MicrosoftThe multi-tenancy solution for deploying ML pipelines across multiple workspaces, each potentially belonging to different Azure subscriptions, involves leveraging Azure Machine Learning (AzureML) Registries as a central hub for sharing data, models, environments, and components among tenants while maintaining version control.
In this setup, each tenant operates within their own workspace using their subscription, ensuring autonomy over computing and storage resources. If a tenant wishes to share resources with another tenant, they first share them with the AzureML registry ("share"). The registry then facilitates the distribution of these resources to other tenants ("push"). Tenants can also retrieve shared resources from the registry as needed ("pull").
This approach accommodates various scenarios:
1. Tenants can keep their models exclusive without sharing.
2. All tenants have the option to share their models.
3. Tenants can pull shared models for retraining or fine-tuning, then share back the updated models.
Sharing, pushing, and pulling of data, models, environments, and components between tenants and the registry can be managed through the AzureML UI and SDK. For example, sharing a model from a workspace to the registry can be done through the AzureML UI, while pulling resources from the registry can be achieved via SDK.
The flexibility of this solution enables tenants to build their ML pipelines using shared resources while utilizing their own resources for execution. Additionally, tenants can access shared resources through AzureML Designer, allowing them to filter and select resources based on the registry creator and incorporate them into their pipelines seamlessly.
Furthermore, tenants can deploy shared models as real-time or batch endpoints in their workspaces for inference tasks. This deployment process involves retrieving the model from the registry and then initiating inference using the deployed endpoint.
Overall, this multi-tenancy solution empowers Solution Providers to efficiently manage ML pipelines across diverse tenant environments, ensuring flexibility, autonomy, and secure resource sharing.
