Repository navigation
Architecture
Susant Sahani edited this page Oct 18, 2026
·
1 revision
Kubernetes is the source of truth; FluxVM owns VMM lifecycle on each host. Full text: docs/architecture.md.
| Component | Role |
|---|---|
kairon-controller |
Placement, MachineMigration state machine, snapshots (-> CSI VolumeSnapshot), quotas, budgets, fencing |
kairon-node |
One per VM host: reconciles Machines into FluxVM (REST on 127.0.0.1:7788), eBPF policy, mTLS migration peer, optional console relay |
kairon-ui |
Optional dashboard (Go backend + React SPA); just another Kubernetes API client |
kaironctl / kubectl kairon
|
CLI, embedded Helm installer, MCP server; never bypasses the controllers |
kairon-csi-controller, kairon-csi-node
|
CSI driver |
kairon-migration-adapter-* |
Optional VMM-specific migration adapters (a stub and a FluxVM one exist under cmd/) |
flowchart TB
clients["kaironctl, dashboard, MCP agents"] --> api["Kubernetes API (source of truth)"]
api <--> ctl["kairon-controller"]
api <--> node["kairon-node (per host)"]
api <--> ui["kairon-ui"]
node -->|REST 127.0.0.1:7788| flux["FluxVM"]
flux --> vmm["QEMU / Cloud Hypervisor / Firecracker / FluxVM"]
node <-->|mTLS peer| node2["kairon-node on another host"]
- Cold:
Pending -> Stopping -> Restarting -> Succeeded. Host-local disk is not copied. - Live:
Pending -> Starting -> Running -> Cutover -> Adopting -> Succeeded, over a TLS 1.3 mutual-auth peer link. - If the transfer succeeds but the destination commit fails, the migration goes to
NeedsRecovery; nothing is auto-resolved.spec.cancelworks only before commit. - Destination sessions are journaled as
0600files (write, fsync, atomic rename).
Controller and node are stdlib-only (exceptions in docs/DEPENDENCIES.md); no libvirt or KubeVirt dependency.
Kairon wiki, drafted for v0.10.2. Source of truth: the repository README and docs/. Apache-2.0.