eugeneyan / eugeneyan-comments

1 stars 1 forks source link

https://eugeneyan.com/writing/feature-stores/ #34

Open utterances-bot opened 3 years ago

utterances-bot commented 3 years ago

Feature Stores - A Hierarchy of Needs

Access, serving, integrity, convenience, autopilot; use what you need.

https://eugeneyan.com/writing/feature-stores/

eggie5 commented 3 years ago

data point for Grubhub recsys: common periodic-offline job creates features for training and serving (eliminating skew). Hive snapshots for access (sharing) and published to Cassandra for serving. heavy run-time feature caching for serving. integrity maintained via ad-hoc monitoring/alerting with datadog and lineage tracking via ml-metadata

Grubhub's recsys use-case is a little more simple than some as we currently don't support real-time features computed in-between offline job runs (typically where you see Flink et al applied). Other groups like logistics might use online-feature generation.

amommendes commented 3 years ago

great framework to think about feature store implementation

hovinh commented 3 years ago

Thank you for the great post, Eugene. I have a few follow-up questions:

By saying features, can you please provide some examples of those need to be centralized? I think statistics measurement are too basic and no need such requirement, except for case a specific metadata used.
I can't help but think that apart from Access level, one must involve human approval to leverage official features to a higher level, otherwise, the cost of duplication or keep the unmeaningful becomes prohibitively expensive. Could you provide some practice are being used?
Could you please explain the term "data leaks"? I have an impression it's not a synonym for "data breach", right?

Thank you :)

eugeneyan commented 3 years ago

I assume you mean centralized as in "stored in a central place". While there's no pressing need to store features in a central location, doing so prevents duplication of effort across multiple teams. In addition, there might be sensitive data (e.g., customer data, financial data) that belongs to a single team and can only be accessed via a central location.
I don't think that's necessarily the case. For example, GoJek found that not having a centralized serving layer led to engineers setting up similar infra and pipelines for each project. That's why they built Feast. Nonetheless, I might be misunderstanding your question.
Data leaks are when label information leaks into the features. For example, trying to predict fraudulent insurance claims while using a feature on the payout—clearly, fraudulent claims would have no payout! More details here.

tobycheese commented 1 year ago

Hi Eugene, would it be possible to add a publication date at the top of your articles? Only with a date statements such as "Last month, Splice Machine, a big data platform, launched its own feature store too." make sense. Thanks!