Feast Community Newsletter #5

[Danny Chiao]

Nov 24, 2021

Announcements

Feast community calendar

We have a new event calendar, which has been sent to feast-dev and feast-discuss. This calendar will: Show upcoming events like community calls and tech talks. Show RFCs in review for 2 weeks. Note that as per RFC-001 that each RFC is in review for at least one week. Show meetings discussing specific topics of interest in the community outside of community calls (e.g. RFC and GitHub issue discussions)  Folks interested in organizing a meeting can do so within Slack (#feast-development) and tag the Feast team to add an event to the calendar.   If you have thoughts on this, or are interested in helping organize community activity, please reach out to the Feast team!

Blog posts and videos

Check out some new content featuring Feast: [Video] Integrating Feast with KFServing (video) [Video] Self-serve feature engineering using Flyte and Feast (video) [Video] Distributed Feature Store with Feast and Dask (blog, video) [Blog] Considerations for Deploying Machine Learning Models in Production (blog) [Blog] Feast and Arize Supercharge Feature Management and Model Monitoring (post) [Tutorial] Securely deploying Feast on Azure (tutorial) [Podcast] The Feast Podcast - Evaluation Store + Feature Store (podcast)

Upcoming events

The community call this week is cancelled for the holidays. Happy Thanksgiving!

Community contribution highlights

Bug fixes for third party integrations / plugins Snowflake offline store plugin (@sfc-gh-madkins) Optimize memory usage during materialization (@judahrand) Add --host as an option for feast serve (@nossrannug) Fix unexpected feature view deletion when applying edited odfv (@ArrichM) Remove tensorflow-metadata folders (@casassg) WIP Trino plugin (@MattDelac)

Work in progress

feast plan

Running feast apply in production requires careful thought today because users cannot preview the changes it'll make. feast plan intends to improve this workflow by outputting validation errors and infrastructure diffs when running feast apply.   RFC-029: Feast Plan RFC-030: Feast Plan Implementation

Documentation on best practices for deploying Feast

Users have expressed confusion on the best way to deploy and integrate Feast into existing systems, especially when there are multiple teams / projects sharing the same Feast registry.   The team is updating Running Feast in production to give more detailed recommendations on best practices depending on the system needs (e.g. varying by latency requirements). See PR for the first update to these docs.

Benchmarks

Users commonly ask how performant Feast is for online retrieval. The answer depends on the number of features being processed, the choice of offline/online stores, request batch sizes, the choice of python vs Java feature servers, etc.   An RFC will be published soon describing a sample benchmark setup, followed by a blog post with results and recommendations on how to meet their latency requirements with Feast.

Improved Feast typing

Feast under the hood uses protocol buffers to represent types and manages type conversions to/from different stores. Several issues have emerged though: Pandas types / custom Feast types are not flexible enough for users (e.g. nested data to be used in an on-demand transform) Type conversion for various systems is hard to maintain and error prone Serialization / de-serialization is expensive in online serving.    There is active exploration on whether switching to Arrow could solve some of these issues. See more discussion in https://github.com/feast-dev/feast/issues/2013

Miscellaneous

Other noteworthy ongoing work: Data quality monitoring (early investigations, @adchia, @pyalex) Batch transformations (early investigations, @achals, @felixwang9817) Scalable historical retrieval (@MattDelac, @mavysavydav) Astra DB online store plugin (@hamzakpt) Alibaba Cloud plugin (@777ki) Trino offline store plugin (@MattDelac)

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