Metrics question

How would you measure reliability and speed in a way customers actually care about?

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Whether the candidate can translate raw infrastructure metrics into what customers actually experience and care about.

Clarify the gap: raw uptime percentage or average latency can look fine while customers still experience real pain, like tail latency spikes during their peak traffic.
Propose tail latency, like p99 response time, over average latency, since customers notice worst-case slowness far more than the mean.
Propose error rate measured during the customer's actual peak usage windows, not a 24-hour blended average that masks bad periods.
Propose a customer-facing composite, like percentage of requests completed within an agreed latency threshold, mirroring how an SLA would actually be evaluated.
Validate with qualitative input: cross-check these metrics against actual support tickets and churn reasons to confirm they correlate with real complaints.

Explicitly rejects average latency and blended uptime as insufficient, replacing them with tail latency and peak-window error rate, which better reflect real customer pain.
Proposes an SLA-style composite metric, like percentage of requests within a latency threshold, framed the way a customer would actually judge reliability.
Recommends validating the chosen metrics against actual support tickets and churn data, to confirm they are not just technically clean but genuinely predictive of customer sentiment.
Notes that reliability and speed must be measured per customer or per workload, since a global average can hide a badly served large account.

Proposing generic uptime percentage or mean latency without addressing tail behavior and peak-time performance.
Ignoring the need to validate metrics against real customer complaints or churn.

How would you set SLA thresholds differently for different customer tiers?
What would you do if p99 latency looked fine but a specific large customer still complained?