Autoscaling

The scaler keeps the fleet the right size and shape. It ingests usage, evaluates rules on a loop, and provisions or removes nodes across cloud providers — all without touching the request path.

Multi-cloud provisioning

The scaler provisions edge nodes through a pluggable provider interface. Built-in targets include AWS, Fly, Hetzner, Vultr, DigitalOcean, E2B, and India-first providers (E2E, Vyom, Utho, Iotamine) — and any provider can be added by implementing the interface. A selector maps a desired (region, ip_type) to a concrete provider and instance.

Self-registration & membership

A freshly provisioned node boots, runs the edge binary, and registers itself in etcd with a short-lived lease (TTL ~30s). If a node crashes, its lease expires and it drops out of the fleet automatically. The scaler keeps an in-memory snapshot of membership by watching the etcd prefix.

A node that doesn’t appear within a boot timeout (~5 min) is treated as failed, and its VM is cleaned up.

Scaling rules

The scaler evaluates rules on a loop (every ~30s), enforcing per-spec and per-node cooldowns so it never thrashes:

Trigger	Action
Constraint gap — a required (region, ip_type) is below its minimum	Scale up: provision a node for that spec
Unhealthy node — error rate > 30% over 5 min, or p99 latency > 5s under low load	Recycle: boot a replacement first, then remove the old node
Idle node — no traffic for ~30 min	Scale down: remove it

Recycling provisions the replacement before removing the old node, so there’s no capacity gap.

Reconciliation

A slower reconcile loop (every ~60s) re-checks for constraint gaps, cleans up orphaned VMs (provisioned but never registered) and ghost nodes (registered but gone), keeping cloud state and etcd in sync.

Metrics feedback to core

The same metrics the scaler computes are exposed over a small gRPC API that core polls every few seconds, so lowest_latency routing and the circuit breaker act on fresh per-node health.