Overview

In this talk, we will describe ShareChat’s feed recommendation infrastructure in detail. The talk will delve into various ml-infrastructure related aspects, such as model training, serving, design, and development of feature-store, and also feature-computation pipelines. The subject matter will also provide insights and learnings that we have obtained via building these large-scale, low latency and fault-tolerant systems. Additionally, we will describe strategies we have employed for disaster management and automated recovery.

Outline of the talk:

Introduction

Scale of our feed system

General outline of the recommendation pipeline

Model Training at ShareChat

• Incremental models
• Real time models
• Ranker related models

Model serving:

• Realtime model inferencing , model versioning, hot reloading and tools that we use.

Feature Store:

• Requirements
• Initial version of feature store at ShareChat, what we tried, what didn’t work.
• Current structure

Feature computation pipelines:

• Batch and real time pipelines with details

Feature logging and model monitoring

Disaster management:

Learning and Conclusion:

Future Directions