Watch: Introducing MLflow & Wix’s Internal Machine Learning Platform

Wix Engineering’s meetups focus on deep technical insights, whether it’s front-end, back-end, mobile, testing, machine learning or anything in between. You can join us on Tel-Aviv, Kyiv, Dnipro and Vilnius.

Below you’ll find videos from one of our recent meetups, where Databricks’ Daniel Haviv and our very own Ran Romano, shared their insights on machine learning.

Simplifying Production Machine Learning with MLflow /Daniel Haviv

Bringing a machine learning project to a successful conclusion is more difficult than it may seem initially. We have to take into consideration the full machine learning lifecycle, which allows us to focus not only on the development of the model, but also on production and monitoring.

Thankfully, there are mature tools available to manage the machine learning lifecycle. One of the tools for end-to-end machine learning lifecycle management is the open-source platform MLflow.

During this session we will deep dive into main MLFlow components:

• MLflow tracking: using an API and UI to track/log/visualize machine learning experiments.
• Mlflow projects: using standardized format to package reusable data science code.
• Mlflow models: using provided tools to deploy common model types to diverse platforms.

Overview of Wix’s Machine Learning Platform /Ran Romano

Machine learning powers a variety of products at Wix. With many ML models — from basic regression and classification methods, to sophisticated recommendation and deep learning based models, the ML engineering team faces a significant challenge of supporting the plethora of models in production.

In this talk Ran presented the internal ML platform designed to address the end-to-end ML workflow:

1. Data management
2. Model training, experimentation and evaluation
3. Model deployment
4. Serving and monitoring predictions

We’ll cover the architecture and main flows of the system, built on top of a mixture of managed services like AWS Sagemaker and open source tools such as Apache Spark, and MLflow.

In addition, we will dive deeply into two of the main components of the system:

1. Machine Learning CI — an MLflow-based CI system, designed for creating reusable and reproducible experiments.
2. Feature store — A single, curated, discoverable, source of truth for features. Features are generated declaratively, in a fashion which facilitates feature reuse. And most prominently, solves one of the hardest problems of using ML in production — training / serving skew.

For more engineering updates and insights:

• Follow us on: Twitter | Facebook | LinkedIn
• Read our blog
• Visit us on GitHub
• Subscribe to our monthly newsletter
• Subscribe to our YouTube channel
• Follow our Medium publication

Photo by Tanner Boriack on Unsplash