ByteCard: Enhancing Data Warehousing with Learned Cardinality Estimation
Core Concepts
ByteCard framework improves cardinality estimation accuracy and query optimization in ByteHouse.
Abstract
ByteCard addresses the bottleneck of cardinality estimation in modern data warehouses, specifically focusing on ByteHouse's query optimization.
The framework integrates learning-based methods to balance accuracy and practicality, resulting in significant speed-ups in query processing.
ModelForge Service automates model training for different scenarios, ensuring accurate estimations for various workloads.
Inference Engine provides high-level interfaces for seamless integration with ByteHouse's query processing.
Multi-stage reader strategy is enhanced by prioritizing highly selective columns and dynamically selecting the optimal materialization strategy based on query selectivity.
Join-order selection plays a crucial role in minimizing I/O overhead during query processing.