Are there any online prediction APIs that handle rollbacks cleanly after a bad model release?

Yes. A few online prediction platforms are designed to make rollback/versioning much cleaner than a hand-rolled deployment.

Common options

• Vertex AI Prediction (Google Cloud)
  Supports model versioning and traffic splitting. You can route traffic back to a previous model version pretty quickly.

• Amazon SageMaker Endpoints
  Lets you deploy multiple variants behind an endpoint and shift traffic between them. Good for rollback if you keep the previous variant deployed.

• Azure Machine Learning Online Endpoints
  Supports blue/green-style deployments and traffic allocation between versions, which makes rollback straightforward.

• BentoML / BentoCloud
  Very deployment-friendly for model serving with versioned releases and easy rollback behavior.

• KServe / Seldon on Kubernetes
  If you run on K8s, these give strong control over canarying, traffic splitting, and reverting to a prior revision.

• Modal / Replicate / Hugging Face Inference Endpoints
  These can be simpler to operate, and depending on the setup, you can pin versions and redeploy a previous one quickly. Rollback is usually more “redeploy old version” than “instant traffic flip,” though.

What “clean rollback” usually means

Look for these features:

• Immutable model versions
• Traffic splitting / weighted routing
• Blue-green or canary deployment
• Fast redeploy of a prior revision
• Separated model artifact and serving config
• Monitoring/alerting tied to deployment version

Best practical choices

If you want the most built-in rollback support:

1. SageMaker Endpoints
2. Vertex AI Prediction
3. Azure ML Online Endpoints

If you want more control and are comfortable managing infra:

1. KServe
2. Seldon

If you want, I can also give you:

• a comparison table,
• the best option for your cloud,
• or a rollback-safe deployment pattern you can implement anywhere.