Asked about Keeping a model working after launch

How would you deploy a new model version safely?

An answer that works

Shadow deployment first, running it on live traffic without serving its output so you can compare per request rather than by distribution. Then a canary on a small share with guardrails, then a full rollout, with the previous version one alias move away.

Then they push further

What makes a rollback real rather than theoretical?

That it is one action, performed without a deploy, by someone who did not build the system. In practice that means the model is behind a registry alias. The test is not whether a plan exists but whether anyone has actually done it this quarter.

Reading an answer is not being able to give one. The section that teaches this is Keeping a model working after launch, 5 hours of MLOps & model lifecycle.

Learn this properly

Also asked about MLOps & model lifecycle

What is model drift and how do you detect it?What do you need to version to reproduce a model?What is training-serving skew?What does a feature store give you?How would you monitor a fraud model whose labels arrive 30 days late?How would you roll out a retrained model?