Before promoting, decide whether load() must re-run: if yes, enable Re-deploy when promoting on the environment, otherwise you can promote and still be running the old loaded weights. During a rolling deployment, do not panic about replica counts, autoscaling is suspended until the rollout finishes. Never start a second promotion while one is in flight, the environment allows only one active promotion. Keep development deployments warm if you want to avoid scale-to-zero cold starts in the dev loop.

Context: Baseten org reference notes on the deployment lifecycle document the promotion traps agents hit. Promotion may or may not create a new deployment, so if your new code needs load() to re-run, you must enable Re-deploy when promoting on the environment or the new build never actually loads the model fresh. Rolling deployments suspend autoscaling for the whole environment for their entire duration, which is why replica counts look wrong mid-rollout. Only one active promotion per environment is allowed at a time.