TechPrufer
Get started
Grab
Grab logoGrab · Feature trail

ML Feature Store Evolution

2 posts · Jul 2025 – Sep 2025

What this trail covers

Grab's ML feature store has evolved from Amphawa to a feature-table centric design using Amazon Aurora Postgres. This new design addresses challenges with high-cardinality features, high-dimensional data, and complex feature management. Key improvements include contextual retrieval for composite keys, atomic updates for entire collections of features, and isolation of read/write capacity. The ingestion workflow now emphasizes atomicity and reproducibility, landing feature tables in the data lake.

All Grab trails
  1. User foundation models for GrabThis post introduces a new foundation model architecture for generating generalized user embeddings. The model jointly trains on tabular and time-series data with diverse modalities (text, numerical, IDs, location) using a transformer architecture. Key innovations include a novel `key:value` tokenization strategy for unifying data types and handling massive entity vocabularies (hundreds of millions of unique entities). The model aims to improve personalization across various downstream tasks by providing holistic user representations.
  2. The evolution of Grab's machine learning feature storeThis post details the evolution of Grab's ML feature store from a feature-centric approach (Amphawa) to a feature-table centric design using Amazon Aurora Postgres. Key technical contributions include: 1. Introduction of 'feature tables' for managing schema and data on a per-table basis, enabling contextual retrieval for composite keys and atomic updates for entire collections of features. 2. Design of a new ingestion framework emphasizing atomicity and reproducibility, involving landing feature tables in the data lake (Parquet on S3), registering and deploying collection summaries, pre-ingestion validation, reverse ETL into Aurora Postgres, and post-ingestion warm-up via shadow reading. 3. Leveraging AWS Aurora for PostgreSQL, specifically its distributed storage architecture, independent scaling of reader and writer nodes (using Serverless for writer and Provisioned for readers), and managed service benefits to meet non-functional requirements like isolation of read/write capacity and cost-effectiveness.