Amazon Backend Engineer Interview Questions

I would first reproduce the crash with a small input and inspect the logs, stack traces, and job payloads. Then I would determine whether the failure is due to bad input, memory pressure, a race condition, or an unhandled exception. The reason I start with the crash path is that worker failures often come from a single unexpected edge case rather than the whole system design. I would also think about retry limits and whether the job should go to a dead-letter queue after repeated failures. Once I know the root cause, I would add validation or better error handling so the same job does not break the worker again.

I would separate file bytes from metadata. The file content would go into object storage, while metadata like owner, filename, permissions, and timestamps would live in a database. I would also think about upload size limits, resumable uploads, and signed URLs for secure downloads. The reason I choose this split is that it keeps the system scalable and makes lookup and authorization easier. I would also talk about antivirus scanning, content validation, and CDN support if the files are downloaded often.

I would first assume that duplicates can still happen because retries and consumer restarts are normal. To make the system safe, I would make the consumer idempotent and record which message IDs have already been processed. The reason I choose this approach is that trying to eliminate every duplicate at the transport layer is usually harder than making processing safe when duplicates appear. I would also think about deduplication windows and storage cost for the processed-ID log. If the interviewer wants more detail, I would explain how this differs for at-least-once versus exactly-once style processing.

I would first define what counts as activity, because that could be messages, requests, session time, or something else. My design would stream activity events into an aggregator, keep counts for the current time window, and maintain the top 10 with a small ranking structure. The reason I choose streaming aggregation is that it avoids recomputing the entire data set every minute. I would also think about sharding, merging partial counts, and how to handle late events. If freshness matters, I would keep the aggregation pipeline low latency.

I would place a rate limiter or gateway in front of the service so it can reject or slow excessive traffic before it reaches the backend. My first choice would be a token bucket because it supports bursts while still enforcing a long-term limit. The reason I choose this approach is that traffic spikes are common and the system should remain stable instead of collapsing under load. I would also discuss whether the limiter state should be shared across instances and what should happen if the limiter store is unavailable. If the interviewer asks, I would mention load shedding and circuit breakers as related protections.