At the AI Summit in New York City today, Cloudera is announcing an initiative around operations of the machine learning (ML) workflow, also known as "MLOps." In a nutshell, the company is extending the Apache Atlas platform to accommodate the governance of machine learning assets, in addition to the more conventional data governance and data lineage capabilities it has had to date. Cloudera provides comprehensive details in a blog post that was actually published over two months ago.
Apache Atlas began life in 2015 as Hortonworks' Data Governance Initiative and was that company's data governance platform of choice, which it used to compete against Cloudera Navigator. Subsequent to Cloudera's acquisition of Hortonworks and the rationalization of the companies' various competing technologies, Atlas has became the go-forward governance standard for the company, and is the technology upon which the current Cloudera Data Catalog product is built.
But Cloudera's announcement is less about throwing down a competitive gauntlet and more a call to action for the industry to get serious about MLOps and land on an industry standard on which to implement it. MLflow -- an open source project (though not under the Apache Software Foundation) being pushed and led by Databricks, and built into the Databricks Unified Analytics Platform, available on Amazon Web Services and in the Azure Databricks service -- focus on the MLOps space too, addressing such capabilities as experiment management and a registry for machine learning models.
There are independent machine learning pure play vendors, like Dataiku, that address these requirements on their own platforms. And data preparation and management player Alteryx attacks the MLOps space head-on with its Promote product, derived from its 2017 acquisition of Brooklyn-based Yhat.
Also read: Alteryx expands product set, makes data science acquisition
The major cloud providers are no strangers to MLOps either. For example, Azure Machine Learning has experiment management built in (and announced in April that it was joining the MLflow project) and Amazon Web Services announced several ML capabilities last week at its re:Invent conference, including the SageMaker Debugger and Model Monitor components that are part of the new SageMaker Studio offering.
With so many competing MLOps initiatives, it's reasonable for Cloudera to be pushing the industry to coalesce around a standard, if for no other reason than to assist many of its customers who have really need the functionality.
Cloudera's Senior Product Marketing Manager for ML Ops and Data Science, Santiago Giraldo, told me that several customers have this need. This is especially the case with customers like Santander Bank, who use machine learning in their fraud prevention efforts. Fraud models are highly susceptible to data drift, since fraudsters are constantly changing their devices and techniques. As such, ML models in the fraud prevention domain are in constant need of monitoring and retraining. Managing that manually is unsustainable; tooling is essential.
The industry as a whole does need to focus on the MLOps problem, and integrate it into more conventional DevOps pipelines and frameworks. Whether a data governance platform is the best one on which the industry to standardize, is another question. Meanwhile, the call to action is overdue and hopefully a consensus can emerge.