Greyship Scaling and sleeping
Services wake on requests, scale up under load and sleep when idle.
Scale to zero
A web service with no requests stays up for a grace window (30 seconds), then pauses. A paused service costs nothing. It stays in memory, so the next request wakes it in milliseconds; that request is simply held until the instance is back, then answered.
After 24 hours without requests, a paused instance is stopped. The next request then starts a fresh one: a cold start, typically well under a second for a Go binary.
Requests are held for up to 30 seconds while a service wakes or starts.
Scaling up
One instance handles up to 100 requests at a time. Beyond that, Greyship starts another instance, up to the
service's maximum (max_instances). Each new instance also goes back to sleep when traffic drops.
Each instance gets 128 MB of memory and up to 1 CPU unless the service asks for more. GOMAXPROCS and GOMEMLIMIT are set to match,
so Go's scheduler and garbage collector work within the limits.
Workers
Workers keep running while deployed during the beta. Waking them only when there's work (a built-in queue, schedules) is coming soon.
What keeps an instance awake
- Requests in flight, including long responses.
- Open WebSockets: an open connection keeps its instance awake.
- Workers, during the beta.
Restarts
Changing an env var restarts the services that use it, one fresh instance at a time, each health-checked before the old one stops. A deploy or rollback starts new instances before stopping the old ones. Either way there's no downtime.