Ecommerce ops · Production

Instacart supercharges ML/AI foundations with Griffin, Feature Store, and Axon in H1 2023

The problem

ML models at Instacart were trained on laptops or siloed team infrastructure with no common patterns, and promoting a model to production often took more than a month, limiting ML velocity and sophistication across the company.

First attempt

The only GPU access path for MLEs was AWS SageMaker, which required multiple rounds of terraform configuration that could take days to weeks for first-time provisioning and was difficult to maintain for dependency management.

Workflow diagram · grounded in source
1
User Modeling signal generation
ai_action
“User modeling signals enable data scientists and MLEs across the company to characterize our customers along a variety of dimensions, including how sensitive they are to fees or category-specific prices, how open they are to product disc…”
2
Feature Store management
integration
“Through a Feature Store, we manage and share features for training ML models. In addition, real-time features from an online feature store support models deployed in production.”
3
Model training and deployment via Griffin
ai_action
“Using the ML productivity tool (Griffin), we train, tune, store, and deploy ML models”
4
Adaptive experimentation via Axon
validation
“Using the adaptive experimental platform (Axon), we increase experimental velocity and facilitate deep personalization via mechanisms such as multi-arm and contextual bandits”
5
Production data feedback loop
feedback_loop
“ML models in production engage with customers to generate new data, which is used as signals to update or train new models, completing the loop”
Reported outcome

Instacart reduced feature onboarding from days to under an hour, achieved up to 85% faster queries, cut TensorFlow P99 latency by 85%, enabled training with 10x or even 100x more data, and halved Feature Store ingestion costs — substantially accelerating ML productivity across the company.

Reported metrics
Feature onboarding timeless than an hour (vs. a few days previously)
Feature query speed85%
Feature Store ingestion costcut its ingestion costs in half
Feature Store database footprint50%
Show all 7 reported metrics
Feature onboarding timeless than an hour (vs. a few days previously)
Feature query speed85%
Feature Store ingestion costcut its ingestion costs in half
Feature Store database footprint50%
TensorFlow P99 inference latency85%
ML model training data scale10x or even 100x more data
Bandit experiment setup timeas little as a few hours, compared to ad hoc processes that can take weeks or months
Reported stack
GriffinFeature StoreAxonFeature MarketplaceTensorFlowTensorFlow ServingRayLightGBMSageMaker
Source
https://tech.instacart.com/supercharging-ml-ai-foundations-at-instacart-d48214a2b511
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Instacart reduced feature onboarding from days to under an hour, achieved up to 85% faster queries, cut TensorFlow P99 latency by 85%, enabled training with 10x or even 100x more data, and halved Feature Store ingesti…

What tools did this team use?

Griffin, Feature Store, Axon, Feature Marketplace, TensorFlow, TensorFlow Serving, Ray, LightGBM, SageMaker.

What results were reported?

Feature onboarding time: less than an hour (vs. a few days previously); Feature query speed: 85%; Feature Store ingestion cost: cut its ingestion costs in half; Feature Store database footprint: 50% (source-reported, not independently verified).

What failed first in this deployment?

The only GPU access path for MLEs was AWS SageMaker, which required multiple rounds of terraform configuration that could take days to weeks for first-time provisioning and was difficult to maintain for dependency man…

How is this ecommerce ops AI workflow structured?

User Modeling signal generation → Feature Store management → Model training and deployment via Griffin → Adaptive experimentation via Axon → Production data feedback loop.

Related ecommerce ops cases