Dragonfly v2.6 Roadmap

Dragonfly

Core Components

Manager

• Enhance service performance and resource utilization while reducing CPU/Memory overhead.
• Enhance user experience and UI design in the Manager Console.
• Support optionally removing the Manager dependency, reducing the minimum cluster deployment unit to Scheduler and Client.

Scheduler

• Enhance service performance and resource utilization while reducing CPU/Memory overhead.
• Optimize the scheduling algorithm to improve bandwidth utilization in the P2P network.

Client

• Enhance service performance and resource utilization while reducing CPU/Memory overhead.
• Implement a bandwidth-aware negotiation protocol to distribute requests across multiple parent nodes,
  preventing single-parent bottlenecks.
• Dfcache/Dfstore support to import persistent cache task to Dfdaemon in Node by UDS.
• Implement RDMA-based distribution of the files.
• Support cache task memory-level download tasks.
• Support reflink to avoid unnecessary data copying.

AI Model/Dataset Distribution

• Writes use Direct IO, reads use Buffered IO.
• Implement Python SDK to provide data distribution for AI Infrastructure.

AI Agent

• Ensure data reliability when asynchronously writing to object storage.
• Implement Python SDK to support snapshotter for use in AI agent.
• Explore integrating the Agent Sandbox, Gymnasium, etc.

Others

Observability

• Improve and refine the monitoring metrics system.
• Optimize the alerting mechanism and enhance issue diagnosis capabilities.

Security

• Implement encrypted data storage.

Testing

• Add more E2E tests and unit tests.

Skills / Agent Capabilities

• Add a Dragonfly skill to enable troubleshooting and diagnosis capabilities.

Documentation

• Enhance the landing page UI.
• Add more documentation on system interactions and implementation details.

No files, tests, or implementation entry points are named. Start by separating the Manager, Scheduler, Client, AI distribution, observability, security, testing, and documentation sections into independently scoped issues, then identify the relevant repository entry points for each. The roadmap alone does not define completion criteria; each resulting issue should state its expected behavior and validation approach.