Back office ops · Production

Griffin: How Instacart's ML platform tripled ML applications in a year

The problem

Instacart's previous ML framework Lore became a bottleneck as the number, diversity, and complexity of machine learning applications grew — its monolithic architecture required refactoring the core design for every new feature.

First attempt

Lore's monolithic architecture became untenable at scale — when the company only had a few ML applications it fulfilled requirements, but as applications multiplied the framework required core refactoring for every new capability.

Workflow diagram · grounded in source
1
MLE generates workflow code
trigger
“Generate ML workflow code from base templates and update the code of their application”
2
Feature engineering via Feature Marketplace
integration
“Backend service schedules pipelines to compute features on a regular cadence specified in the FD and index the computed features in Feature Store, a storage layer to provide consistent access to features”
3
Training pipeline orchestration
integration
“Workflow Manager schedules and manages machine learning pipelines. It leverages Airflow to schedule containers and utilizes an in-house abstraction - ML Launcher - to containerize task execution”
4
Hyperparameter tuning
ai_action
“tune the hyperparameter by scheduling multiple runs and tracking the metrics/metadata in the MLFlow”
5
Model deployment to inference
output
“deploy the model version with the best performance based on offline metrics to the inference service using Twirp(an RPC framework) and AWS ECS (a managed container orchestration service)”
6
Continuous retraining cycle
feedback_loop
“All-new workflows are automatically synchronized, scheduled, and executed for continuous training”
Reported outcome

Griffin enabled Instacart to triple the number of ML applications in one year by providing an extensible, self-serving platform that lets MLEs quickly iterate on models, effortlessly manage product releases, and closely track production applications.

Reported metrics
ML applications growthtriple the number of ML applications in one year
Reported stack
GriffinLoreMLCLIMLFlowRayKubeflowFeastSnowflakeAWSDatabricksSagemakerAirflowDockerRedisScyllaS3SparkFlinkTensorflowPytorchSklearnXGBoostFastTextFaissTwirpAWS ECSKerasScikit-learnPostgres
Source
https://tech.instacart.com/griffin-how-instacarts-ml-platform-tripled-ml-applications-in-a-year-d3d4dcae3690
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Griffin enabled Instacart to triple the number of ML applications in one year by providing an extensible, self-serving platform that lets MLEs quickly iterate on models, effortlessly manage product releases, and close…

What tools did this team use?

Griffin, Lore, MLCLI, MLFlow, Ray, Kubeflow, Feast, Snowflake, AWS, Databricks.

What results were reported?

ML applications growth: triple the number of ML applications in one year (source-reported, not independently verified).

What failed first in this deployment?

Lore's monolithic architecture became untenable at scale — when the company only had a few ML applications it fulfilled requirements, but as applications multiplied the framework required core refactoring for every ne…

How is this back office ops AI workflow structured?

MLE generates workflow code → Feature engineering via Feature Marketplace → Training pipeline orchestration → Hyperparameter tuning → Model deployment to inference → Continuous retraining cycle.

Related back office ops cases