Feathr: LinkedIn’s feature store is now available on Azure | Azure Blog and Updates

This weblog publish is co-authored by David Stein, Senior Workers Software program Engineer, Jinghui Mo, Workers Software program Engineer, and Hangfei Lin, Workers Software program Engineer, all from Feathr group.

Function retailer motivation

With the advance of AI and machine studying, firms begin to use complicated machine studying pipelines in numerous functions, resembling suggestion methods, fraud detection, and extra. These complicated methods often require tons of to 1000’s of options to help time-sensitive enterprise functions, and the characteristic pipelines are maintained by totally different group members throughout numerous enterprise teams.

In these machine studying methods, we see many issues that eat numerous vitality of machine studying engineers and information scientists, particularly duplicated characteristic engineering, online-offline skew, and have serving with low latency.

Determine 1: Illustration on issues that characteristic retailer solves.

Duplicated characteristic engineering

• In a company, 1000’s of options are buried in numerous scripts and in numerous codecs; they aren’t captured, organized, or preserved, and thus can’t be reused and leveraged by groups apart from those that generated them.
• As a result of characteristic engineering is so essential for machine studying fashions and options can’t be shared, information scientists should duplicate their characteristic engineering efforts throughout groups.

On-line-offline skew

• For options, offline coaching and on-line inference often require totally different information serving pipelines—guaranteeing constant options throughout totally different environments is dear.
• Groups are deterred from utilizing real-time information for inference as a result of issue of serving the best information.
• Offering a handy manner to make sure information point-in-time correctness is essential to keep away from label leakage.

Serving options with low latency

• For real-time functions, getting characteristic lookups from database for real-time inference with out compromising response latency and with excessive throughput may be difficult.
• Simply accessing options with very low latency is essential in lots of machine studying eventualities, and optimizations must be accomplished to mix totally different REST API calls to options.

To unravel these issues, an idea known as characteristic retailer was developed, in order that:

• Options are centralized in a company and may be reused
• Options may be served in a synchronous manner between offline and on-line atmosphere
• Options may be served in real-time with low latency

Introducing Feathr, a battle-tested characteristic retailer

Creating a characteristic retailer from scratch takes time, and it takes way more time to make it steady, scalable, and user-friendly. Feathr is the characteristic retailer that has been utilized in manufacturing and battle-tested in LinkedIn for over 6 years, serving all of the LinkedIn machine studying characteristic platform with 1000’s of options in manufacturing.

At Microsoft, the LinkedIn group and the Azure group have labored very intently to open supply Feathr, make it extensible, and construct native integration with Azure. It’s out there on this GitHub repository and you may learn extra about Feathr on the LinkedIn Engineering Weblog.

Among the highlights for Feathr embody:

• Scalable with built-in optimizations. For instance, based mostly on some inside use case, Feathr can course of billions of rows and PB scale information with built-in optimizations resembling bloom filters and salted joins.
• Wealthy help for point-in-time joins and aggregations: Feathr has excessive performant built-in operators designed for Function Retailer, together with time-based aggregation, sliding window joins, look-up options, all with point-in-time correctness.
• Extremely customizable user-defined capabilities (UDFs) with native PySpark and Spark SQL help to decrease the training curve for information scientists.
• Pythonic APIs to entry every thing with low studying curve; Built-in with mannequin constructing so information scientists may be productive from day one.
• Wealthy kind system together with help for embeddings for superior machine studying/deep studying eventualities. One of many frequent use instances is to construct embeddings for buyer profiles, and people embeddings may be reused throughout a company in all of the machine studying functions.
• Native cloud integration with simplified and scalable structure, which is illustrated within the subsequent part.
• Function sharing and reuse made straightforward: Feathr has built-in characteristic registry in order that options may be simply shared throughout totally different groups and increase group productiveness.

Feathr on Azure structure

The high-level structure diagram beneath articulates how would a consumer interacts with Feathr on Azure:

Determine 2: Feathr on Azure structure.

1. An information or machine studying engineer creates options utilizing their most well-liked instruments (like pandas, Azure Machine Studying, Azure Databricks, and extra). These options are ingested into offline shops, which may be both:
   • Azure SQL Database (together with serverless), Azure Synapse Devoted SQL Pool (previously SQL DW).
   • Object storage, resembling Azure BLOB storage, Azure Knowledge Lake Retailer, and extra. The format may be Parquet, Avro, or Delta Lake.

2. The info or machine studying engineer can persist the characteristic definitions right into a central registry, which is constructed with Azure Purview.
3. The info or machine studying engineer can be part of on all of the characteristic dataset in a point-in-time appropriate manner, with Feathr Python SDK and with Spark engines resembling Azure Synapse or Databricks.
4. The info or machine studying engineer can materialize options into a web-based retailer resembling Azure Cache for Redis with Energetic-Energetic, enabling multi-primary, multi-write structure that ensures eventual consistency between clusters.
5. Knowledge scientists or machine studying engineers eat offline options with their favourite machine studying libraries, for instance scikit-learn, PyTorch, or TensorFlow to coach a mannequin of their favourite machine studying platform resembling Azure Machine Studying, then deploy the fashions of their favourite atmosphere with providers resembling Azure Machine Studying endpoint.
6. The backend system makes a request to the deployed mannequin, which makes a request to the Azure Cache for Redis to get the net options with Feathr Python SDK.

A pattern pocket book containing all of the above circulate is positioned within the Feathr repository for extra reference.

Feathr has native integration with Azure and different cloud providers. The desk beneath reveals these integrations:

Desk 1: Feathr on Azure Integration with Azure Companies.

Set up and getting began

Feathr has a pythonic interface to entry all Feathr elements, together with characteristic definition and cloud interactions, and is open sourced right here. The Feathr python consumer may be simply put in with pip:

pip set up -U feathr

For extra particulars on getting began, please seek advice from the Feathr Quickstart Information. The Feathr group will also be reached within the Feathr group.

Going ahead

On this weblog, we’ve launched a battle-tested characteristic retailer, known as Feathr, which is scalable and enterprise prepared, with native Azure integrations. We’re devoted to bringing extra functionalities into Feathr and Feathr on Azure integrations, and be happy to provide any suggestions by elevating points in Feathr GitHub repository.