Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cross-Layer Decision Timing Orchestration in Cost-Based Database Systems: Resolving Structural Temporal Misalignment

Ilsun Chang

TL;DR

A cross-layer decision timing orchestration framework that shifts final decision authority from the compile-time optimizer to the runtime executor via selective late binding of operator-level choices is proposed and improves execution stability, reducing P99 latency by up to 20x under severe estimation drift while maintaining comparable median latency.

Abstract

This paper analyzes execution instability in traditional cost-based database management systems (DBMS) and identifies a structural timing misalignment between optimization and execution stages that contributes to tail-latency amplification. Beyond estimation accuracy and raw execution throughput, we argue that decision timing and the availability of runtime signals materially affect robustness under uncertainty. In conventional DBMS architectures, the optimizer relies on historical statistics, the executor observes runtime data distributions and resource states, and accelerators impose up-front transfer costs and amortization constraints. This temporal asynchrony can lead to rigid early-bound decisions that fail under input-scale shifts or stale statistics. We propose a cross-layer decision timing orchestration framework that shifts final decision authority from the compile-time optimizer to the runtime executor via selective late binding of operator-level choices. A Unified Risk Signal (URS) integrates optimizer uncertainty, execution-time observations, and accelerator cost signals without collapsing them into a single static cost model. Experiments on a modified PostgreSQL prototype evaluate (i) input-scale shift, (ii) stale-statistics drift, and (iii) GPU offload break-even regimes using controlled microbenchmarks. The proposed orchestration improves execution stability, reducing P99 latency by up to 20x under severe estimation drift while maintaining comparable median latency.

Cross-Layer Decision Timing Orchestration in Cost-Based Database Systems: Resolving Structural Temporal Misalignment

TL;DR

A cross-layer decision timing orchestration framework that shifts final decision authority from the compile-time optimizer to the runtime executor via selective late binding of operator-level choices is proposed and improves execution stability, reducing P99 latency by up to 20x under severe estimation drift while maintaining comparable median latency.

Abstract

This paper analyzes execution instability in traditional cost-based database management systems (DBMS) and identifies a structural timing misalignment between optimization and execution stages that contributes to tail-latency amplification. Beyond estimation accuracy and raw execution throughput, we argue that decision timing and the availability of runtime signals materially affect robustness under uncertainty. In conventional DBMS architectures, the optimizer relies on historical statistics, the executor observes runtime data distributions and resource states, and accelerators impose up-front transfer costs and amortization constraints. This temporal asynchrony can lead to rigid early-bound decisions that fail under input-scale shifts or stale statistics. We propose a cross-layer decision timing orchestration framework that shifts final decision authority from the compile-time optimizer to the runtime executor via selective late binding of operator-level choices. A Unified Risk Signal (URS) integrates optimizer uncertainty, execution-time observations, and accelerator cost signals without collapsing them into a single static cost model. Experiments on a modified PostgreSQL prototype evaluate (i) input-scale shift, (ii) stale-statistics drift, and (iii) GPU offload break-even regimes using controlled microbenchmarks. The proposed orchestration improves execution stability, reducing P99 latency by up to 20x under severe estimation drift while maintaining comparable median latency.
Paper Structure (21 sections, 1 equation, 7 figures)

This paper contains 21 sections, 1 equation, 7 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. Adaptive Plans in Commercial Systems
  4. Mid-query and Progressive Re-optimization
  5. Continuous Adaptivity
  6. Parametric Optimization
  7. Distributed Runtime Adaptation
  8. Differentiation
  9. The Core Problem: Structural Temporal Misalignment
  10. Temporal Coordinates of DBMS Layers
  11. Structural Limitations of Single-Point Decision
  12. Proposed Architecture: Decision Timing Orchestration
  13. Decision Authority Shift: From Optimizer-Centric to Executor-Centric
  14. Unified Risk Signal (URS)
  15. Execution Time-Based Orchestration Flow
  16. ...and 6 more sections

Figures (7)

  • Figure 1: Cross-Layer Decision Timing Mismatch and Orchestration Structure. This figure visualizes the temporal misalignment and the proposed shift in decision authority.
  • Figure 2: Exp 1: Execution Time Distribution (CDF) - large_noselect. The Orchestrated system maintains a tight distribution compared to the Baseline.
  • Figure 3: Exp 1: P99 Latency Comparison - large_noselect.
  • Figure 4: Exp 2: Execution Time Distribution (Stale Statistics).
  • Figure 5: Exp 2: Tail Latency Comparison (P95 vs P99). The Orchestrated system reduced P99 latency from 3836ms to 333ms.
  • ...and 2 more figures