Why do you need it?
Agent sandboxes are long-lived but bursty, so they need suspension and rapid resumption.
Currently, a Sandbox is effectively mapped 1:1 to a Pod. Resume recreates the Pod, introducing scheduling, image pulling, container startup, and Agent initialization latency.
How could it be?
Is there any plan to optimize resume latency, for example by decoupling the Sandbox lifecycle from the Pod lifecycle, using pre-warmed runtimes, or supporting process/ memory hibernation?
Kubernetes P50/P95 resume latency benchmarks would also be helpful.
Other related information
- #1448 — Auto idle-pause and resume-on-revisit for long-lived Agent sandboxes
- #1169 — Volume-based pause/resume as an alternative to OCI image commit
- #1354 — Clarification of cold-start and warm-start latency
- #1324 — Lifecycle latency metrics for pause, resume, and kill
Why do you need it?
Agent sandboxes are long-lived but bursty, so they need suspension and rapid resumption.
Currently, a Sandbox is effectively mapped 1:1 to a Pod. Resume recreates the Pod, introducing scheduling, image pulling, container startup, and Agent initialization latency.
How could it be?
Is there any plan to optimize resume latency, for example by decoupling the Sandbox lifecycle from the Pod lifecycle, using pre-warmed runtimes, or supporting process/ memory hibernation?
Kubernetes P50/P95 resume latency benchmarks would also be helpful.
Other related information