A multi-cloud MLOps framework improves AI service reliability through automated deployment, canary releases, and cross-cloud failover. By reducing recovery time and increasing service availability, the research demonstrates how resilient deployment strategies can strengthen large-scale commercial AI systems and support dependable digital infrastructure.
— As companies move artificial-intelligence systems into everyday products, keeping those systems available and safe to update has become as important as the models themselves. In the research paper MLOps Model Deployment System for Multi-Cloud Environments and Improvement of Commercial AI Service Availability, presented at the IEEE International Conference on Communication Systems and Computing (CNC 2025), a multi-cloud deployment architecture is examined as a way to improve the reliability and release safety of commercial AI services.
The work addresses a weakness in how many AI services are run today. A model deployed in a single cloud region is exposed to localized outages and to interference from neighboring workloads, which can translate into missed service-level agreements and slow rollbacks when a release goes wrong. The paper frames this as a problem of operational resilience, asking how to release and recover machine-learning services quickly without sacrificing stability or cost.
To address this, the research builds an MLOps platform spanning three United States cloud regions (AWS us-east-1, Google Cloud us-central1, and Azure eastus2), connected through a multi-cluster service mesh and the KServe inference framework. It adds automated CI/CD/CT pipelines, site-reliability error-budget controls, canary and blue-green releases, and chaos testing, so releases are gated by measured error rates and rolled back automatically when a threshold is crossed.
The system was evaluated over 30 days across three public datasets of different types, spanning Yelp 2024 review text, the UCI Adult tabular data, and the LISA vision data. Latency, error rate, and throughput were collected through Prometheus and OpenTelemetry and measured against service-level objectives and error budgets, letting single-cloud and multi-cloud strategies be compared on the same terms.
Across these tests, the multi-cloud approach reduced mean time to repair by roughly 41 to 58 percent and raised the 30-day service-target achievement rate by about 2.7 to 4.9 percentage points over single-cloud baselines. The proposed active-active configuration reported 99.6 percent service-level-agreement compliance and a 12-minute mean time to repair, against 97.8 percent and 28 minutes for a self-managed single-cloud deployment. Throughput rose close to linearly with concurrency, showing the added control layer did not constrain capacity.
The deployment study builds on a broader interest in applying forecasting and optimization to commercial decisions. In a separate paper, Research on Dynamic Price Prediction of E-commerce Based on Time Series Modeling (International Journal of Business Management and Economics and Trade), a SARIMA-GARCH time-series model is paired with multi-objective optimization to guide dynamic pricing and overseas-warehouse siting for cross-border e-commerce, folding customer time satisfaction in alongside cost.
The author of this research, Huisheng Liu, is a data scientist whose work spans operations research, forecasting, and large-scale experimentation. Drawing on his training in operations research at Columbia University and economics at the University of Washington, his research and applied work connect quantitative modeling with real-world data systems across several areas, including price prediction and demand forecasting using gradient-boosting and deep-learning methods, retrieval-augmented and multi-agent systems for automating data workflows, and customer segmentation for marketing optimization. His work on multi-cloud model deployment and time-series price prediction for cross-border e-commerce further reflects this focus on building scalable, reliable, and commercially relevant data-driven solutions. More recently, he has applied this expertise to user-safety and integrity systems at a major social media platform, where he designs and analyzes large-scale experiments to evaluate platform interventions and improve system performance.
By treating deployment, monitoring, and recovery as part of the machine-learning system rather than an afterthought, Liu’s work offers a practical example of how commercial AI services can be made more dependable as they take on a larger role in digital infrastructure. Its significance extends beyond a single architecture, pointing toward broader use of multi-cloud operations for large-scale and, in future work, large-model and multimodal AI systems.
Contact Info:
Name: Huisheng Liu
Email: Send Email
Organization: Huisheng Liu
Website: https://scholar.google.com/citations?hl=en&user=fTrWD2gAAAAJ
Release ID: 89197413
If you detect any issues, problems, or errors in this press release content, kindly contact [email protected] to notify us (it is important to note that this email is the authorized channel for such matters, sending multiple emails to multiple addresses does not necessarily help expedite your request). We will respond and rectify the situation in the next 8 hours.