- Support temporality through Kueue
- Once data migration is completed, run data preheating, triggering the running of machine learning tasks (such as tfjob, mpiJob, pytorchJob, sparkJob)
- After computation is completed, data migration and cache cleaning can be carried out
- Choose data access methods based on the scheduling results of the Kubernetes scheduler (default scheduler, Volcano, YuniKorn)
- If scheduled to ordinary nodes with shared operating system kernels, adaptively use csi plugin mode
- If scheduled to Kata container nodes with independent operating system kernels, you can use the sidecar mode adaptively and support scalable modifications by cloud vendors
- Support basic data operation
- Combine with Hugging face and Pytorch to support transparent data acceleration through pre-reading and multi-stream reading
- Support defining automated data flow operations
- Integrate with Kubeflow Pipelines to accelerate datasets in the pipeline
- Integrate with Fairing for model development and deployment in the notebook environment
- Integrate with KServe to facilitate model deployment
- Minimum container permission (remove the privileged permission of FUSE Pod)
- Minimum rbac permission
- Minimal container image installation
- Continuously provide best practice documentation
- Simplify deployment
- Merge Dataset/Runtime controllers into one binary package
- Simplify usage
- Support Runtimeless, Dataset as the single API entry for users to use Fluid
- Improve code quality
- Reduce repetitive code
- Improve test coverage
- Security hardening
- Minimize the permissions of controller's RBAC
- Regularly review and update the permissions when new runtime is introduced
- Enhance observability
- Provide monitoring and alerts for Datasets
- Enhance the quality of documentation
- Organize the documentation so users can navigate it easily and find the information
- Provide more practical examples and tutorials can significantly improve the user's comprehension and learning process.
- Maintain consistency in language, style, and formatting throughout the documentation