Powering Real-time Recommendations with Machine Learning on GCP
summary
Real-time recommendation stack on GCP uses LiftIgniter logic with Aerospike key‑value clusters to serve millisecond lookups and serve a continuous update pipeline. The deployment stores about 1 billion unique objects (~2 TB unique per cluster, five nodes), handles a few thousand TPS, and runs with less than one full‑time engineer for maintenance. User‑history updates are near‑instant, item‑level signals propagate in 5–30 minutes, heavier UDF updates run on a separate cluster, and Kubernetes/small‑node approaches proved counterproductive.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:10What is Maven
- 3:13Our journey
- 4:57Realtime first
- 6:35How we use Aerospike
- 7:43High Compute Intensity Updates
- 8:33Data Storage
- 11:53Challenges
- 13:14Application robustness
- 15:12Failed experiments
- 18:54Moving from AWS to GCP
- 19:58Experience with GCP
- 20:32Realtime recommendations
- 23:51Model updates
- 27:17Leveraging Kubernetes
- 29:38Moving query logic to the database
related talks
32mResurrecting a Recommendations Platform
24mLarge Scale On-Demand Low-Latency Near Real-Time Predictions
20mDataEngBytes 2023 - PER-T1-06 - Kamal Hossain
12mQuickUpdate: a Real-Time Personalization System for Large-Scale Recommendation Models
41mBig Data Serving: The Last Frontier. Processing and Inference at Scale in Real Time
33mBuilding Recommendation Systems for Global E-commerce Marketplaces
28mThe Power of Personalized Customer Experiences Through AI Innovations
14mMaximizing Aerospike Administration in GCP