The canary that only hurt the slowest requests

A recommendation service runs on ECS Fargate behind an Application Load Balancer, twelve tasks across three Availability Zones. Each request reads a user's feature vector from DynamoDB, calls an in-process model, and returns ranked items.

You deploy a new model version to 10% of tasks as a canary. The new model is architecturally identical to the old one — same framework, same input shape — but was trained on six months more data and has roughly 30% more parameters.

After the canary goes live: median latency is unchanged at 42ms. p99 latency rises from 180ms to 1,400ms. The p99 rise is visible on ALL tasks, not only the canary tasks. Error rate is unchanged. CPU utilisation across the fleet averages 40%. The DynamoDB table reports no throttling. Rolling back the canary returns p99 to 180ms within two minutes.

The p99 degradation appears on tasks running the OLD model as well as the new one. Give me your ranked hypotheses for how a canary on 10% of tasks degrades the other 90%, and for each one the specific measurement that would confirm or eliminate it.