app-lb

A Pingora-based load balancer that boots Firecracker microVMs on demand: requests route by hostname and path, cold starts hold the request while a VM boots, and idle pools scale to zero.

app-lb is one binary with two listeners:

• The proxy (0.0.0.0:6188, TLS on :6189) — the data plane. Every request is routed to a deployment: a pool of managed microVMs, a fixed set of upstream addresses, or a directory of files served straight off disk.
• The admin API (127.0.0.1:9090) — the control plane: register deployments, watch the dashboard, mint tokens, run exec and shell inside VMs. Loopback and plaintext by design; reach a remote one over an SSH tunnel or through the proxy itself.

Quick start

# start app-lb (it finds the heyvm daemon on the host)
app-lb

# register a deployment from a spec file
serverctl apply -f web.json

# watch it
serverctl get deployments
serverctl describe web

A minimal managed deployment:

{
  "id": "web",
  "routes": [ { "host": "web.example.com" } ],
  "vm": { "driver": "firecracker", "image": "web-rootfs", "port": 8080 },
  "scaling": { "min_replicas": 0, "max_replicas": 3 }
}

With min_replicas: 0 the pool sits empty until a request arrives; the proxy holds the request (up to cold_start_timeout_secs, default 120) while the autoscaler boots a VM, then completes it. After scale_to_zero_after_secs (default 300) of idleness the pool drains back to zero.

Configuration

Everything is environment variables; nothing is required to start.