Skip to content

NRI plugin failure-mode hardening (fail-closed / fail-open) #434

Description

@ArangoGutierrez

Part of the NRI node-wide injection epic (v0.3.0 flagship), building on PR #433.

An NRI plugin sits in the container-creation path for the whole node, with two failure modes that cut opposite ways:

  • Fail-closed: a plugin that returns an error on a CreateContainer request fails that container's creation — blocks every workload on the node.
  • Fail-open: a plugin that times out gets unregistered by containerd, and subsequent containers silently lose injection — tests keep running against an unmocked environment with no signal.

Scope:

  • Liveness/readiness probes for the nvml-mock-nri pod so a wedged plugin restarts and re-registers.
  • Document both failure modes and containerd plugin_request_timeout guidance in the chart README.
  • Decide and document the intended posture per mode (e.g. fail-closed acceptable on dedicated test clusters; fail-open must be detectable — see the e2e self-check issue).

Acceptance: killing/wedging the plugin mid-run is observable and recovers; behavior is documented.

Metadata

Metadata

Labels

kind/featureFeature request or enhancement.priority/p1P1: works but not as documented; small impact; workaround exists. Max fix SLA: 6 months.

Type

No type

Fields

No fields configured for issues without a type.

Projects

No projects

Relationships

None yet

Development

No branches or pull requests

Issue actions