Machine Learning Operations (MLOps) — Part1
Machine Learning Operations often referred to as MLOps, sounds like DevOps but trust me — it’s beyond the traditional Ops framework. MLOps builds on the concepts of DevOps and adds missing pieces like Data Quality, Model Output, and managing model hyperparameters.
It’s a unified set of practices for collaborating and communicating between Data Engineers, Data Scientists, and Operations teams to manage the production ML lifecycle. This collaboration leads to tremendous benefits in terms of scalability, centralization, and governed means to automate and scale the development of the right ML applications deployment and management in production systems.
Sharing one of the best illustrations i have come across to help understand MLOps
MLOps combines Machine Learning, DevOps, and Data Engineering, with an objective of reliable and efficient — developing, deploying, and maintaining ML systems in production catering to the business needs.
Need for MLOps
Until recently, we were dealing with manageable amounts of data and a very small number of models at a small scale. Additionally, the adoption of these models and their dependencies was less in the organizations. With the rapid exponential growth in the data every second and increased adoption of ML solutions leading to decision automation (increasing prevalence of decision making that happens without human intervention), machine learning models are becoming more critical but come with a cost as managing model risks becomes more important.
Some of the major challenges I have come across in my overall experience from developing to deploying machine learning models and monitoring them in production.
1. Data Dependency: One major challenge is changing data. Data reflect real-world behavior and changes over time. Due to continuous changes in data, models in production must adapt to these changes to continue to help organizations meet their goals.
2. Missing Common Language: ML projects involve people from various teams with varying backgrounds. This consists of Data Scientists, Data Engineers, Business, and other IT teams. These groups do not share common skills to communicate about data and their alignment is crucial to deploying models in production and evaluating their success.
3. Data Scientists as Data Engineers: As organizations are adopting more and more models, some of them could be complex as well, data scientists are facing challenges in re-developing & iterating models and monitoring them in production. Data scientists are experienced in developing and evaluating ML models but not in writing applications to be deployed.
Traditional approaches to managing these challenges add a lot of collaboration and communication efforts, additional resources, and time. Eventually, impacting the timelines and business goals.
There could be more challenges present with the increased adoption of Machine Learning in every organization. But the key observation I am trying to highlight is: that various challenges in managing ML models make MLOps more important going forward in every organization.
The above figure shows a realistic picture of a machine learning model life cycle inside an organization today (without MLOps practice), which involves many different people with completely different skill sets and who are often using entirely different tools.
Benefits of MLOps
· Automated Retraining of ML Models –Data reflect real-world behavior and changes over time. Due to continuous changes in data, models in production must be up to date. Frequent retraining is critical. The manual checks process adds more time. Through automation, it's easy to discover drifts in input data and model results in reduced associated costs.
· Shorter development cycles, and as a result, shorter release cycle.
· Improved collaboration between teams across all levels of technical expertise.
· Increasing reliability, performance, scalability, and security of ML systems.
· Streamlining overall operational and governance processes.
· Increased ROI of ML products.
What does it take to Master MLOps
Future of MLOps
The past couple of years has seen tremendous growth in MLOps popularity. This signals the importance of MLOps. With the persistent growth in data and technologies continuing and reaching new heights, Organization’s ability to develop strong ML strategies will be a key factor in the future.
In my next post, I will talk about MLOps Components and Drifts (as this is critical for ML models) and some open source tools for MLOps.