Wake on message

# Things to try

• Press Play and let the readouts settle. Expect about 21 actors in memory and 6.7 cold starts a second. The median message takes 8 ms; the p99 sits at 128 ms, because roughly one message in three finds its actor asleep and waits 120 ms for a load.
• Stretch the idle timeout to 10 s. Cold starts fall to about 1.7 a second, but about 48 actors now stay in memory. Pull it down to 0.5 s and the trade flips: 61% of messages arrive cold, so even the median waits for a load.
• Pause, then click a dashed cell three times (or pick it with the arrow keys and press Enter three times), and Step twice. The actor wakes, loads its state and saves its three messages one at a time. Show the trick to see when the eight busiest actors were awake.

# How it works

Each actor is a counter that lives in a store until something needs it. A message for a sleeping actor wakes it: its state loads, which takes the activation cost, and the message waits in the mailbox. An awake actor works through its mailbox strictly one message at a time. It adds one, saves the result, then takes the next. When nothing has arrived for the idle timeout it drops out of memory; its state is already saved, so nothing is lost.

Each message goes to an actor drawn from a Zipf distribution, where actor k is chosen with weight 1 / k^s. At s = 1 the busiest actor gets 21% of the traffic and the quietest well under 1%. Busy actors never sleep. Quiet ones wake for almost every message.

An actor that receives r messages a second is awake when one arrives with probability close to 1 - exp(-r * timeout). Summed over the grid, that predicts the readouts, and it shows the trade: a longer timeout keeps more actors in memory so that fewer messages pay for a wake. With evenly spread traffic, at s = 0, more than half of them pay.

The costs are illustrative: 8 ms to process and save a message, 60 ms to go back to sleep, 4 kB of memory per awake actor.

A toy model of the pattern. aktorz, one of my public repositories, provides lazily activated, single-threaded, stateful services with private durable storage for Zig, with SQLite persistence and TypeScript bindings.

# Keyboard