MLOps for Dummies
“87% of data science projects never make it into production.”
The machine learning community is now preparing for a new challenge: deployment. But why so many stories about something so obvious? You’re building something to deploy, right? Well not really. Many models of machine learning never see the light of day. And those that go into production make little noise. That’s why production issues have been pushed under the rug for much of the AI hype cycle of the past decade. We read about cutting edge algorithms and Unicorn AI startups, but how produced is ML? Organizations are finally facing these challenges. And they found a hero in MLOps – a meticulous marriage between machine learning and software engineering. At the third edition of Rising hosted by Analytics India Magazine, Hamsa Buvaraghan of Google Cloud gave the audience a glimpse into how MLOps are powering the machine learning pipelines of the future. Hamsa leads the Data Science and MLOps Solution team at Google Cloud to create breakthrough software solutions for business problems using Google Data Analytics and AI / ML products.
In her talk, she showed how MLOps solutions fit perfectly into the ML pipeline automation aspirations.
- Organizations barely succeed beyond pilots and proofs of concept.
- 72% of organizations that started AI pilots couldn’t even deploy a single application to production.
- According to a recent survey, 55% of companies have not deployed an ML model.
- Models don’t make it into production, and if they do, they break.
- Teams do not have reusable or reproducible components and their processes involve transfer difficulties between data scientists and IT.
- Deployment, scale-up and release management efforts always create headaches.
Who needs MLOps
MLOps bridges the glaring gap between the development and deployment of machine learning, in the same way that DevOps and DataOps support application engineering and data engineering. According to Google Cloud, successful deployments and efficient operations are a bottleneck in leveraging AI. MLOps is an engineering culture and practice that aims to unify ML systems development (Dev) and ML system operations (Ops). Hamsa pointed out that the ML code is a small part. From configuration to monitoring, service infrastructure to resource management, building production-grade machine learning systems requires more than just code.
Creating an ML-compatible system is a multi-faceted endeavor that combines data engineering, ML engineering, and application engineering tasks. “It takes a village to build an MLOps pipeline,” Hamsa said. For example, some basic functionality is required to support any IT workload, such as a reliable, scalable, and secure compute infrastructure. MLOps capabilities include experimentation, data processing, model training, model evaluation, model delivery, online experimentation, model monitoring, ML pipeline, and model registry. An ideal MLOps pipeline can support machine learning development, training operationalization issues, model deployment issues, data and model management, and more. Hamsa believes that organizations are now moving towards automated end-to-end pipelines and that MLOps will have applications in many industries. One of the most important features of MLOps is its ability to leverage ML metadata and artifact repository as well as data and entity repository. Artifacts can be anything: splits of processed data, patterns, statistics, hyperparameters, models, or model evaluation metrics, to name a few.
“These ML pipelines are the result of a combined effort of data scientists, data engineers, SRE and ML engineers.”
The principle Change Everything Changes Everything, or CACE, refers to the reliance on minor changes in a software engineering pipeline. In the context of machine learning, this principle extends to hyper-parameters, learning parameters, sampling methods, convergence thresholds, data selection and essentially all other possible adjustments. Various ML artifacts are produced in different MLOps lifecycle processes, including descriptive statistics and data schemas, trained models, and evaluation results. There is also metadata, which is the information about these artifacts.
Benefits of MLOps
- MLOps allows shortened development cycles.
- MlOps leads to increased reliability, performance, scalability and security of ML systems produced.
- MLOps enables the creation of the entire ML lifecycle.
- It also helps manage risk as organizations adapt their ML applications to more use cases in changing environments.
Model governance, version control, explainability, and other details of deploying a machine learning model present a nightmarish scenario for an ML practitioner unaware of the labels of software engineering. MLOps, with its myriad of options and a growing developer community, is today the best possible solution to deal with the realities of model production.
Join our Telegram group. Be part of an engaging online community. Join us here.
Subscribe to our newsletter
Receive the latest updates and relevant offers by sharing your email.