Development Roadmap (v0.6.0)

[Syspretor]

Here is the development roadmap for v0.6.0. Contributions and feedback are welcome.

P1: Critical Path Items

• In-place Update Enhancement

  • [ ] Warm-up stage: Introduce pre-pulled images and pre-warmed models to reduce service downtime (Owner: @Syspretor)
  • [ ] In-place recreate: Fallback mechanism when standard in-place updates fail
  • [ ] Resource reservation: Ensure deterministic scheduling during non-in-place updates
  • [ ] Redundant capacity: Warm up spare capacity to accelerate MaxSurge readiness and GPU utilization

• Coordinated Update Improvements

  • [ ] Clarify trigger conditions
  • [ ] State machine tracking
  • [ ] Dependency configuration interactions

• Workload Management

  • [ ] InstanceSet stateful mode: Enable as default workload with Stateful/LWS compatibility
  • [ ] Template optimization: Reduce duplication via templateRef

• Documentation

  • [ ] Mooncake Deployment Guide
  • [ ] End-to-End Upgrade Practices
  • [ ] InstanceSet Deployment Procedures (single-node/multi-node)
  • [ ] Coordination Update Specifications

---

Coordination

• [ ] Coordinated Scaling: Scale specific roles by defined ratios during scaling events

Schedule

• [ ] Flexible Topology Scheduling:
  Multi-level scheduling with hard/soft constraints and weighted preferences
• [ ] Multi-level Gang Scheduling:
  Co-scheduling for dependent pod groups
• [ ] Coordinated Scheduling:
  Enforce affinity/anti-affinity policies between coordinated roles

RoleBasedGroupSet

• [ ] RBGS-level RollingUpdate implementation
• [ ] State machine refinement and status reporting

CLI (rbgctl) Enhancement

• [ ] SLA-driven configuration: Integrate Dynamo AIConfigurator for initial RBG recommendations
• [ ] Lifecycle management improvements

[bcfre]

/assign Coordinated Scaling: Scale specific roles by defined ratios during scaling events