Using Redis As Your Online Feature Store: 2021 Highlights. 2022 Directions
0 likes198 views
This document discusses using Redis as an online feature store and provides examples of companies that use Redis for this purpose. It highlights that Redis is well-suited for feature stores due to its ultra-low latency, support for multiple data structures, and high throughput. Major companies like DoorDash, AT&T, and Uber are highlighted as using Redis for online feature stores to power applications like fraud detection, recommendations, and personalization that have strict latency requirements. The document predicts increased adoption of online feature stores in 2022 and beyond to support real-time machine learning applications.
![• After successful launch of Michelangelo in 2017, the
engineering team at Uber spent a couple of years
building a scalable platform.
• Key Requirements:
- Low Latency: under 10 ms (P99)
- High throughput: Up to tens of millions QPS
- Extreme Personalization: With millions of users,
restaurants, menu items and even
restaurant-specific models
- Choice of Serving Infra: One size doesn’t fit all. Local
Java Cache, Redis (restaurants), Cassandra (users)
- Efficient batch data writes
- High Availability and Scalability
17
Uber - Michelangelo Palette at Scale
Michelangelo Palette at Scale session by Uber Engineering here
[Redis] It has proven far more cost effective than
Cassandra [for some feature sets]... Today, Redis
handles some UberEats feature sets like
restaurants”
Nicholas Marcott
Sr Software Engineer, Uber](https://image.slidesharecdn.com/applymeetupfeb22-redis-220210194236/85/Using-Redis-As-Your-Online-Feature-Store-2021-Highlights-2022-Directions-17-320.jpg)
Using Redis As Your Online Feature Store: 2021 Highlights. 2022 Directions
- 1. Using Redis As Your Online Feature Store: 2021 Highlights. 2022 Directions Guy Korland, CTO of Incubations, Redis Ed Sandoval, Sr. Product Manager AI/ML, Redis
- 2. Redis, Real-time apps & Feature Stores 2
- 3. Redis at a Glance 3 Open source, in-memory, data structure store, used as a cache, message broker and database. StackOverflow’s Most Loved Database 2017-2021
- 4. Redis at a Glance 4 Open source, in-memory, data structure store, used as a cache, message broker and database. Docker Hub 2022 1B+ docker pulls
- 5. Redis at a Glance (The company) 5 ● “Home of Redis” open source ● Commercial: ○ Redis Cloud ○ Redis Enterprise ● 8,000+ customers worldwide.
- 6. Hashes Bitmaps Strings Bit Field Streams Hyperloglog Sorted Sets Sets Geospatial Search Graph TimeSeries AI JSON Data-Structures Modules BloomFilter Redis beyond cache Lists Redis Enterprise core Linear scalability HA Geo-distribution Durability Backup & restore Tiered-memory access Security Multi-tenant Gears 6
- 7. Redis’ 10,000 foot view 7
- 8. Let’s define real-time 8 User interaction < 100 msec Instant applications decisions
- 9. Redis enables real-time apps 9 User interaction < 100 msec Instant applications decisions
- 10. So…Why Redis in a Feature Store? ● Feature serving to meet ultra low latency & high throughput requirements in real time predictive applications ● Streaming feature support: real time event data ingestion, aggregation & transformation and storage ● Native support for multiple data structures: documents, timeseries, graphs, embeddings, lists, sets, geo-location… ● High availability, scalability and cost efficiency 10 Super fast storage for your online feature store
- 12. 12 Nov 2020 Redis Updates Market Observations DoorDash publishes blog on Redis based feature store Apr 2021 RedisConf 2021 sessions on “Redis as an online feature store” “Redis Vector Similarity Search” Redis collaborated performance improvements to Feast v0.11 Udaan, large B2B platform company. Feast based feature store on Azure Redis and Microsoft collaborate on “Feast on Azure” https://github.com/Azure/feast-azure Numerous examples of companies deploying Redis as an online feature store Jun 2021 2021 - Redis organic adoption & collaborations Uber reports Redis cost effectiveness for some feature sets Jul 2021 Oct 2021 Feb 2022 Aug/Sep 2021 “Serving Features in ms” benchmark blog post
- 14. Fraud Detection Store Ranking Redis is powering ONLINE feature stores 14 Recommendations
