ralf: Real-time, Accuracy Aware Feature Store Maintenance
summary
Feature-store maintenance splits into two concrete systems: ralf uses a scheduling policy that tracks per-key cumulative regret (error accumulated from stale features) and prioritizes updates so the same resource budget yields higher downstream accuracy; experiments on real time-series show higher accuracy for equal cost. Sky Plane is a CLI data-replication tool (works across clouds, marketed as pip install skyplane) that uses overlay routing, striping, and on-the-fly compression to cut transfer time up to 113x faster and costs from about $7.27 to $1.17 for example transfers.
AI-written summary · about
related talks
30mHow Feature Stores Enable Operational ML
20mNavigating the Skies of Data: A Flight Through Feature Stores
33mEvolution and Scaling of Feature Store at Uber
34mSolving the ML Pain You Forgot to Mention
44mFeature store
15mFeature Stores: Powering Machine Learning Infrastructure with Apache Flink
32mA feature Store for Machine Learning
44mFrom Mistakes to Success: Evolution of the ML Feature Store at Flocktory