A look at MLOps with Best Practices for Operational Excellence
MLOPs is a key component in the transformative tech landscape. By operationalizing ML, business can look to reduce costs, increase revenue and improve data science efficiencies through improved workflow automations and scalability. Kamal Jain, Principal Data Engineer Manager, BT discusses the best practices within MLOPs to drive operational excellence in this interview with Brad Shimmin, Chief Analyst, AI & Data Analytics, Omdia, and hosted by Victoria Wilcox.
What breakthroughs from the past year are you most excited about (either in your organization or in the wider field of data engineering and AI)?
One of the most recent breakthroughs that everyone in the industry is talking about is ChatGPT, launched by OpenAI. It's having a tremendous effect on the technology industry in terms of the way we can ask the questions to AI systems, and have that system provides complete answers to users, all within the relevant application they are using.
What are the key considerations around achieving success with MLOps when implementing this in a business or organisations? Challenges/barriers to success etc?
One of the key challenges I tend to see when implementing MLOPs is getting the support and buy-in from the right stakeholders by ensuring we showcase the value that MLOPs is going to provide to the users. This comes as part of the complete change management process. It’s important that when we are building an MLOPs ecosystem we make sure the correct governance and security is integrated into the whole life cycle.
What kind of use cases appear best aligned with MLOps practices; what aspects of these use cases supported successful outcomes? Can you also speak to those use cases that did not fare as well under MLOps principles?
MLOPs is very pertinent across all industries, however there is a huge issue with data drift happening. With trends and behaviours of users changing over time, we need to ensure that our models are giving accurate results, so it is very important that we have MLOps in place to support this. If we can automate the complete pipeline of MLOps, it will ensure that results coming from the machine learning models are more accurate.
On both a technological and organizational level what is your advice to for achieving operational excellence through MLOps?
My advice for achieving operational excellence through MLOps is to make sure we take care of data drift, so that when this inevitably occurs, we can ensure that teams are being alerted well in advance and can manage this. We can harness the latest techniques including cloud to automate CI/CD pipelines. We need to ensure that we are not breaching any security requirements and have complete governance in place. It’s also about ensuring we make the complete data versioning and the model versioning more predictable.
What is new in this area and what does 2023 look like in terms of how these tools might influence MLOps practitioners such as yourself?
There are two things which are still going to be very important to watch out for in 2023 and beyond. One is the issue of data drift as I have already mentioned. For example, how can we predict the data drift which is being fed into machine learning models? That's going to determine how successfully we can proactively train our machine learning models so that they keep giving accurate results.
We also need to watch out for the automation of MLOPs. We currently have something like AutoML, which is being used to automate the machine learning tasks. On the similar lines, how can we put in place something like Auto-MLOPs so that the processes we are building into the MLOPs stream can be automated and made much more predictable for organizations to use.
You can join Kamal Jain at The AI Summit London this June 14-15, at the Tobacco Dock, where he will be talking in-depth as a panellist on this as part of a panel discussion on the Practitioners’ Stage, day two at 11:55am-12:35pm. Find out more.