Skip to main content
A stateful actor instance is activated on demand, kept in memory while it’s busy, evicted when idle, and can be restarted or relocated at any time. You don’t manage any of this — but you design around it, because only what you write to storage survives.

​
Where it Runs

An instance is placed on first touch — the first method call to a name brings it up. After that you address it by name, never by location; the platform routes every call to wherever the instance currently lives. (idFromName accepts a locationHint option, reserved for regional placement — see Actor Namespace.)

​
Activation, Sleep, and Eviction

  • Activated on demand. Naming an instance is cheap; the first method call to a name brings the instance up and routes the call to it.
  • Kept warm while busy, evicted when idle. Between calls, an idle instance may be removed from memory to free resources, then re-activated on the next call.
  • Scales to zero like functions. Where the actor runtime runs on Knative, an idle instance sleeps after an idle period — the same way a function container scales to zero after idling — unless something is holding it resident: an actor with an open WebSocket stays active between frames and does not sleep. (Alarms also wake a sleeping instance — below.) The first request to a sleeping instance pays a cold start (pod boot + activation) before the method runs, just like a function’s first invocation after idle. Duration billing is unchanged in spirit: you pay for active seconds only — a sleeping instance does not accrue.
  • Alarms wake a sleeping instance. Where scale-to-zero is enabled, a scheduled alarm() fires even if the instance is scaled to zero — the alarm delivery brings the pod back up. Alarms are the one trigger (besides requests) that wakes a sleeping instance.
  • Restarted on deploy or host failure. A new code version or a host issue tears the instance down and re-creates it.

​
What Survives a Restart

In-memory state — instance fields like this.x — is a cache. It vanishes on eviction, restart, or relocation. Your actor’s storage is the truth: it is durable and reloaded when the instance next activates. Treat memory exactly like a process that can be SIGKILLed at any instant — only what you flushed to storage is real. (This is guarantee 4 in Execution model.) A common pattern is to load from storage once on activation and cache the result in memory; just never assume the cache is still there.

​
Next Steps