A traffic ramp after a quiet period saturates a managed network component that scales automatically but not fast enough. Degradation cascades. Analyse.

What the interviewer is testing

Whether you recognise that a managed component's scaling behaviour is part of your capacity model.

The shape

Traffic ramped steeply after a period of unusually low volume. A managed component scaled — but its ramp rate could not match the traffic's. Saturation produced packet loss, which caused health checks to fail for instances that were actually healthy, which caused service discovery churn and redistributed load onto already-degraded components.

A network capacity problem became a systemic one. This is the shape of Slack's January 2021 outage, and their published postmortem is unusually detailed.

The findings

Autoscaling has a reaction time, and that reaction time is a design assumption. A capacity plan depending on scaling must have enough standing capacity to survive the ramp until scaling catches up.

The worst ramps follow troughs. Going from a holiday or weekend low to a Monday peak is a harder engineering problem than sustained high traffic, because everything has scaled down and has the least standing capacity precisely when demand arrives.

Managed components scale according to their own rules. Their ramp rate is part of your capacity model and is frequently undocumented until you exceed it. It should be measured, not assumed.

Health checks that fail under network degradation amplify. An instance removed because a probe was dropped concentrates load on the remainder, worsening congestion. Health checking must distinguish a failed instance from a failed network path.

The changes

Pre-scale ahead of predictable events — the first Monday of a year, a marketing launch, a seasonal peak. Unglamorous and repeatedly cheaper than relying on reaction.

Set scale-down floors so the estate does not shrink to a level it cannot ramp from.

Measure the managed component's actual ramp rate with a load test, so it is a known number.

Make health checks resilient to transient network degradation — require multiple consecutive failures, and compare against peers rather than treating each in isolation.

What a strong answer adds

The general principle: model the worst ramp, not the peak. Capacity planning that examines only the maximum sustained load misses the transition, and the transition is where systems break.

Common weak answers

Increasing the scaling limits, which does not change the ramp rate. Blaming the managed service.