Recently Cevo were engaged to help build an automated machine learning pipeline that created the ability to rapidly and accurately generate predictive models. All aspects of training a model were automated. We refer to this as MLOps.
Wait, What’s MLOps?
MLOps is DevOps applied to machine learning.
DevOps is a paradigm that attempts to remove silos and build bridges between all teams involved in delivering software. This is achieved by having cross-functional teams and automating as many tasks as possible that traditionally were manual such as building, packaging, testing and deploying a solution. When fully realised, DevOps can create rapid feedback loops for developers allowing them to deploy changes that are automatically verified to meet the highest standards in a matter of minutes instead of weeks or months.
When applying this mindset to machine learning, you have a cross-functional team where every member strives to automate their speciality. Through automation, Data Scientists have the ability to create and deploy high-quality models rapidly.
The benefits of MLOps are similar to that of DevOps:
- Rapidly adapt to the market and deliver value quickly.
- Fully automated model testing and verification ensuring that models are of the highest quality
- Automatic rollback of deployed models if things go wrong potentially saving hundreds of thousands of dollars
Through MLOps, teams can have greater confidence that the changes they are making have increased quality – and due to the reduced cycle times greater experimentation and market responsiveness can be achieved.
MLOps Team structure
In a typical machine learning project, you will find three groups: data scientists, data analysts and data engineers. These three groups perform their own dedicated role and may be part of separate organisational branches. Each group is in a sense a silo. The data scientist will request something from each member and wait for that task to be fulfilled in order to continue their work.
For example, they may ask a data engineer to create features or cleanse data.
MLOps attempts to remove silos and empower the data scientists. Data engineers aim to develop simplified self-service tooling that data scientists can use themselves to develop their required features. Machine learning engineers try to simplify the machine learning process and apply continuous delivery practices to deploy trained models to endpoints. All of these elements are wrapped together into an end to end automated pipeline that data scientists can kick off at will to quickly train high-quality models which are immediately available.
Traditionally, a data engineer would be required to generate features. By instead applying MLOps approaches we empower data scientists and work hard to remove silos. The data engineer creates tools for the data scientists to generate the features themselves.
Cevo recently helped one of our customers build an automated machine learning pipeline that created the ability to rapidly and accurately generate predictive models. We leveraged our experience in developing automated delivery pipelines for infrastructure and applications and applied these same techniques to the Machine Learning domain.
Using an MLOps approach, we were able to rapidly create models and deploy them to production within days, as opposed to months. This gives the organisation the ability to experiment, assess and gain value much quicker than the traditional approach. In fact, during the engagement, we were able to train a model that gave significant business value in less than a week from inception.
This is an impressive capability proving that applying MLOps principles enables rapid agile delivery of machine learning-based outcomes for organisations.
The future of machine learning is MLOps: automating as much as possible for data scientists so that they can iterate and get rapid feedback. The project we undertook proved the immense value that machine learning can bring to an organisation in such a short time.