How MLOPs implements AI-enabled products differently than a DevOps cycle
With the pandemic increasing the pace of digital transformation, but how do we move from proof of concept and MVP to AI in production, at scale. To achieve this we are leveraging MLOps, DataOps, and DevOps.
Some might wonder how these are different. This talk aims to examine the various ways in which these two practices differ.
Understanding MLOps and DevOps
MLOps developed out of DevOps, thus it uses its principles and workflows by applying them to machine-learning operations. However, they are understood as a different set of practices:
- MLOps — a practice for collaboration and communication. It allows data scientists, ML engineers, and operations professionals to manage everything from model training to deployment.
- DevOps — a set of practices that pools software development and IT operations. It helps reduce life cycle development and provides high-quality, uninterrupted delivery.
Since machine learning systems are software systems, DevOps principles also apply to MLOps. And while there is overlap it is not that simple. MLOps stands for “machine learning” and “operations” but it goes beyond these confines. MLOps is not just DevOps applied to machine learning because we are not developing software but rather managing software and implementing AI into a product.
Differences in developing AI-enabled applications
DevOps applies only to code whereas we need to consider code, data, and models. Therefore, the development lifecycle needs to be described as think, design, build, train, test, deploy, and run. There are many aspects of MLOps and DevOps that require different focus or implementation whereas there is one pipeline for code integration and possibly a second one for deployment. MLOps has many pipelines integrated into an automated feedback loop for machine learning workloads.
- Exploration needs
- Industrialization
- Production
Regarding competencies, MLOps includes ML engineers, Data Scientist, and Operations. Their priorities are to evolve models, achieve more stability, and most importantly have a best-in-class product enabled by AI.
The feedback loop for MLOps looks for quality data by automating as much as possible from unit testing to performing functional tests. By going step by step the data citizen can ensure that all issues are addressed in a timely fashion. So there is no need to rush to cram all the tests at the end. The deployment enables multiple strategies such as A/B testing and the monitoring focus on detecting drift and bias in data and models.
Depending on the results of the continuous monitoring the cycle will be be done again as many times as necessary. An automated feedback loop aims to confirm all the steps from start to finish: from the choice of data set and automated ingest subject to data governance, as well as the design and industrialization of use cases and models. A typical feedback loop includes multiple steps as described below.
Data trust — To build anything you must have trust in data and know where it comes from, its quality, and the regulations applied to it.
Train — Training is required for validation, debugging, strategy, and visualization. Once you train it you need to act on it.
QA — Once you have trained you need to evaluate the training to find the strategic plan it needs to visualize what’s going on. This means running acceptance tests to see if it complies with business requirements and regulations.
Deploy — This phase does not exist in the DevOps model. Deployment means getting your model launch signed-off and some kind of packaging of the model. We will put it in a repository where it can be accessed with a data set and then release management can be done of the model.
A/B test — Just as in DevOps you would do some beta testing to see how your model compares to others. We then analyze which people this model can best be applied to.
Serving — In the MLOPs cycle you need to offer an explainable decision and API support inference
Monitor- Like in DevOps you have monitoring of the application, its performance, business, fairness but a crucial difference your model must be based on what data comes in. If your data is not expected, we will not be able to make predictions so we must be very aware of data & model drifting, and constantly observe any shifts in data to maintain accuracy.
Trigger — Enables trigger redeployment and retraining using mostly root cause analysis
Nowadays more and more AI-enabled products are being built. MLOps takes the idea best practices of software engineering and DevOps to code, but industrialization needs to go further including data, and training models to adjust to the ever changing environment and rules adaptation. Therefore, there is a need for an automated feedback loop that can differ from the DevOps lifecycle.
