Identify the client-visible effect of earliest eStargz cache-write failures

Determine whether the earliest ENOSPC cache-write failures produced by stargz-snapshotter v0.18.2 under cache exhaustion make model files unreadable by clients immediately, or whether the failed writes instead trigger successful refetches.

Background

The paper observes a period during which stargz-snapshotter logs real ENOSPC failures while Kubernetes readiness, health checks, and prediction requests remain successful. The authors explicitly distinguish this early observability gap from the later pressure-matrix experiments, which directly establish widespread client-visible file failures.

Whether the first cache-write failures themselves already cause client-visible unreadability remains unresolved because a failed cache write may degrade to a refetch. This uncertainty concerns the onset mechanism, not the existence of corruption under stronger cache pressure.

References

Second, whether those first \lieeio{} cache-write failures made any file client-unreadable at all is not established by this experiment (a failed cache write can degrade to a refetch); client-visible corruption under the same mechanism is established directly by the pressure matrix below, where 67--94\% of files fail.

— The Lazy Pod That Lies: Deferred Cost and Failure Semantics of Lazy Container Image Pulling for Model Serving on Kubernetes  (2608.19412 - Kliukovkin, 19 Aug 2026) in Section 6.2, “The lying-pod timeline”

We characterize this failure phenomenologically; the internal path by which a failed cache write becomes a persistent read error in v0.18.2 was not root-caused (planned configuration micro-experiments were abandoned for environment-stability reasons; see the artifact).

— The Lazy Pod That Lies: Deferred Cost and Failure Semantics of Lazy Container Image Pulling for Model Serving on Kubernetes  (2608.19412 - Kliukovkin, 19 Aug 2026) in Section 6.1, “A threat-free failure model”

We attempted and could not cleanly reproduce the cache-exhaustion failure on SOCI within our time budget---persistent local caching resisted isolation. We therefore claim the failure mode for eStargz only and explicitly do not claim SOCI is immune; its proactive background fetcher gives it the same unbounded-write ingredient.

— The Lazy Pod That Lies: Deferred Cost and Failure Semantics of Lazy Container Image Pulling for Model Serving on Kubernetes  (2608.19412 - Kliukovkin, 19 Aug 2026) in Section 8, “Threats to Validity,” subsection “SOCI failure parity”