ScaleSense: Cost-Intelligent Scaling in Alibaba AnalyticDB
A new research paper introduces ScaleSense, a proactive query-level resource scaling framework designed to address the costly 'provisioning trap' in cloud-native serverless data warehouses. The framework, developed for Alibaba AnalyticDB, uses a multi-faceted query encoder that models plan topologies and hardware specifications, and a quantile-based resource predictor to estimate multi-dimensional physical footprints. An auto-scaling controller then navigates resource allocation to avoid over-provisioning and alleviate non-CPU bottlenecks like I/O saturation. The paper, available on arXiv (2608.07945), analyzes production workloads and proposes a solution to optimize resource allocation for heterogeneous ad-hoc queries, potentially saving significant monetary budgets. The work is authored by researchers from Alibaba, and the paper was announced as a cross-type submission.
Key facts
- ScaleSense is a proactive, query-level resource scaling framework.
- It is designed for Alibaba AnalyticDB, a cloud-native serverless data warehouse.
- The framework addresses the 'provisioning trap' where users over-provision resources due to fear of resource depletion.
- Over-provisioning wastes monetary budgets and does not alleviate non-CPU bottlenecks like I/O saturation.
- ScaleSense features a multi-faceted query encoder that models plan topologies and hardware specifications.
- A quantile-based resource predictor estimates multi-dimensional physical footprints.
- An auto-scaling controller navigates resource scaling decisions.
- The paper is available on arXiv with ID 2608.07945.
- The research is based on analysis of production workloads in Alibaba AnalyticDB.
Entities
Institutions
- Alibaba
- arXiv