feat(container-runner): sleep on startup idle timeout

[abcxff]

No description provided.

[abcxff]

Stack for rivet-dev/actors

• #5594 feat(container-runner): drain engine sleeps unless the idle …
• #5595 feat(container-runner): jitter idle timeout to avoid teardow …
• #5585 feat(container-runner): self-sleep on repeated actor start
• #5586 feat(container-runner): sleep on startup idle timeout 👈
• #5589 chore(container-runner): tighten comments
• #5587 feat(container-runner): exit process when the last child sto …
• #5588 feat(container-runner): drain child before SIGTERM on engine …
• #5584 feat(container-runner): drain children and engine concurrent …
• #5565 fix(envoy-client): ack terminating stop commands so pegboard …
• main

Get stack: forklift get 5586
Push local edits: forklift submit
Merge when ready: forklift merge 5586

_{change ontuyypk}

[claude]

Code Review

Reviewed container-runner/src/actor.rs / main.rs changes adding a one-shot startup idle timeout (RIVET_IDLE_TIMEOUT_SECS) that sleeps the actor if no request arrives within a window.

Bugs

1. on_sleep skips drain for all sleep causes once idle timeout is configured, not just idle-triggered sleeps (container-runner/src/actor.rs:317-321)

async fn on_sleep(self: Arc<Self>, ctx: Ctx<Self>) -> Result<()> {
    if idle_timeout().is_some() {
        self.stop_child(ctx.actor_id(), "actor sleeping (idle)").await;
    } else {
        self.drain_then_stop_child(ctx.actor_id(), "actor sleeping").await;
    }
    Ok(())
}

The branch is keyed on whether RIVET_IDLE_TIMEOUT_SECS is configured, not on whether this particular sleep was caused by the idle timer. The method's own doc comment notes the engine can sleep an actor for reasons other than idle (dashboard action, eviction, crash policy). With idle timeout enabled, any of those non-idle sleeps now skip the drain grace period entirely and SIGTERM the child after only effective_stop_grace() (default 9s) instead of drain_grace() (default 900s) — abruptly killing long-running in-flight connections (e.g. an active game match) even though the actor had long since received traffic and moved past IDLE_REQUESTED.

The existing idle_state field already distinguishes "still armed, never got a request" from "got a request" — consider gating the fast-stop path on idle_state.load(...) != IDLE_REQUESTED (or similar) instead of on whether idle-timeout mode is globally enabled, so drain still applies once the actor has served traffic.

2. The duplicate-start early-return path never arms the idle timer for the new Ctx (container-runner/src/actor.rs:154-167)

if let Some(existing) = children().read_async(&actor_id, |_, c| c.clone()).await {
    if !existing.has_exited() {
        ...
        register_ctx(&actor_id, &ctx).await;
        *self.child.lock().await = Some(existing);
        return Ok(());
    }
}

Compare with the normal-start path, which calls self.arm_idle_timeout(&ctx, actor_id) after spawning. If the engine retries a Start (the comment above this block explicitly calls out that this must be idempotent for retries), the new GameServer/Ctx generation takes over the shared child but never arms its own idle timer. Meanwhile the original generation's timer, if it never itself observed a request, is still running against its own idle_state/ctx. When that fires, ctx.sleep() is called on the stale generation, which (per finding #1's stop_child path) removes the actor from the registry and kills the child process — the same process the new generation is actively serving from. Consider arming (or re-arming) the idle timer in this branch as well, or explicitly documenting why it's intentionally skipped.

Minor / style

3. idle_state as AtomicU8 with two named constants is more than the state needs (container-runner/src/actor.rs:27-36)

const IDLE_ARMED: u8 = 0;
const IDLE_REQUESTED: u8 = 1;
...
idle_state: AtomicU8,

The field is only ever compared for equality with IDLE_REQUESTED or stored as IDLE_REQUESTED — a one-shot flag. A plain AtomicBool (idle_requested: AtomicBool, default false, store(true, ...) on request) expresses the same semantics without the constant pair and is more idiomatic for a two-state flag.

Other notes

• Good use of tokio::select! with the abort signal in arm_idle_timeout to cancel the timer early on shutdown.
• idle_timeout() correctly treats 0/unset as disabled via (secs > 0).then(...).
• No test coverage included for the new idle-timeout behavior (arming, disarming on request, sleep-triggering, or the drain-skip interaction in on_sleep). Given the drain-skip bug above, a test exercising "actor receives a request, then engine sends an unrelated sleep" would have caught it.

No security concerns; this is a container-runner-local lifecycle feature and doesn't cross the client/engine or envoy/pegboard-envoy trust boundaries.