2 / 4 Deploys, Canary & Rollback

Warm Against Cold

You deploy a canary: 3 freshly started instances of the new build take 2 percent of traffic while 150 long-running instances of the old build take the rest. The canary shows p99 latency 40 percent worse than the fleet average, but its CPU and error rate look normal. What is the most likely explanation, and the standard fix for the comparison?

A

The load balancer sends newer instances a costlier mix of requests, so pin the canary to a fixed set of endpoints

B

Two percent of traffic is too small a sample to measure p99 at all, so raise the canary to 50 percent and re-read it

C

The new build is genuinely slower, because latency regressions reliably show up before error-rate regressions do

D

The canary instances are cold, so compare them against a baseline group of old-build instances started alongside them