Serving ML Models at a High Scale with Low Latency // Manoj Agarwal // MLOps Meetup #48

MLOps community meetup #48! Last Wednesday, on his birthday, we talked to Manoj Agarwal, Software Architect at Salesforce.
// Abstract:
Serving machine learning models is a scalability challenge at many companies. Most applications require a small number of machine learning models (often less than 100) to serve predictions. On the other hand, cloud platforms that support model serving, though they support hundreds of thousands of models, provision separate hardware for different customers. Salesforce has a unique challenge that only very few companies deal with; Salesforce needs to run hundreds of thousands of models sharing the underlying infrastructure for multiple tenants for cost-effectiveness.
// Takeaways:
This talk explains Salesforce hosts hundreds of thousands of models on a multi-tenant infrastructure to support low-latency predictions.
// Bio:
Manoj Agarwal is a Software Architect in the Einstein Platform team at Salesforce. Salesforce Einstein was released back in 2016, integrated with all the major Salesforce clouds. Fast forward to today and Einstein is delivering 80+ billion predictions across Sales, Service, Marketing & Commerce Clouds per day.
//Relevant Links
engineering.sa...
engineering.sa...
---------- ✌️Connect With Us ✌️------------
Join our Slack community: go.mlops.commu...
Follow us on Twitter: @mlopscommunity
Sign up for the next meetup: go.mlops.commu...
Catch all episodes, Feature Store, Machine Learning Monitoring and Blogs: mlops.community/
Connect with Demetrios on LinkedIn: / dpbrinkm
Connect with Manoj on LinkedIn: / agarwalmk
Timestamps:
[00:00] Happy birthday Manoj!
[00:41] Salesforce blog post about Einstein and ML Infrastructure
[02:55] Intro to Serving Large Number of Models with Low Latency
[03:34] Manoj' background
[04:22] Machine Learning Engineering: 99% engineering + 1% machine learning - Alexey Gregorev on Twitter
[04:37] Salesforce Einstein
[06:42] Machine Learning: Big Picture
[07:05] Feature Engineering
[07:30] Model Training
[08:38] Machine Learning: Big Picture
[08:53] Model Serving Requirements
[11:09] Model Serving: Able to Predict
[12:52] Model Serving Requirements
[13:01] Do you standardize on how models are packaged in order to be served and if so, what standards Salesforce require and enforce from model packaging?
[14:29] Support Multiple Frameworks
[16:16] Is it easy to just throw a software library in there?
[19:20] Support Multiple Frameworks/Versions
[22:50] What does gRPC stand for? (Google Remote Processing Core)
[24:36] Support Multiple Model Versions
[27:06] Along with that metadata, can you break down how that goes?
[28:27] Low Latency
[30:29] Scaling for RPS (Request Per Second): Replication
[31:15] Scaling Number of Models: Sharding
[32:30] Model Sharding with Replication
[33:58] What would you do to speed up transformation code run before scoring?
[35:55] Model Serving Scaling
[37:06] Noisy Neighbor: Shuffle Sharding
[39:29] If all the Salesforce Models can be categorized into different model types, based on what they provide, what would be some of the big categories be and what's the biggest?
[44:18] Avoid Hotspots: Per Model Scaling
[46:27] Retraining of the Model: Does that deal with your team or is that distributed out and your team deals mainly with this kind of engineering and then another team deals with more machine learning concepts of it?
[47:26] Do you track the quality of the data going into the model and if it changes over time that would probably be on the other side of the machine learning?
[49:15] Do you think this will be open-sourced or is it going to be that makes us all jealous?
[50:13] How do you ensure different models created by different teams for data scientists expose the same data in order to be analyzed?
[52:08] Are you using Kubernetes or is it another registration engine?
[53:03] How is it ensured that different models expose the same information?