Logistics ops · Production

DoorDash builds a centralized ML platform quadrupling model count and achieving 5x prediction throughput

The problem

DoorDash's rapid hypergrowth required a centralized ML platform to abstract infrastructure complexity for data scientists, and the existing prediction service could not keep pace with surging food order volumes during COVID-19.

First attempt

The manual model testing process—requiring data scientists to hand-write Python gRPC scripts per migration—was not scalable as the team grew, and early feature quality monitoring required an onboarding step that hindered adoption.

Workflow diagram · grounded in source
1
Data scientist initiates ML workflow
trigger
“That highly iterative process is a scientific endeavor that requires ongoing experimentation over the course of multiple steps”
2
Declarative feature engineering
integration
“the platform provides a declarative way of performing feature engineering logic, during which it figures out how to execute the logic, orchestrate the necessary computations, and secure the necessary compute resources”
3
Sibyl online prediction serving
ai_action
“Sibyl prediction service to perform online predictions at high throughput and low latency. Among its notable capabilities are batch predictions, model shadowing, and feature fetching”
4
Feature quality monitoring
validation
“Model predictive performance can decay with time or show unexpected results. We want our users to be able to know about decay, manage it, and take corrective actions quickly to resolve underlying issues”
5
Customer feedback drives iteration
feedback_loop
“our biannual customer survey revealed a need for a proper onboarding experience for new data scientists so they can be productive during their first three months at DoorDash”
Reported outcome

DoorDash's ML platform quadrupled the number of models and achieved 5x prediction throughput; a feature store optimization cut costs three-fold and reduced feature fetching latencies by 38%, while the platform now handles billions of predictions per day.

Reported metrics
Model count growthquadruple the number of models
Prediction throughput growth5x the number of predictions
Feature store cost reductionreduced costs three-fold
Feature fetching latency reduction38%
Show all 5 reported metrics
model count growthquadruple the number of models
prediction throughput growth5x the number of predictions
feature store cost reductionreduced costs three-fold
feature fetching latency reduction38%
daily predictions scalebillions of predictions per day
Reported stack
SibylRedisgRPC
Source
https://careersatdoordash.com/blog/3-principles-for-building-an-ml-platform/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

DoorDash's ML platform quadrupled the number of models and achieved 5x prediction throughput; a feature store optimization cut costs three-fold and reduced feature fetching latencies by 38%, while the platform now han…

What tools did this team use?

Sibyl, Redis, gRPC.

What results were reported?

Model count growth: quadruple the number of models; Prediction throughput growth: 5x the number of predictions; Feature store cost reduction: reduced costs three-fold; Feature fetching latency reduction: 38% (source-reported, not independently verified).

What failed first in this deployment?

The manual model testing process—requiring data scientists to hand-write Python gRPC scripts per migration—was not scalable as the team grew, and early feature quality monitoring required an onboarding step that hinde…

How is this logistics ops AI workflow structured?

Data scientist initiates ML workflow → Declarative feature engineering → Sibyl online prediction serving → Feature quality monitoring → Customer feedback drives iteration.

Related logistics ops cases