By Animesh Singh (IBM), CD Foundation MLOps SIG Chair
Machine Learning (ML) is quickly becoming a key technology in a wide range of industries and organizations. One key thing happening in ML landscape is that more and more models are getting produced, but are they actually getting deployed?
Looking at the number of steps needed to be performed in Data and AI lifecycle, currently in a lot of cases the process remains bifurcated amongst various teams, and at every single step there are manual handoffs. Besides just slowing down the production and deployment of models, this also poses a strong challenge for traceability, governance and risk management.
This signifies utilizing Continuous Delivery (CI/CD) in ML may increase the speed of deployments, but does it increase the quality? How do you know if you’re deploying models which are giving ethical, fair and unbiased predictions? Are the actions performed by the ML code verifiable? Can the code be maintained and tested rigorously? Are there clear guidelines for lineage tracking, metadata collection, experiment tracking, data versioning, ETL operations, and more?
To address these challenges, the MLOps SIG managed under the CD Foundation has been formed with the following goals
- MLOps Definition and Roadmap: Create vision and roadmap for MLOps, what it means, and how do we envision its role within the CI/CD ecosystem. Terry Cox, a co-leader in SIG has done an excellent job of compiling a draft vision and roadmap for this
- Reference Architecture and Design Patterns: Create reference architecture, design patterns and implementations and processes for MLOps. A use case is integration between MLOps Pipelines and Tekton e.g. Kubeflow Pipelines working with Tekton as a backend. The SIG has already delivered significant results and has kicked off a project in to start enabling some of these capabilities on top of Tekton. Some of the artifacts which have been produced are listed here:
- Argo and Tekton Comparison
- Argo-Tekton Yaml for Flip Coin
- KFP Compiler for basic sequential pipeline in Tekton
- Kubeflow Pipelines-TFX Pipelines
3. AI Governance and Risk Management: Define architecture and guidelines around lineage tracking, metadata collection, experiment tracking, data versioning, ETL operations, etc. which a typical Data and ML Pipeline shall support to enable Ethical AI
MLOps SIG communication happens via a public mailing list: https://lists.cd.foundation/g/sig-mlops
You can join in the discussion on Slack with us to collaborate.
The MLOps SIG meets every other week at 9:30 AM Pacific, on Thursdays
Meeting agendas, minutes, and documentation are here: https://github.com/cdfoundation/sig-mlops
Download this invitation to add the meeting to your calendar: https://zoom.us/meeting/u5Iqduutpj8o7fVIT1pePLk5wv4H9XpojQ/ics
Full details on agenda, members, meeting times, mailing list sign up, and more are available here: https://lists.cd.foundation/g/sig-mlops
All are welcome to join the mailing list and attend meetings. Please add your voice to this important new area in CI/CD development.