FuseML aims to provide an MLOps framework as the medium dynamically integrating together the AI/ML tools of your choice. It's an extensible tool built through collaboration, where Data Engineers and DevOps Engineers can come together and contribute with reusable integration code.
The 0.3.0 FuseML release brings significant improvements to installer and workflow extensions/components. The range of extensions has been extended to integrate more 3rd party AI/ML tools, such as Seldon Core, Intel OpenVino and NVidia Triton. The existing workflow extensions have also been enhanced with new options and parameters.
This is a summary of extension additions and improvements available in this release:
The documentation has been updated with new tutorials featuring newly added extensions and more complex MLOps workflows.
As usual, the built-in components, as well as the 3rd party tools that can be installed through the FuseML installer have been updated to their latest and greatest version:
Another important improvement that comes with v0.3.0 is ARM support for the installer, the FuseML CLI, the FuseML core container image and the container images built for all workflow components with the exception of the OpenVINO converter. This allows FuseML to be installed in Kubernetes clusters running on an ARM CPU architecture.
New sample ML projects that can be used with FuseML are also available in this release:
Implemented enhancements:
Fixed bugs:
upgrade: new config map does not have right values #242
kfserving extension fails to install with FuseML #218
fuseml-installer (re-run to incrementally move on) #206
Assigning the same codeset to different workflows results in app name conflict #183
fuseml upgrades fail with volume mount errors #232
can't list workflow runs in case of PVC failures #221
fuseml-installer lacking arm64 support #207
installer: rerunning the installer with extensions option fails #184
First mlflow-sklearn-e2e workflow run stuck for several minutes in the build step #134
When deleting a codeset, also removes its assignments to workflows #113
Closed issues:
mlflow-builder image not supported by glibc compiled libraries #262
mlflow-e2e job taking too long to complete #174
Overview
The FuseML 0.2 release brings a lot of improvements in the area of 3rd party AI/ML tools installation and integration. We've made it easier for FuseML users to extend the FuseML installer so that it is installing AI/ML tools running on kubernetes.
FuseML now also features an extension registry that stores information about the external AI/ML tools and services that are part of the tool stack that FuseML integrates with. This information is used to dynamically connect workflow steps to the AI/ML services that they need consume. Decoupling connectivity information from workflows allows FuseML users to configure workflows independently of how backing tools and services are deployed.
The 0.2 release also introduces data persistence, a much needed feature that brings FuseML closer to a production grade solution. The information that FuseML stores about projects, codesets, applications, workflows and so on is now stored in persistent storage and will survive accidental pod restarts.
With the new release, we've made some important changes to the architecture. We've updated all built-in components (Gitea, Tekton) as well as 3rd party tools (MLFlow, KNative, KFServing) to their latest version. The docker-registry service has been replaced by Trow, an OCI registry with a lower footprint. The Quarks component has also been removed.
On the CLI side of things, FuseML 0.2 now provides better support for projects and defaults for user, passwords and codesets.
We've also updated the MLFlow workflow extension to avoid rebuilding environment container images unnecessarily.
Implemented enhancements:
Fixed bugs:
Closed issues:
Implemented enhancements:
Closed issues:
Implemented enhancements:
Fixed bugs:
Closed issues: