How we start a Linux microVM in under 90ms
Cold start is a product feature, not an implementation detail. A walk through the snapshot format, the warm pool, and where the remaining milliseconds actually go.
Most sandbox providers quote a cold start that quietly excludes image pull, network attach, and agent handshake. We measure the number a caller actually experiences: the time from API call to a shell that can run a command.
The snapshot format
Every template is booted once, allowed to settle, then frozen. We capture guest memory alongside device state, so a restore is a page-table remap rather than a kernel boot. The kernel never runs its init path twice.
Where the milliseconds go
At steady state the budget splits roughly three ways: claiming a machine from the warm pool, restoring the snapshot, and attaching the network namespace. Only the last of these scales with concurrency, which is why the pool is sized against arrival rate rather than peak load.
The result holds under load. Claiming from a warm pool lands under 200ms at p99, and a paused sandbox costs nothing while it waits.