Logistics ops · Production

DoorDash builds an ML Platform to standardize and scale machine learning across logistics and marketplace

The problem

As ML usage grew across DoorDash—spanning fraud prediction, search ranking, delivery time predictions, and recommendations—there was no holistic ML Platform to standardize frameworks, manage model lifecycles, or compute and serve features at scale, limiting engineering and data science productivity.

Workflow diagram · grounded in source

Real-time event ingestion

trigger

“Realtime Feature Aggregator - Listens to a stream of events and aggregates them into features in realtime and stores them in the Feature Store. These are for features such as historic store wait time in the past 30 mins, recent driving s…”

Historical feature computation

integration

“Historical Aggregator - This runs offline to compute features which are longer-term aggregations like 1W, 3M, etc. These calculations run offline. Results are stored in the Feature Warehouse and also uploaded to the Feature Store.”

Model training and storage

ai_action

“Model Training Pipeline - A build pipeline where models will be trained for production usage. Once a model training script is submitted into git repo, this pipeline takes care of training the model and uploading the artifacts to the Mode…”

Online prediction serving

ai_action

“Prediction Service - Serves predictions in production for various use cases. Given a request with request features, context (store id, consumer id, etc) and prediction name (optionally including override model id to support A/B testing),…”

Prediction log capture

feedback_loop

“Prediction Logs - This stores the predictions made from the prediction service including the features used when the prediction was made and the id of the model used to make the prediction. This is useful for debugging as well as for trai…”

Reported outcome

DoorDash standardized on LightGBM for tree-based models and PyTorch for neural network models, and designed a four-pillar ML Platform architecture—Modeling Library, Model Training Pipeline, Features Service, and Prediction Service—to support end-to-end model lifecycle management at scale.

Reported metrics

ML shipping productivityincrease the productivity of shipping ML-based solutions

Model accuracy across evaluated frameworksaccuracy of models were almost the same

Reported stack

LightGBMPyTorchTorchScript

Source

https://careersatdoordash.com/blog/doordash-ml-platform-the-beginning/

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

DoorDash standardized on LightGBM for tree-based models and PyTorch for neural network models, and designed a four-pillar ML Platform architecture—Modeling Library, Model Training Pipeline, Features Service, and Predi…

What tools did this team use?

LightGBM, PyTorch, TorchScript.

What results were reported?

ML shipping productivity: increase the productivity of shipping ML-based solutions; Model accuracy across evaluated frameworks: accuracy of models were almost the same (source-reported, not independently verified).

How is this logistics ops AI workflow structured?

Real-time event ingestion → Historical feature computation → Model training and storage → Online prediction serving → Prediction log capture.