Feast Community Newsletter #2

Skip to first unread message

Danny Chiao

Oct 11, 2021, 1:41:01 PM10/11/21
to Danny Chiao

Oct 11, 2021


Feast 0.14 release

We released a new version of Feast last week! A big thanks to those contributing to the release. This release introduces:

  • [Experimental] AWS Lambda feature servers, which allow you to quickly deploy an HTTP server to serve online features on AWS Lambda. GCP Cloud Run and Java feature servers are coming soon! (see docs)
  • Several bug fixes around performance. The core online serving path is now significantly faster (~30+% speedup in local tests).
  • Several improvements for developer experience. The integration tests are now faster, and temporary tables created during integration tests are immediately dropped after the test.

Experimental features are subject to API changes in the near future. If you have feedback, please reach out to the Feast team!

Blog posts and videos

Check out some new content featuring Feast:

  • [Video] How Robinhood Built a Feature Store Using Feast (video)
  • [Blog] Getting started with Feast on AWS Managed Services (blog post)
  • [Blog] SingleStore comes to the Feast (blog post)
  • [Blog] Operationalize vector search with Pinecone and Feast (blog post)
  • [Blog] An overview of Feast concepts in Chinese (blog post)
  • [Blog] Feast: a feature store practitioner's perspective (blog post)
  • [Tutorial] Feast integration with Flyte (tutorial link)
  • [Documentation] How the Feast test suite works, and how to add / reuse tests (docs)

Upcoming events

Times below are in Pacific time: 
  • Oct 12 @ 10AM: Feast community call (agenda, zoom link)
    • Please add agenda items to discuss! Current topics include Feast 0.14 and Redis benchmarks.
  • Oct 12 @11:05AM: Twitter's Management of ML Features (using Feast) featuring @mavysavydav (talk link)
  • Oct 13 @ 9AM: Feast Deep Dive Meetup featuring @felixwang9817 (event link)

Community contribution highlights

  • Allowing BigQuery store to use non-US GCP regions (@loftiskg)
  • Caching provider for feature retrieval, resulting in local online serving improvements of ~30% (@DvirDukhan)
  • Type conversion fixes during materialization

Work in progress

On demand transforms

Earlier in Feast 0.13, we introduced on demand feature views. Work is underway to build a Feature Transformation Server for scalable on demand feature transformations.


Additionally, in response to community feedback, we are adding the ability to register request data as a separate feature view. 


RFC-021: On-demand Transformations

Streaming support

Initial support for pushing stream data into the online store has begun! Design discussions have shifted the recommended M1 to be a push API client due to complications around authentication in an intermediate ingestion server.


RFC-026: Streaming Support in Feast

Low latency serving

Feast Java's data model has been forward-ported to work with the latest version of Feast. More work is underway to better support plugins for the java repo. We expect to move much of this logic back into the main Feast repo.


Additionally, in a collaboration with Redis, optimizations to python serving and to the Redis online store implementation are underway. Specific benchmarks are being established as well.


RFC-025: Online JVM Client

RFC: Low Latency Serving

Shadow entities

There are use cases where there are multiple aliases for the same entity across different feature views. Initial work around adding feature view aliases and feature view projection cleanups has completed. Shadow entities implementation is in review.


RFC-023: Shadow Entities Mapping

Additional connectors

Join the conversation on Slack, give feedback on our roadmap, or checkout our GitHub!

Reply all
Reply to author
0 new messages