PodcastCalibration for ML at EtsyClaimErica Greene said the team’s move from heavily engineered features to neural networks was an early major win, with most of the work involving operations and data engineering rather than modeling.14:43
Feature stores, and second thoughts in 2022
22 sessions
PodcastCalibration for ML at Etsy
PodcastPlatform Thinking: A Lemonade Case StudyClaimLemonade's platform provides point-in-time data directly from Snowflake and gives researchers access to large dimension tables with pre-created features.10:16
MeetupTrustworthy Data for Machine LearningClaimConvoy’s data platform team owns data infrastructure from instrumentation and ETL through Snowflake, orchestration, data discovery, experimentation, machine learning deployment, and a feature store.9:34
Reading groupFeature Stores at Shopify and SkyscannerPushed backMatt Delacour says feature engineering is outside the feature store project, while the question concerned whether feature services contain feature-engineering code.25:26
PodcastLessons from Studying FAANG ML SystemsClaimThe five fundamental ML platform components are a feature store, workflow orchestration, a model registry, model serving, and model monitoring.6:04
Reading groupMachine Learning Engineering in ActionClaimBen Wilson says deployment should include monitoring, failover logic, controls for nonsensical predictions, and checks for bias and feature-data problems.11:37
MeetupBuilding a Modern Data Analytics StackClaimThe feature-engineering query groups events by user and creates columns containing event counts, while the apply-button event is treated as a binary target.22:17
MeetupModern Data Science with VaexClaimVaex ML uses Vaex expressions and transformers to build machine-learning features and pipelines without materializing every intermediate result in memory.48:05
Scaling Real-time Machine Learning at ChimeClaimThe Pay Friends inference service combines transaction details with feature information about senders and recipients before making a risk decision.9:27
MeetupOn Juggling, Dr. Seuss and Feature Stores for Real-time AI/MLClaimAn end-to-end real-time prediction includes network latency, the application pipeline, feature serving, and model scoring.13:40
MeetupBuilding a Movie Recommendation System on Tecton with SnowflakePushed backDavid Hershey says Tecton does not currently provide built-in automatic drift detection, while external tools are commonly used to profile feature data.47:52
PodcastMLOps + BI?ClaimEntity-centric data modeling associates a clear entity with many features, attributes, facts, and metrics.7:15
MeetupSo Fresh and So Data CleanClaimMage combines interactive notebook features with text-editor practices such as clean code, reviewability, versioning, modularity, and reproducibility.3:47
PodcastTurning Redis into a Composable, ML Data PlatformClaimRedis is being expanded beyond traditional website caching into a composable database with modules for JSON, full-text search, vector search, and online feature stores.5:12
PodcastMLOps at DoorDashClaimThe DoorDash machine learning platform aims to cover the applied machine learning lifecycle end to end and at scale, from feature engineering through model serving.4:23
PodcastFeathr: LinkedIn's High-performance Feature StorePushed backReal-time features should not be treated as universally necessary because many signals and prediction targets change slowly.40:04
PodcastRecommender System: Why They Update Models 100 Times a DayPushed backGleb Abroskin rejected the description of FunCorp's system as a unified feature store with declarative transformations and one API for offline and online stores.15:00
PodcastLet's Continue Bundling into the DatabasePushed backEthan Rosenthal argued that streaming databases may support feature stores and model monitoring, while Mike Del Balso said they were not yet a complete replacement for feature stores and remained early for large-scale requirements.30:52
Podcast"Real-Time" ML: Features and InferencePushed backSkylar Payne initially presents himself as skeptical of real-time features, while later saying they should be made broadly usable if infrastructure makes them simple.2:23