AWS Site Reliability Engineer Interview Questions

I would describe the incident, the user impact, and the first signal that alerted the team. Then I would walk through how I diagnosed the issue, what mitigation I chose, and how I kept communication clear while the problem was unfolding. The reason I choose this structure is that AWS SRE interviews care a lot about ownership and incident response under pressure. I would also mention the postmortem actions and what changed afterward to reduce the chance of recurrence. If the interviewer wants more detail, I would explain what I would do differently next time.

An SLI is the metric that measures service behavior, such as latency or availability. An SLO is the target we want that metric to meet, while an SLA is the formal promise we make to customers. Error budget is the allowed gap between perfect reliability and the SLO target, and it helps balance release speed with stability work. The reason I choose to explain them this way is that the terms matter most when tied to operational decisions. If the interviewer wants more detail, I would give a concrete example from a service and show how the error budget affects behavior.

I would first determine whether the slowdown is in the application, database, network, or infrastructure layer. Then I would compare the issue against recent deploys, traffic changes, saturation metrics, and traces. The reason I choose this approach is that latency problems are usually caused by one or two bottlenecks rather than the whole stack. I would also think about whether the problem is isolated to one region, one endpoint, or one type of request. If the interviewer wants more detail, I would explain how I would narrow the problem step by step.

I would base alerts on symptoms that users actually feel, not just internal component metrics. Then I would use thresholds, burn-rate alerts, and deduplication so the team gets pages only when the service is truly at risk. The reason I choose this design is that noisy alerts make on-call less effective and reduce trust in the system. I would also think about severity levels, suppression during incidents, and periodic review of alerts that never lead to action. If the interviewer wants more detail, I would discuss how I would test alerts against past outages.

I would immediately reduce the blast radius by slowing retries, enabling backoff, or disabling the most aggressive retry path. Then I would check whether the dependency is actually recovering or whether the retries are just adding load. The reason I choose this approach is that uncontrolled retries can turn a partial failure into a full outage. I would also think about circuit breakers, timeouts, and fallback responses so the system does less work while the dependency is unhealthy. If the interviewer wants more detail, I would explain where retry logic should live and how I would prevent this from happening again.