[Roadmap] Q2 2026

[peterschmidt85]

Kubernetes

• Support network and instance volumes
• Support private container registries (registry_auth)
• [Feature]: Multiple Kubernetes clusters #3871 Support multiple clusters per project

Inference

• Support PD-disaggregation with NVIDIA Dynamo/vLLM
• Support multi-replica gateways (high availability)
• Gateways on SSH fleets - requires research

Technical debt

• Implement instance health (incl. GPU health) via Events - create a new event if health status or message per instance changes
• Migrate to Pydantic V2
• Multi-tenancy /SSH proxy docs & UI/CLI integration - 1) document better the current isolation; 2) document better how to use SSH proxy; 3) polish SSH proxy UI/CLI integration

Documentation

• A dedicated API guide with examples - cover all the CLI functionality (in addition to the reference documentation)
• Skills guide (improve skills, plus add a dedicated guide on how to use dstack via agents)
• Distributed training examples, incl. RL - refresh existing examples or/and add better examples, incl TRL; look at Miles

Benchmarks

• PD-disaggregation

Other / Minor

• Orphaned resources - allow dstack server to detect orphaned instances (and other related resources)
• Sandboxes - consider supporting a run configuration type - requires research
• Monarch integration
• CLI performance
• Benchmark the overhead the gateway adds