Overview
Deploying Fuse for real workloads: orchestrator, hosts, and GPU hosts
Fuse is self-hosted: you run the orchestrator (the control plane) and one or more hosts (each running a host agent) yourself. This section covers bringing all of that up for real workloads, beyond the in-memory stub used in the Quickstart.
What runs where
Deploying the orchestrator
One HTTP API process, the scheduler, and the source of truth: Postgres, or in-memory for small/local deployments.
Firecracker host setup
Bring up a host running fc-agent, which drives Firecracker microVMs.
GPU host setup
Bring up a host running qemu-agent, driving QEMU/VFIO GPU passthrough: whole cards or MIG slices.
Monitoring
Prometheus metrics, health checks, and what to alert on.
Security hardening
TLS, CIDR allowlisting, fail-closed mode, and secrets encryption.
Troubleshooting
Common install and setup problems, and known limitations, with workarounds.
The simplest production topology co-locates the orchestrator on a Firecracker host
and talks to fc-agent over loopback: a single command installs both as systemd
services. Scaling out from there is just registering more hosts; the orchestrator
itself doesn’t change.