Overview
This series examines how Kata Containers and Firecracker can provide strong guest isolation while leaving memory, storage, eviction, and billing boundaries misaligned. It moves from sandbox cgroups and delegated I/O to first-touch page charging, path-based storage gaps, combined economic effects, and co-tenant harm. Its bounded conclusion is diagnostic: identify the physical resource, host owner, observer, governing control, and failure mode; local measurements and modeled savings are not universal deployment claims.
Learning outcomes
- MicroVM isolation does not automatically make guest use, host charge, eviction, and billing share one boundary.
- Kata’s cgroup modes are governance contracts; separate sandbox overhead still needs an owner, budget, and control.
- Delegated I/O can charge memory to a host backend and leave storage outside a path-based observer’s view.
- The study’s roughly 500× memory result and near-42% modeled bills are configuration-bound, not universal rates.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
Do Kata Containers and Firecracker resource limits include all host work?
Not automatically: guest isolation can remain strong while VMM, virtio, page-cache, or storage work is charged or observed outside the intended tenant boundary. The answer depends on backend, cgroup placement, storage observer, and billing model.
Why can Kubernetes show low usage while a microVM consumes host resources?
Memory may be charged to the host backend that first touches file-backed pages, while path-based storage accounting may miss deleted-but-open or passthrough objects. The guest caused the use, but Kubernetes sees a narrower cgroup or path boundary.
How can operators distinguish sandbox overhead from an accounting gap?
Check cgroup placement, backend growth, guest access, and page or storage-object ownership. Expected separate overhead should be sized and governed; unexplained backend growth or retained objects outside the observer’s domain indicates an attribution or visibility problem.
What do the study’s billing and co-tenant results prove?
They show, under selected modeled and local conditions, that accounting gaps can reduce modeled payment to about 42% of the corresponding native comparison and degrade co-tenant memory or I/O performance. They do not establish universal savings, amplification, or current production impact; cloud checks were intentionally limited.