How to deploy a recommendation service with a BERT-like model to high-load production
summary
Describes deploying a high-load production recommendation service built with a BERT-like model at Rabota.ru to improve job search user experience. Covers engineering challenges and practical steps for rolling out the model-backed recommendation service in a high-throughput platform environment.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:10Why a custom language model
- 2:13Tokenization issues in HR texts
- 10:01Model training process
- 15:41Vector representation validation
- 18:33Simple starter architecture
- 20:21Cain Index architecture
- 25:22Event-driven architecture
- 35:23Architecture comparison
- 40:25Model updates and monitoring
- 42:38Model implementation results
related talks
34mThe Billion-Request Content Recommendation System Challenge
01h 14mBuilding a Recommendation System with a Hybrid Search From Scratch
35mIntegrating a Spark ML Pipeline into a High-Load Low-Latency Recommendation Service
41mMicroservices for Machine Learning
27mEnhancing Personalization Using Recommendation Systems & Multi-Armed Bandits
23mML-Powered Search & Recommendation 101
41mDynamic Ranking of Search and Recommendation Results in High-Traffic Systems
28mBuilding a Recommendation Engine for Bus Search Based on Real-Time Feedback of Users