than sending every event directly into a database. I would also discuss replay, backpressure, and schema evolution because event streams tend to grow and change over time. If the interviewer asks for an end-to-end flow, I would explain how the raw events get turned into aggregates later.

I would store the file contents in object storage and keep the metadata in a database. For large uploads, I would support chunked or resumable uploads so users do not have to start over if a transfer fails. The reason I choose this split is that it keeps the backend scalable and makes metadata queries much easier. I would also think about signed access links, virus scanning, and CDN support for popular files. If the interviewer wants more depth, I would talk about deduplication and partial failure cleanup.