AutoSLO Framework for Latency Management in Cloud Data Warehouses

This document is a technical report that presents AutoSLO, a framework designed for managing latency service-level objectives (SLOs) in multi-cluster cloud data warehouses. The framework addresses the challenges of performance isolation among diverse workloads by enabling each workload to meet its latency SLO more reliably. AutoSLO operates across three timescales through its key components: a Policy Tuner for proactive cluster scaling, an SLO-aware Autoscaler for adjusting active clusters based on workload behavior, and a Query Router that manages query placement to avoid SLO violations. The report details how AutoSLO successfully meets varying latency SLOs while reducing costs by an average of 26.4% compared to traditional methods. Evaluations indicate significant reductions in SLO violation rates, with the Query Router and Autoscaler achieving reductions of 47.8% and 93.7%, respectively. The document outlines the operational efficiency of each component and their roles in ensuring that cloud data warehouses can adapt to workload changes effectively.