- 15. • DoorDash uses Machine Learning (ML) at various places like store ranking (personalized consumer search results), menu item recommendations and many others • Key Requirements: - Low Latency and persistent storage for billions of records: Millions of entities (consumers, merchants, menu items) - High read throughput: “Store ranking” alone generates millions of predictions per second! - Realtime feature aggregation: Listen to a stream of events, aggregates them into features and store them - Fast batch writes: Features are refreshed periodically in batches. - Heterogeneous data types: strings, numeric, lists or embeddings. Serialization and compression made a huge difference in cost - High availability and scalability 15 DoorDash - Gigascale ML Feature Store ”We ran a full-fledged benchmark evaluation on five different key-value stores to compare their cost and performance metrics. Our benchmarking results indicated that Redis was the best option so we decided to optimize our feature storage mechanism, tripling our cost reduction”. Full blog post here
- 16. • Significant drop in “fraud events NOT stopped” as the company moved from rules-based fraud detection approach to 25+ ML model approach • Key Requirements: - Scoring throughput: 10 Million transactions per day - Scoring Latency: Under 50ms - Hundreds of real time features: Constantly monitor real-time events, aggregate and turn into features readily available for scoring. 4x more online features than offline features. Features have a time to live - High Availability and Scalability 16 AT&T - Combating Fraud with Realtime ML “Deploy and Serve AI” component of AT&T ML platform Taken from AT&T session on AI Modernization here
- 17. • After successful launch of Michelangelo in 2017, the engineering team at Uber spent a couple of years building a scalable platform. • Key Requirements: - Low Latency: under 10 ms (P99) - High throughput: Up to tens of millions QPS - Extreme Personalization: With millions of users, restaurants, menu items and even restaurant-specific models - Choice of Serving Infra: One size doesn’t fit all. Local Java Cache, Redis (restaurants), Cassandra (users) - Efficient batch data writes - High Availability and Scalability 17 Uber - Michelangelo Palette at Scale Michelangelo Palette at Scale session by Uber Engineering here [Redis] It has proven far more cost effective than Cassandra [for some feature sets]... Today, Redis handles some UberEats feature sets like restaurants” Nicholas Marcott Sr Software Engineer, Uber
- 18. So…Why Redis in a Feature Store? ● Feature serving to meet ultra low latency & high throughput requirements in real time predictive applications ● Streaming feature support: real time event data ingestion, aggregation, transformation and storage ● Native support for multiple data structures: documents, timeseries, graphs, lists, sets, geo-location, embeddings ● High availability, scalability and cost efficiency 18 Super fast storage for your online feature store
- 19. So…Why Redis in a Feature Store? ● Support for larger feature sets with Redis on Flash ● Global geo-distribution: Active-Passive (additional HA/scalability) & Active-Active (business continuity) ● Flexible deployments: Database-as-as-Service on Cloud, On-Prem or Hybrid ● Fully supported by Redis (the company) 19 Additional operational and cost efficiencies with Redis Enterprise
- 21. Feature store vendors will target ultra-low latency scenarios 21 Benchmark is out TL;DR Redis for Lower Latency Read the full benchmark report https://feast.dev/blog/feast-benchmarks/ “Serving Features in ms with Feast Feature Store”
- 22. Online Feature Store Adoption in 2022+ - Expect increased adoption in traditional mass consumer markets - Financial Services: Fraud detection + Open Banking + New De(centralized) Fi(nance) - Media & Advertising - E-commerce platforms - Digital natives brands demands for personalization at scale (*) - Digital native brands account for 15% of new unicorns funded in 2020 - Sales revenue growing at 3x rate of total e-commerce - Deep knowledge of their customer base & online behavior - Extreme personalization drives consumer engagement, differentiation and loyalty 22 (*) Mckinsey Report: Digital Native Brands born digital but ready to take on the world
- 23. Growing importance of low-latency feature serving in 2022+ - Near real-time features are the “MVP signal” for real-time ML-based apps. - Valuable source of user behavior and intent - “Latency is the new outage” mentality - When latency goes up from single-digit ms to 50ms-70ms, this has a significant knock-on effect on user experience - Geo-distribution for additional scalability, high-availability and business continuity 23
- 24. Thank you. 24 Any Questions? (Other follow up information can also go here.)