RECOMMENDER SYSTEM: Why They Update Models 100 Times a Day // Gleb Abroskin // MLOps Coffee Sessions #123

MLOps.community

MLOps Coffee Sessions #123 with Gleb Abroskin, Machine Learning Engineer at Funcorp, RECOMMENDER SYSTEM: Why They Update Models 100 Times a Day co-hosted by Jake Noble.
// Abstract
FunCorp was a top 10 app store. It was a very popular app that has a ton of downloads and just memes. They need a recommendation system on top of that. Memes are super tricky because they're user-generated and they evolve very quickly. They're going to live and die by the Recommender System in that product.
It's incredible to see FunCorp's maturity. Gleb breaks down the feature store they created and the velocity they have to be able to create a whole new pipeline in a new model and put it into production after only a month!
// Bio
Gleb make models go brrrrr. He doesn't know what is expected in this field, to be honest, but Gleb has experience in deploying a lot of different ML models for CV, speech recognition, and RecSys in a variety of languages (C++, Python, Kotlin) serving millions of users worldwide.
/ MLOps Jobs board
https://mlops.pallet.xyz/jobs
MLOps Swag/Merch
https://mlops-community.myshopify.com/
// Related Links
Putting a two-layered recommendation system into production -
https://medium.com/@FunCorp/putting-a-two-layered-recommendation-system-into-production-b8caaf61393d
Practical Guide to Create a Two-Layered Recommendation System -
https://medium.com/@FunCorp/practical-guide-to-create-a-two-layered-recommendation-system-5486b42f9f63
Ten Mistakes to Avoid When Creating a Recommendation System -
https://medium.com/@FunCorp/ten-mistakes-to-avoid-when-creating-a-recommendation-system-8268ed60aeba
Applying Domain-Driven Design And Patterns: With Examples in C# and .net 1st Edition by Jimmy Nilsson:
https://www.amazon.com/Applying-Domain-Driven-Design-Patterns-Examples/dp/0321268202
--------------- ✌️Connect With Us ✌️ -------------
Join our slack community: https://go.mlops.community/slack
Follow us on Twitter: @mlopscommunity
Sign up for the next meetup: https://go.mlops.community/register
Catch all episodes, blogs, newsletters, and more: https://mlops.community/
Connect with Demetrios on LinkedIn: https://www.linkedin.com/in/dpbrinkm/
Connect with Jake on LinkedIn: https://www.linkedin.com/in/jakednoble/
Connect with Gleb on LinkedIn: https://www.linkedin.com/in/gasabr/

Timestamps:
[00:00] Introduction to Gleb Abroskin
[00:50] Takeaways
[05:39] Breakdown of FunCorp teams
[06:47] FunCorp's team ratio
[07:41] FunCorp team provisions
[08:48] Feature Store vision
[10:16] Matrix factorization
[11:51] Fairly modular fairly thin infrastructure
[12:26] Distinct models with the same feature
[13:08] FunCorp's definition of Feature Store
[15:10] Unified API
[15:55] FunCorp's scaling direction
[17:01] Level up as needed
[17:38] Future of FunCorp's Feature Store
[18:37] Monitoring investment in the space
[19:43] Latency for business metrics
[21:04] Velocity to production
[23:10] 30-day retention struggle
[24:45] Back-end business stability
[27:49] Recommender systems
[30:34] Back-end layer headaches
[32:04] Missing piece of the whole Feature Store picture
[33:54] Throwing ideas turn around time
[36:37] Decrease time to market
[37:41] Continuous training pipelines or produce an artifact
[39:33] Worst-case scenario
[40:38] Realistic estimation of a new model deployment
[41:42] Recommender Systems' future velocity
[43:07] A/B Testing launch - no launch decision
[46:32] Lightning question
[47:08] Wrap up