AnalyticDB Ray enables fully managed Ray-based Data+AI with heterogeneous CPU/GPU scheduling (Qwen-VL fine-tuning case)

AnalyticDB Ray is a fully managed Ray service on Alibaba Cloud AnalyticDB for MySQL that supports cloud-native Data+AI workloads with hybrid CPU/GPU resource scheduling. The article describes PB-level multimodal processing and distributed fine-tuning scenarios, including Qwen-VL multimodal model fine-tuning through LLamaFactory integrated into AnalyticDB Ray.

Industry
Tech & Comms
Location
China
Published
July 2026

Reported outcomes

+80%

GPU utilizationProductivity & throughput

1.8xoverall performance5 mstask scheduling latency4,000,000 tasks/daydaily task schedules80,000 tasks/secondpeak task scheduling throughput

Strategic outcomes

Cost efficiencyLower barrier to entry for Ray-based data and AI workloadsScale & capacityEnabled integrated Data+AI architecture for multimodal workloadsOther strategic outcomeSupported distributed fine-tuning of Qwen-VL multimodal models

Catalog median for productivity & throughput deployments: +45% across 225 reported metrics. Compare benchmarks →

Primary read

Use case focus

Showing 3 of 3

  • 1Data platform modernization
  • 2Multimodal analytics
  • 3AI model training
  • Enterprises processing multimodal Clip data needed to handle complex pipelines across video, point clouds, radar, GPS, and control signals.
  • Traditional solutions had complex configuration, low efficiency, inflexible resource scheduling, and difficult multimodal data management.
  • Alibaba Cloud launched AnalyticDB Ray as a fully managed Ray service with one-click deployment, full-link observability, automatic elasticity, and fine-grained heterogeneous CPU/GPU scheduling.
  • The service integrates with AnalyticDB for MySQL, Ray Object Store, Lance, and LLamaFactory to support distributed multimodal processing and Qwen-VL fine-tuning.
The article reports overall performance improved by 1.8 times, GPU utilization exceeded 80%, scheduling latency reached 5 ms to 20 ms, and one customer scenario handled nearly 4 million task schedules per day with peaks near 80,000 per second.
Architecture

AnalyticDB Ray is described as a fully managed Ray service on AnalyticDB for MySQL with a cloud-native Data+AI architecture. It uses a unified CPU/GPU resource pool, a pipeline stream computing engine, Ray Object Store for intermediate data, adaptive scheduling based on job profiles, built-in high-performance video processing operators, and runtime fencing via separate conda environments. The service integrates with AnalyticDB lakehouse storage, Lance, and LLamaFactory for multimodal workloads and model fine-tuning.

Sources & evidence1
Groundedness: 5/5Type: Blog PostPublished: Jul 11, 2026Publisher: Alibaba CloudEvidence: VendorConfidence: Medium

AI-generated summary. Verify important details with the linked sources before relying on this case.

Explore related AI use cases

Was this useful?