TechPrufer
Get started
Grab
Grab logoGrab · Feature trail

Ad Server Infrastructure

1 post · Feb 2022 – Sep 2023

What this trail covers

Grab has evolved its ad serving infrastructure to include a scalable lookalike audience platform. This platform leverages user and audience embeddings for efficient audience creation and retrieval. Key improvements include a high-performance in-memory retrieval service with passenger embedding compression, and an embedding-based audience creation pipeline using a small neural network. The system aims to reduce SLA for audience creation, improve scalability, increase updating frequency, and lower operational costs. The ad server itself is a microservice-based system that handles targeting, capping, pacing, scoring, ranking, and pricing of ads, utilizing Elasticsearch for ad targeting and ScyllaDB for storing real-time statistics like budget spent and ad impression counts. Latency, graceful fallbacks, and global optimization are key design principles. Future work includes optimizing real-time ad decisions with ML, expanding ad types, and scaling services.

All Grab trails
  1. Stepping up marketing for advertisers: Scalable lookalike audienceThis post details the development of a scalable lookalike audience platform for Grab's advertisers. The solution addresses challenges of long SLA, low scalability, low updating frequency, and high cost of audience creation by leveraging user and audience embeddings. A real-time in-memory retrieval service with a passenger embedding compression algorithm (reducing memory needs by ~90%) and an embedding-based audience creation/updating pipeline using a low-cost small neural network are introduced. The audience creation process involves averaging passenger embeddings to generate audience embeddings, and audience serving utilizes cosine scores between user and audience embeddings stored in memory. Embedding compression is achieved using hash tricks inspired by Bloom Filters. The impact includes a 15-minute audience visibility, doubled impressions/clicks, improved CTR/conversion/ROI, and a 98% decrease in audience generation costs.
  2. How Grab built a scalable, high-performance ad serverThis post details the design and implementation of Grab's in-house ad server, a microservice-based system built to scale the ads business. It outlines the ad serving flow, including targeting (using Elasticsearch), capping and pacing (using ScyllaDB for stats), scoring, ranking, and pricing. Key principles of latency, graceful fallbacks, and global optimization are emphasized. The post also discusses the rationale for building a custom solution over off-the-shelf options, highlighting the need for integration with Grab's ecosystem and ML personalization capabilities. Future challenges and opportunities in ML-driven personalization and auction strategy are mentioned.