MLOps at DoorDash (Data+A.I. summit 2022)

This post is a summary of a insightful tech talk by Hien Luu, Head of ML platform in Doordash about how DoorDash’s MLOps infrastructure, motivation, strategy and learnings.

Link: session video.

At a glance

1. DoorDash’s formula of MLOps success: mlops_success = combine(user_case, culture, technology, people)
2. DoorDash’s journey to MLOps success was driven by building the blocks that matters the most to the “use case”.
3. Adopt MLOps as a team sport by paying attention to organizational alignment upfront.
4. Who owns the model, scientist, MLE or ML Infra Eng? Scientist owns the model E2E taking the model to the production; MLE helps integrate the model into the serving infra (say search/ranking query) and manage training lifecycle; ML Infra Eng owns the infra to enable the own experiences to MLE and scientists.

DoorDash’s MLOps Journey

Formula: mlops_success = combine(user_case, culture, technology, people)

• Use case: identify the game you are playing to define project priorities. (Say data governance for Banking/Health care use case vs. velocity for personalized marketing, voice assistance)
• Culture: risk tolerance, velocity (relative to customer’s), decision making process, collaboration
• Technology: the maturity of data infra, experiment infra, CI/CD, compute infra
• People: align requirements among science (MLE, scientist, analyst), business (PM) and engineering (MLOps) teams.

Way to MLOps success: MLOps Maturity Model

Most of the companies or organizations out there use the maturity model proposed by Microsoft and Google to position themselves when adopting ML in their business. Hien presented a practical tool (proposed by Algorithmia) to define your organization’s ML maturity-level and generates priorities.

DoorDash AI Infrastructure High-level Architecture

If you have spent enough time in the MLOps world, DoorDash’s ML system would not be anything surprisingly flashy but mostly very common architecture most companies follows (velocity-optimized I supposed?). Here are some great slides Hien shared.