[Roadmap] Q1 2026 #3452
Description
Higher priority
- Performance (background processing) The research phase is already done, and work can be done incrementally.
- Events (finish pending work) Some essential unfinished parts of event handling (e.g., instance termination) need to be finalized.
- Kubernetes: improve offers #3481 Kubernetes offers (minimum UX improvements) Improve how dstack handles Kubernetes offers, at least for common usage scenarios. Also check what other UX issues may currently prevent users from using Kubernetes via dstack.
- Simplified cluster configuration on GPU clouds Explore how to simplify backend configuration for top GPU clouds with cluster setups (AWS/GCP). For AWS, pay attention to private VPC scenarios.
- Production-grade inference capabilities Prefill–decode disaggregation and high-availability gateways.
- Local disks for high-end GPU clouds Support for local disks can be important, for “high-value” backends (AWS / GCP / Nebius).
Experimental
- K8S-native GPU clouds (experiment with one provider) Explore whether a managed K8S GPU provider (e.g., CoreWeave or similar) can be supported as “Managed K8S” (fleet creates/manages cluster, etc.).
- SSH reverse proxy (experimental) Experiment with allowing the dstack server to proxy SSH traffic inside containers to enable multi-tenancy without host access, remove bastion keys for SSH fleets, improve developer UX, and even possibly simplify private VPC usage?