跳转至

ScaleSense:阿里巴巴 AnalyticDB 中基于学习型资源评估的成本智能扩展框架

文章背景与核心概要

云原生无服务器(Serverless)数据仓库通过存算分离架构实现了细粒度的弹性扩展。然而,如何为异构的即席查询(Ad-hoc queries)确定最优资源配置,一直是工业界的一大难题。本文通过对阿里巴巴 AnalyticDB 生产工作负载的深入分析,揭示了“配置陷阱”(provisioning trap)现象:用户因担心资源耗尽而长期过度配置硬件,这不仅造成了巨大的预算浪费,且往往无法解决非 CPU 瓶颈(如 I/O 饱和)问题。

为了解决这一挑战,作者提出了 ScaleSense,这是一个主动的、查询级的资源扩展框架。通过采用多面查询编码器(Multi-faceted query encoder)和基于分位数的资源预测器(Quantile-based resource predictor),ScaleSense 能够精确估计多维物理资源占用。结合自动扩展控制器,该框架成功在性能与成本的帕累托前沿(Pareto frontier)上实现了最优平衡,在严格满足用户性能要求的前提下,将货币成本降低了最高 5.22 倍。


📌 执行摘要

云原生无服务器数据仓库依赖于存算分离来实现细粒度的弹性。然而,为异构的即席查询确定最优资源分配仍然是主要的工业障碍。在本文中,作者分析了阿里巴巴 AnalyticDB 的生产工作负载,并揭示了“配置陷阱”——即由于担心灾难性的资源耗尽,导致用户长期过度配置硬件。这在浪费巨额财务预算的同时,并未解决非 CPU 瓶颈(如 I/O 饱和)问题。

Cloud-native serverless data warehouses rely on decoupling storage from compute to achieve fine-grained elasticity. However, determining optimal resource allocations for heterogeneous ad-hoc queries remains a major industrial hurdle. In this paper, the authors analyze production workloads in Alibaba AnalyticDB and uncover the "provisioning trap"—where fear of catastrophic resource depletion leads users to chronically over-provision hardware. This wastes immense financial budgets without resolving non-CPU bottlenecks (such as I/O saturation).

为了解决这个问题,作者引入了 ScaleSense,这是一个主动的、查询级的资源扩展框架。通过利用多面查询编码器和基于分位数的资源预测器,ScaleSense 可以准确估计多维物理足迹。结合自动扩展控制器,该框架成功地导航了性能-成本帕累托前沿,在严格满足用户定义的性能要求的同时,将货币成本降低了高达 5.22 倍

To solve this, the authors introduce ScaleSense, a proactive, query-level resource scaling framework. By utilizing a multi-faceted query encoder and a quantile-based resource predictor, ScaleSense accurately estimates multi-dimensional physical footprints. Combined with an auto-scaling controller, the framework successfully navigates the performance-cost Pareto frontier, cutting monetary costs by up to 5.22x while strictly meeting user-defined performance requirements.


📊 关键亮点与成果

  • 规模与验证: 在超过 136 万条生产查询上进行了广泛评估。
  • 预测准确性: 实现了具有稳健预测区间覆盖率的最先进预测准确性。
  • 资源配置选择: 在选择最优资源配置方面,比最佳基准方法有 76.7% 的相对提升
  • 成本效率: 在满足自定义性能优化策略的同时,将运营货币成本降低了 最高 5.22 倍
  • 低开销: 保持了极低的推理延迟,证明了其在高吞吐量企业生产部署中的可行性。
  • Scale & Validation: Evaluated extensively on over 1.36 million production queries.
  • Prediction Accuracy: Achieves state-of-the-art prediction accuracy with robust prediction interval coverage.
  • Resource Configuration Selection: Delivers a 76.7% relative improvement over the best baseline approach in choosing optimal resource configurations.
  • Cost Efficiency: Satisfies custom performance-optimization policies while reducing operational monetary costs by up to 5.22x.
  • Low Overhead: Maintains minimal inference latency, proving its viability for high-throughput enterprise production deployments.

🛠️ ScaleSense 的核心架构

  1. 多面查询编码器: 将复杂的查询执行计划拓扑与底层硬件规范进行联合建模。
  2. 基于分位数的资源预测器: 准确预测多维物理资源足迹,作为防止资源耗尽的可靠安全网。
  3. 自动扩展控制器: 动态调整分配,以平衡帕累托前沿上的性能与成本偏好,无需针对变化的业务优先级进行模型重训练。
  1. Multi-Faceted Query Encoder: Jointly models complex query execution plan topologies alongside underlying hardware specifications.
  2. Quantile-Based Resource Predictor: Predicts multi-dimensional physical resource footprints accurately, acting as a dependable safety net against resource exhaustion.
  3. Auto-Scaling Controller: Dynamically tailors allocations to balance performance against cost preferences along the Pareto frontier, requiring no model retraining for changing business priorities.

🔗 链接与资源

license icon