Real-time Personalization Scenarios

6 posts · Jan 2021 – Dec 2025

What this trail covers

Grab's CDP Scenarios feature has evolved to enable event-driven, real-time personalization. This includes integrating user-initiated events, geo-fencing, historical CDP data, and on-the-fly predictions. The system empowers marketers with a self-serve GUI and APIs for creating complex event sequences and validating them with synthetic data. Real-time Flink processing evaluates user historical data via StarRocks and performs predictions using pre-trained models. Outcomes are synced to Kafka or Amphawa (DynamoDB). Future enhancements include optimized Scenario sharding for scalable processing, enhanced signal distribution across multiple destinations, and advanced scheduling and delayed triggering of actions.

All Grab trails

Dec 2025How Grab is accelerating growth with real-time personalization using Customer Data Platform scenariosThis post details the evolution of Grab's CDP Scenarios feature, focusing on its technical architecture and impact. It introduces the self-serve GUI for marketers to define event triggers, configure geo-fencing, and incorporate historical attributes or predictive models. The technical architecture highlights real-time Flink processing for evaluating user historical data via StarRocks and performing predictions using pre-trained models. Outcomes are synced to Kafka or Amphawa (DynamoDB). The post also outlines future areas of exploration, including optimized Scenario sharding for scalable processing, enhanced signal distribution across multiple destinations, and advanced scheduling and delayed triggering.
Dec 2023Sliding window rate limits in distributed systemsThis post details the implementation of frequency capping for marketing communications to prevent user information overload and potential spam perception. It addresses two key sub-problems: efficient storage of user segment data and persistent tracking of communications sent per user. For user segment storage, Roaring Bitmaps were chosen over Bloom filters due to better compression, faster set operations, and the ability to handle mixed datasets, with a custom microservice developed to abstract their implementation. For tracking communications, a distributed rate limiter using Redis was implemented. The sliding log rate limiter algorithm was selected for its accuracy, utilizing Redis sorted sets to store request timestamps. Lua scripts were optimized for Redis Cluster mode to ensure atomicity and efficiency of operations like `zcount`, `zadd`, and `zremrangebyscore` for rate limiting decisions and data eviction.
Aug 2023Streamlining Grab's Segmentation Platform with faster creation and lower latencyThis post details the evolution of Grab's Segmentation Platform to address challenges with long segment creation times and high read latency. The core technical contributions include: 1. Adoption of Roaring Bitmaps for efficient storage of user IDs within segments, significantly reducing storage size and improving compression for sparse data. 2. Development of an SDK that handles segment retrieval, decoding, caching (with LRU eviction), and watching for updates, enabling low-latency membership checks (<1ms p99) for downstream services. This SDK allows for set operations on segments on the fly and has enabled new use cases like controlling communication frequency and lifting segment size restrictions in experimentation.
Apr 2022Supporting large campaigns at scaleThis post details the development of a batch job service, part of Grab's real-time automation engine (Trident), to support large-scale marketing campaigns. It addresses the need for quick execution, high turnover rate, and maximized resource utilization by moving from a naive single-threaded approach to a distributed system using Kafka. Key technical contributions include: 1. Using Kafka for distributing batches of users (approx. 100 users per batch) to multiple servers for parallel processing. 2. Implementing partition keys for Kafka to ensure even distribution of batches. 3. Developing batch endpoints in downstream services (Rewards, Messaging) to process multiple users and actions in a single API call, significantly reducing QPS and network latency. 4. Optimizing database queries within batch endpoints by grouping users for batch updates. 5. Further optimizations include sharding Kafka streams by action type and country to enable parallel processing of different campaign types and geographical regions. 6. Making messaging actions asynchronous to avoid blocking subsequent user batch processing.
Sep 2021Automating Multi-Armed Bandit testing during feature rolloutThis post introduces a novel Multi-Armed Bandit System for Recommendation that automates MAB testing simultaneously with new feature rollouts. It details the system architecture, including a stream processing framework, a feedback loop utilizing Kafka Streams and Spark Jobs for metric calculation, and a dynamic value receiver. The Multi-Armed Bandit Optimiser modules (Reward Update, Batched Multi-Armed Bandit Agent, Monte-Carlo Simulation, Adaptive Rollout) are explained, along with strategies for optimizing model traffic distribution. The application of this system to the GrabFood 'Recommended for You' widget is described, focusing on the Effective Conversion Rate as the optimization objective and the use of the mean strategy in Monte-Carlo Simulation for stable distribution. Fallback logic for job failures is also included.
Advertisement
Jan 2021Trident - Real-time Event Processing at ScaleThis post introduces Trident, Grab's in-house real-time IFTTT engine for automating business workflows and campaigns. It details the system's architecture, which consumes events from Kafka, ensures exactly-once processing using Redis and MySQL, and scales to handle billions of events daily. Key technical challenges and solutions discussed include load distribution for Kafka partitions and server auto-scaling, and performance optimizations for rule evaluation through event prefiltering and weighted rule evaluation with lazy loading. The post highlights the system's ability to process over 2,000 events per second and deliver massive volumes of rewards and communications.