Skip to content

Secure, Fast, and Extensible Sandbox runtime for AI agents.

7.2 KB
Raw Blame

title: Auto-Pause and Wake-on-Access description: The ingress-driven lifecycle loop for Fast Sandbox — idle observation, automatic checkpoint pause, and wake-on-access with bounded request parking (OSEP-0024).

Auto-Pause and Wake-on-Access

Fast Sandbox sandboxes occupy their full CPU and memory allocation for as long as they live, and agent workloads are idle most of that time. Auto-pause and wake-on-access turn the existing checkpoint machinery into an automatic cycle: a sandbox that opts in is paused after X seconds without ingress traffic, and the next request to it wakes it — the gateway parks the request for up to a budget Y while the sandbox restores, then forwards normally. Both halves default to off; scope is the fast-sandbox backend only.

Full proposal: OSEP-0024. Enablement and configuration: the user guide.

Lifecycle

Auto-pause and wake-on-access lifecycle

Three properties of this state machine carry most of the design:

  • Pausing is cancelable. A wake arriving mid-checkpoint flips the sandbox back to Running — a wrongly-timed pause never completes.
  • Pausing does not stop the expiry clock. A paused sandbox whose expires_at passes becomes Terminated (CR retained) and wake then fails fast with 404.
  • Absence of policy is a hard no-op. A sandbox that never declares lifecycle.idlePolicy is never auto-paused and never auto-woken, regardless of deployment flags.

Design principles

  1. One observation point. All client traffic transits the ingress, so it is the only activity producer and the only wake authority. What the ingress cannot see does not keep a sandbox alive; activeProbe is the single app-declared exception.
  2. Desired state over coordination. PauseSandbox / ResumeSandbox are idempotent CAS patches on the Sandbox CR. Pause needs only a per-sandbox lock (checkpoints must not double-run); wake needs none — duplicate resumes converge because the restore is driven by spec.state, not by callers.
  3. Three-party activation. Server config, ingress flags, and the per-sandbox policy must all agree. An unreadable or malformed policy is treated as absent — never as an implicit opt-in.
  4. Fail-safe direction. Redis down → no pauses (wake does not use Redis). Missing activity key → unknown, tombstoned: data loss can delay a pause, never accelerate one. Any pause failure leaves the sandbox Running.
  5. Bounded wake. One resume flight per sandbox per replica (singleflight); a parking lot and restore permits admit before any RPC; budget exhaustion answers 503 + Retry-After and never cancels an in-flight restore.
  6. Provenance-gated replay. The gateway-owned OpenSandbox-Wake: expired marker proves a request was never forwarded — only marked responses are SDK-auto-retried; post-forward re-entry retries idempotent methods once, everything else keeps today's 503.

How it works

Auto-pause and wake-on-access sequence

👉 Interactive sequence diagram — walk the three phases with guided views, theme switching, and relationship tracing.

Observe. Every proxied request produces one activity observation: SET opensandbox:activity:{id} = now_ms EX ttl, a monotonic-max write issued asynchronously and coalesced per sandbox — fire-and-forget, never on the forwarding hot path. Create and resume write immediately, so the idle clock starts at the latest lifecycle transition. Since X ≤ activity_ttl_seconds is enforced at create time, a present-but-stale key is idle by value; a missing key is resolved by tombstones, not assumptions. A sandbox that declares activeProbe adds a pull-mode producer: when the sweeper is otherwise ready to pause, it probes the app through the gateway (sentinel-carried, never counted as activity) — a busy answer is recorded as ordinary activity, an explicit not-active lets the pause proceed, and any failure drops the candidate.

Pause. Each server replica sweeps every sweep_interval_seconds: enumerate Running sandboxes, batch-read activity keys, then apply the gates in order — opt-in annotation → Running → idle ≥ X → not expired → NX lock → re-verify → probe. Surviving candidates get PauseSandbox (idempotent CAS). There is no leader election: the lock admits one replica per idle episode and idempotency absorbs lost races, so a crashed replica leaves no stuck state. The worst case of the whole subsystem is a missed pause.

Wake. A request whose route resolution reports FailedPrecondition (or that surfaces a stale-route error) reads the sandbox state and policy; Paused/Pausing/Resuming start or join a resume flight. The flight — one per sandbox per replica — issues ResumeSandbox fenced on expected_checkpoint_id, parks its requests, and polls GetSandbox with backoff until Ready or budget Y. Parked requests hold no upstream connection; WebSocket upgrades complete after wake. Budget exhaustion answers the wake-timeout contract while the restore converges in the background — the next request is served on the fast path.

Failure behavior

Failure Outcome
Redis unreachable Sweeps skipped — no pauses; wake unaffected (no Redis on that path)
Redis data loss Tombstone ⇒ pause delayed by ≥ X from first observation, never accelerated
Replicas race to pause Lock + idempotent CAS ⇒ exactly one pause per idle episode
Resume races a re-pause expected_checkpoint_id fence ⇒ Aborted; the flight restarts within budget
Slow (cross-node) restore Budget Y bounds the client; 503 + Retry-After while the restore continues
Burst to many paused sandboxes Lot + permits bound; overflow shed before any ResumeSandbox

Security

The ingress gains exactly two FastPath calls on the channel it already holds (GetSandbox, ResumeSandbox) plus one read-only Kubernetes grant to read the idle-policy annotation. It can never pause. Parking happens after the existing host/route-scope/signature verification, so wake is reachable only for sandboxes a client can already address.

Observability

The loop exposes wake flights, park wait durations, shed count, active restores, dropped activity writes, and probe outcomes. The flap ratio (pauses followed by a quick wake — the primary tuning signal for X) is derived per sandbox from pause and wake telemetry. See the Telemetry guide.

References