How Wix Uses Amazon SageMaker to Solve Multitenant Machine Learning at Scale

Headquartered in Tel Aviv, Israel, Wix helps users create, manage, and grow a digital presence. What began as a website builder in 2006 is now a complete platform providing users with enterprise-grade performance, security, and reliable infrastructure. With access to a wide range of commerce and business solutions and advanced SEO and marketing tools, its users have full ownership of their brand, data, and customer relationships so that they can build a powerful digital presence. Wix places a high priority on continuously innovating and delivering new features and products, and the company wanted to take advantage of its user and product data to develop more sophisticated solutions. However, it needed a robust continuous integration and continuous deployment framework to support the ML life cycle.

In 2018, Wix turned to AWS and used several AWS services, including Amazon SageMaker, a service that makes it simple to build, train, and deploy ML models for nearly any use case, to build the Wix ML Platform. On AWS, Wix could access advanced capabilities to accelerate the ML life cycle, such as one-click deployment for Amazon SageMaker and out-of-the-box compatibility with MLflow, an open-source solution used for ML development.

“Most of Wix’s infrastructure is built on AWS, making the integration of the Wix ML Platform with internal services and systems much smoother,” says Itamar Keller, research and development team leader at Wix.

The Wix ML Platform handles critical steps in Wix’s ML life cycle, from data management to model building, training, deployment, and monitoring. Using this solution, Wix data scientists can quickly spin up research environments using Amazon Elastic Compute Cloud (Amazon EC2), a service that provides secure and resizable compute capacity for virtually any workload, without requesting any development resources. They can also seamlessly deploy ML models to make predictions using Amazon SageMaker inference capabilities, which provide a broad selection of ML infrastructure and model deployment options to help meet all of Wix’s ML inference needs. “Amazon SageMaker inference helps us deploy models across multiple Availability Zones and runs predictions at scale, either online or in batch mode,” says Keller.

On AWS, Wix has automated critical tasks in the ML pipeline, making it simpler for data science teams to deploy advanced solutions at scale. For example, the team built a language model for its customer support chatbot. Now, the bot can provide faster and more accurate support to thousands of users per day, powered by Amazon SageMaker training.

“The language model was trained to understand the language specifics of both the domains of Wix in general and customer care specifically,” says Keller. “For example, it understands that Velo is a Wix product. It knows that ‘booking’ in the general language is either the well-known company Booking.com or the verb booking, but in the Wix language, it is a major product that, if you’re a fitness trainer, lets you manage your appointments.”

In order to adapt to Wix’s dynamic business environment, the chatbot content teams have the freedom to retrain the language model as often as needed, without relying on support from data science or engineering teams. They can effortlessly incorporate new data, evaluate its effectiveness, and train the ML model within a user-friendly environment. Once the new ML model is prepared for deployment, chatbot content team members can swiftly replace the old model and redeploy the new model in production, facilitating uninterrupted availability through blue/green deployment. This approach makes sure that the new (green) fleet is fully operational before transitioning traffic from the old (blue) fleet.

The Wix team also used the platform to optimize its website categorization service, an artificial intelligence–driven solution that automatically determines the business-oriented categories of users’ sites based on their content. Wix has an extensive, highly curated taxonomy of websites; a single website can belong to multiple categories, which is also known as the multilabel classification problem. Now, the Wix team can automatically collect websites that were updated by users within the last 24 hours, run batch predictions using Amazon SageMaker batch inference, and assign the websites to the relevant categories.

“The Wix ML Platform can support hundreds of independent ML models running simultaneously in production,” says Keller. “On AWS, we can modify, train, build, evaluate, and redeploy the full set of models without any involvement from the engineering team and with very minor operations made by data scientists.”