Ilmu Komputer & AI editorial
AutoSLO: Practical Latency SLOs on Cloud Data Warehouses -- Extended Version
The core problem
Modern cloud data warehouses decouple compute from storage, which makes it easy for organizations to access the same underlying data with multiple compute clusters. This flexibility is frequently used for performance isolation among diverse workloads so that each workload meets its latency service-level objective (SLO) more reliably. For example, interactive dashboards, ad hoc analysis, and batch jobs can each run on separate clusters.
However, the dedicated-cluster approach requires each compute cluster to be continuously scaled to adapt to workload evolution. Over-provisioning wastes resources, while under-provisioning risks SLO violations. The paper addresses this tension by asking how a multi-cluster warehouse can dynamically right-size its active cluster set and route queries so that latency SLOs are met at minimum cost.
The authors, Markos Markakis and Tim Kraska, present AutoSLO, a latency-SLO-aware workload management framework for multi-cluster cloud data warehouses. AutoSLO operates across three timescales through three key components: a periodic Policy Tuner, an SLO-aware reactive Autoscaler, and an online Query Router. The extended version reports evaluations on reali
Innovation
The authors evaluate AutoSLO on realistic Redbench workloads. The headline result is that AutoSLO successfully meets latency SLOs of varying strictness while reducing cost by a mean of **26.4%** compared to the per-scenario next-best baseline.
Component-level evaluations isolate the contribution of each part of the framework:
- The **Query Router** reduces SLO violation rates by a mean of **47.8%** relative to its corresponding alternative.
- The **Autoscaler** reduces SLO violation rates by a mean of **93.7%** relative to its corresponding alternative.
- The **Policy Tuner** can reduce the SLO violation rate by a mean of **44.6%** using a single day of workload history.
The paper also reports that each component is efficient given its intended operating timescale, indicating that the three-timescale decomposition does not introduce prohibitive overhead. The results span SLOs of varying strictness, suggesting that the approach is not tuned to a single operating point.
Why it matters
The results support the central claim that latency SLOs in multi-cluster cloud data warehouses are best managed by separating decisions by timescale. The Autoscaler's large reduction in violation rates (93.7%) shows that reactive correction of forecast deviation is the dominant lever for SLO compliance, while the Query Router's 47.8% reduction demonstrates that per-query, concurrency-aware placement matters even after cluster-level capacity is set.
The Policy Tuner's ability to reduce violations by 44.6% from a single day of history is practically significant: it suggests that organizations can obtain meaningful proactive benefits without long historical traces. The 26.4% mean cost reduction versus the per-scenario next-best baseline indicates that meeting SLOs need not come at the expense of efficiency; the framework can trade off cost and latency more effectively than dedicated-cluster over-provisioning.
Several limitations and open questions remain. The evaluation is conducted on Redbench workloads, so generalization to other workload mixes and warehouse engines is not established. The framework assumes that workload forecasts and a concurrency-aware latency predictor are available; forecast quality and predictor drift could affect performance. Finally, the interaction between the three components is tuned for the studied timescales, and different operating regimes may require re-tuning.
Overall, AutoSLO provides a practical template for SLO-aware workload management: proactive policy tuning, reactive autoscaling, and online routing, each matched to the timescale at which it is most effective.
Who should read this
Opening member content…