Webinar Lessons: Scaling Machine Learning on AWS with Cloudability
Earlier this week, we sat down with DataVisor and AWS for a webinar to chat about scaling machine learning on the cloud. In the webinar, David Ting, VP of Engineering at DataVisor, and Gideon Wulfsohn of AWS gave us some great best practices for building machine learning infrastructure.
As an industry leader in fraud detection using machine learning, DataVisor knows just how important it is to correctly optimize your cloud infrastructure when scaling machine learning. Their system needs to be able to identify fraud as fast as possible — even when the attack is an entirely new approach. Pulling that off means optimizing their cloud as much as possible. It also means building a pipeline that can quickly process the massive amounts of data fueling their machine learning engine.
Let’s take a quick look at some of the key insights from the webinar.
#1: Using Unsupervised Machine Learning on AWS to Combat Fraud
Fraud techniques are constantly evolving, but so is DataVisor’s Unsupervised Machine Learning (UML) engine. The UML engine pulls from a Global Information Network with over 4 billion users and more than 800 billion events — which translates to a massive amount of data. That data goes through a three-step process that includes feature extraction, correlation analysis and result categorization. David gave us a rich view into their approach that included a look at their clustering analysis method.
A particularly interesting insight was how infrastructure was much more of a challenge than coding for machine learning implementation. Much of the machine learning code is well-documented and developed, with availability either on AWS or through open source. Thus the greater challenge was how to construct a pipeline and infrastructure to create the UML engine. David walked us through DataVisor’s approach, including how they built their pipeline to handle massive batch processing of data.
#2: Optimizing Cloud Cost & Usage Is Essential for Their Systems
As with many technology companies, DataVisor’s cloud represented one the largest cost centers for their company — on par with their labor costs. Beyond the cost factor, there was also a drive to optimize their cloud to reduce the amount of time it took to process data. The more their cloud was optimized, the faster their UML engine ran and the lower their costs were. In turn, this freed up funds, giving them more to invest in further innovation of their UML engine to make it even better. It’s a self-perpetuating cycle that makes the company and product continuously stronger.
David gave us a deep dive into their optimization process. He started by going over exactly how Cloudability gave them the visibility they needed to improve their process, then flipped over to show the changes they made as a result. Digging into their use of the RI Planner and Rightsizing tools, he showed how they lowered costs on key infrastructure while also showing how their homegrown SparkGen tool (utilizing Apache Spark) helped them use Spot Instances for their pipeline to further savings.
DataVisor is a great example of the kind of innovative approach to cloud cost optimization we love to see at Cloudability. Our tools helped them directly lower costs while also providing them the insights they needed to refine their individual process to increase efficiency, further lowering costs and driving innovation.
#3: Optimization Led to Incredible Results
DataVisor’s optimization was an aggressive strategy that paid off in big ways. To finish his presentation, David showed the results that they were able to achieve with their optimization, including:
- Drastically lowered pipeline latency
- Cloud cost lowered by millions of dollars
- Weekly ops dev-hours cut to a small fraction of what they were
- Massively increased peak scale capabilities
Overall, their optimization lowered their cloud costs by 50% while at the same time increasing the performance of their UML engine.
Want to get the whole story and all the details? Listen to the recorded webinar to get machine learning optimization best practices that can help you drastically lower costs and fuel innovation.