Flyte Contributors Meetup - May 9, 2024

Project isolation implementation Robert reviewed the current implementation of the project isolation feature, which uses an interceptor to check the user's access permissions based on claims in the token. The team discussed potential weaknesses in the current approach, such as the need to manually maintain a list of all endpoints, and explored alternative solutions like using the user ID instead of claims. Extending project isolation to other dimensions Eduardo suggested extending the project isolation feature to support other dimensions like environment, in order to create a more comprehensive role-based access control (RBAC) system. The team agreed this could be a valuable enhancement, while being mindful of not conflicting with union.ai's commercial offering. Shared memory configuration for distributed training Niels proposed improving the configuration for shared memory used in distributed training setups, to provide a more user-friendly default experience. The team discussed technical details around mounting, timeouts, and integrating this with the flight-run fast registration process. Documenting multi-cluster deployments Fabio raised questions about the documentation for multi-cluster deployments, specifically around how the control plane components (like admin and data catalog) should be accessed by the data plane clusters. The team agreed to add a disclaimer in the docs clarifying that users need to handle the networking and authentication for these cross-cluster connections. Promoting community-maintained plugins David proposed introducing a "contrib" tier for community-maintained plugins, to help users better understand the support and maintenance levels for different plugins. The team agreed this was a good idea and will explore the details of how to implement this in a future RFC.