March 2021 Summaries

The ML Data Engineering conference `apply()` will be held on April 21 and 22, bringing together industry thought leaders and practitioners from over 30 organizations to share and discuss the current and future state of ML data engineering. The complexity of ML data engineering is a significant barrier between most data teams and transforming their applications with operational ML. The conference aims to address this by sharing findings and best practices on topics such as online serving, data observability, feature stores, model performance management, and MLOps best-practices. Speakers from various industries and backgrounds will share their experiences and perspectives on ML data engineering, including innovators like Wes McKinney and Chip Huyen from companies like Confluent and Stanford. The event is entirely virtual, making it easier to bring together participants from across the country and the globe.

The text discusses the challenges and strategies of sharing and reusing features across different machine learning (ML) models in an organization. It highlights that companies with tens of thousands of ML models in production, like Uber, Twitter, or Google, have successfully shared and reused features to scale their operational ML capabilities. The key strategy for successful feature reuse is to share and reuse features across models and use cases, which allows teams to leapfrog over the most difficult parts of putting new models into production. However, this requires addressing common challenges such as lack of knowledge about existing features, unclear ownership, and poor pipeline visibility. The text proposes three approaches for sharing features: publishing feature data to a shared location, sharing feature pipelines, and implementing a feature store. Each approach has its pros and cons, and the choice of approach depends on the organization's specific needs and requirements.

The authors of the article, Lior Gavish and Kevin Stumpf, argue that traditional on-prem deployment models offer security and compliance benefits but require significant operational overhead from customers, while SaaS solutions provide ease of deployment and convenience but introduce anxiety for security and compliance teams. They propose a hybrid deployment architecture as a better solution, which marries the best of both worlds by providing a control plane managed by the vendor in their environment, and a data plane in the customer's environment, allowing customers to maintain full control over their sensitive data while still benefiting from the ease of deployment and convenience of SaaS. This approach enables faster onboarding, quicker time-to-value, and greater flexibility for customers, while also providing vendors with a secure and compliant solution that can be easily managed by the vendor.

Tide has successfully deployed machine learning (ML) models 2x faster using Tecton's feature platform, reducing its time to deploy new models from 2-4 months to just 1 month. The company achieved this significant improvement by repurposing resources that were previously dedicated to maintaining an internal feature store. With the help of Tecton, Tide was able to accelerate the development of ML features and triple the number of features used per model, while also saving a substantial amount of engineering time and headcount. This deployment has enabled Tide's data scientists and engineers to focus on making data-driven decisions for their customers, ultimately helping them save time and money.