Risk-Aware GPU-Assisted Cardinality Estimation for Cost-Based Query Optimizers

TL;DR

This paper introduces GACE, a GPU-assisted hybrid approach for more accurate and stable cardinality estimation in query optimizers, especially under challenging workload conditions, by selectively using GPU measurement.

Contribution

The paper proposes GACE, a novel hybrid architecture that integrates GPU-based measurement into query optimization, improving stability and reducing latency in risky estimation scenarios.

Findings

GACE improves plan stability in skewed and correlated data scenarios.

GACE reduces tail latency (P99) in problematic workload conditions.

Selective GPU measurement balances overhead and accuracy effectively.

Abstract

Cardinality estimation is a cornerstone of cost-based optimizers (CBOs), yet real-world workloads often violate the assumptions behind static statistics, degrading decision stability and increasing plan flip rates. We empirically characterize failures caused by stale statistics, skew, join correlations, hidden distributions in bind variables, and sampling bias, and quantify the overhead and break-even points of hardware-accelerated measurement. We propose GACE (GPU-Assisted Cardinality Estimation), a hybrid auxiliary architecture that augments rather than replaces the optimizer. GACE selectively invokes GPU-based measurement only in risky intervals via a Risky Gate that detects estimation uncertainty, and a GPU Measurement Engine that performs high-speed probing with explicit cost accounting for the measurement itself. This design preserves low overhead in stable regions while improving plan stability and reducing tail latency (P99) in problematic scenarios.