Ilmu Komputer & AI editorial
Open AccessOA2026
A Resource-centric Analysis and Optimization of NoSQL Workloads using Distressed Resource Volume Metric
Introducing DRV, LoadStar, Luna, and Orbit: an open benchmark pipeline and PAM algorithm for Cosmos DB that cuts resource usage by up to 35% while improving tail reliability at cloud scale.
Gunika Verma; Aashutosh A; Pooja Srinivas; Yogesh Simmhan; Ayush Choure; Harshit Shah; Mayukh Das; Prashant Sasatte; Chetan Bansal; Abhijit Pai; Suraj Dixit; Achint Agrawalยท 2026ยท DOI 10.48550/arXiv.2608.09173
The core problem
Large-scale managed cloud databases rely on sophisticated Load Packing and Migration (PAM) algorithms to run efficiently at scale. Yet research into optimizing resource usage and reliability for these systems is constrained by a scarcity of public NoSQL workloads. The authors address this gap in the context of Cosmos DB, Microsoft's flagship cloud-hosted NoSQL database. Their central contributions are fourfold: (1) open-source NoSQL workloads collected from real Cosmos DB clusters, (2) a novel reliability metric called Distressed Resource Volume (DRV) that captures the quality of service experienced by end users, (3) an open-source policy simulation framework named LoadStar, powered by a non-parametric statistical model of real traffic QoS, and (4) a resource optimization problem for placing Cosmos DB replicas onto VM nodes, together with the Luna forecasting model and the Orbit PAM algorithm. The work forms a reusable benchmark pipeline for validating resource-centric NoSQL policies. Experiments validated with LoadStar demonstrate that Orbit outperforms both the existing Cosmos DB policy and a worst-fit optimized baseline, delivering higher load at lower error rates and reducing r
Innovation
The experiments, validated using LoadStar on the released workloads, compare Orbit against two baselines: the existing Cosmos DB policy and a worst-fit optimized baseline. Across these evaluations, Orbit consistently delivers higher load at lower error rates. Quantitatively, the authors report up to reduction in resources relative to the baselines. This improvement is achieved while maintaining or improving service reliability, as measured by the DRV metric and tail-error behavior. The results indicate that forecast-driven rebalancing of stressed replicas can simultaneously improve efficiency and QoS, breaking the usual trade-off between resource minimization and reliability. The reported gains are not confined to simulation: the approach has been deployed in production, with potential savings of \100M$s per year while improving service reliability for millions of customers. The combination of open traces, a simulation framework, and a deployable PAM algorithm provides a reusable benchmark pipeline for future resource-centric NoSQL research.
Large-scale managed cloud databases rely on sophisticated Load Packing and Migration (PAM) algorithms to run efficiently at scale. Yet research into optimizing resource usage and reliability for these systems is constrained by a scarcity of public NoSQL workloads. The authors address this gap in the context of Cosmos DB, Microsoft's flagship cloud-hosted NoSQL database. Their central contributions are fourfold: (1) open-source NoSQL workloads collected from real Cosmos DB clusters, (2) a novel reliability metric called Distressed Resource Volume (DRV) that captures the quality of service experienced by end users, (3) an open-source policy simulation framework named LoadStar, powered by a non-parametric statistical model of real traffic QoS, and (4) a resource optimization problem for placing Cosmos DB replicas onto VM nodes, together with the Luna forecasting model and the Orbit PAM algorithm. The work forms a reusable benchmark pipeline for validating resource-centric NoSQL policies. Experiments validated with LoadStar demonstrate that Orbit outperforms both the existing Cosmos DB policy and a worst-fit optimized baseline, delivering higher load at lower error rates and reducing resources by up to . These results have been deployed in production, with potential savings of \100M$s per year while improving reliability for millions of customers.
The methodology proceeds in four stages. First, the authors collect and open-source NoSQL workloads from real Cosmos DB clusters, providing a trace corpus that reflects production traffic patterns. Second, they derive the Distressed Resource Volume (DRV) metric from these traces to quantify end-user QoS. DRV aggregates the volume of resources operating under distress, i.e., in states that degrade user experience. Formally, if denotes the distress indicator for replica at time and its resource consumption, then a representative formulation is:
Why it matters
The paper's core insight is that reliability in managed NoSQL databases should be measured in resource terms, not only in latency or error counts. DRV operationalizes this by weighting distressed resources by their volume, making it a user-centric QoS signal that can be directly optimized. LoadStar addresses the reproducibility crisis in cloud database research by offering a non-parametric statistical model of real traffic, allowing policies to be tested without access to proprietary production systems. Luna and Orbit then close the loop: Luna forecasts future load distributions, and Orbit uses those forecasts to trigger and rebalance stressed replicas before tail errors materialize. This proactive stance contrasts with reactive PAM policies that respond only after degradation occurs. The reported resource reduction and \100M$s/yr potential savings demonstrate that the approach scales economically. However, the work also implies open questions: how sensitive is DRV to the definition of distress, how well does Luna generalize across workload shifts, and what are the migration costs and stability guarantees of Orbit under adversarial or highly bursty traffic? The open-source release of workloads and LoadStar invites the community to probe these questions, extending the taxonomy of resource-centric optimization for NoSQL and adjacent distributed systems.
Who should read this
CS practitioners and researchers
Opening member contentโฆ