Skip to content

Concepts

The spec remains the authoritative contract.

What InferenceIdentityBinding Is

InferenceIdentityBinding is the namespaced intent object owned by kleym.

It tells the controller:

  • which InferenceObjective should receive an identity
  • whether identity should be scoped to the whole pool or to one objective
  • which safety selectors must always be present
  • how to narrow a per-objective identity down to one container when needed

kleym then resolves the objective's poolRef, derives workload selection from the referenced InferencePool, and writes one or more ClusterSPIFFEID resources.

Identity Modes

kleym has two identity boundaries.

Mode Meaning Typical use
PoolOnly One identity for the serving pool. The whole pool is one workload boundary and model-level separation is not needed.
PerObjective One identity per InferenceObjective. Multiple objectives share a pool but still need distinct identities.

PoolOnly answers "which serving pool is this pod part of?"

PerObjective answers "which model endpoint is this container serving?"

If you do not set mode, the controller defaults to PerObjective.

Why Container Discrimination Exists

PerObjective only makes sense if kleym can prove that one objective identity lands on one workload slice.

When several objectives share the same InferencePool, the pod selector alone is not enough because every objective may resolve to the same pods. The container discriminator adds a narrower selector so the identity applies to the intended serving container instead of the whole pod.

Current discriminator types:

  • ContainerName, which is preferred because container names are explicit and stable within a pod template
  • ContainerImage, which is a fallback when container names are not a useful discriminator

Without this extra boundary, distinct objective identities could collapse onto the same workload selection.

Safety Selectors

Safety selectors are the controller's proof that the rendered identity stays inside the intended tenant boundary.

Every rendered identity must include:

  • the binding namespace selector
  • the workload service account selector
  • selectors derived from the referenced pool
  • the container discriminator when the mode is PerObjective

kleym refuses to reconcile when it cannot prove those constraints. That is why selector handling is intentionally narrower than raw Kubernetes label selection.

What kleym Does Not Do

kleym is an identity registration compiler.

It does not:

  • deploy inference workloads
  • create InferencePool or InferenceObjective objects
  • route traffic
  • configure gateways, meshes, or policy engines
  • issue certificates itself

SPIRE and SPIRE Controller Manager remain responsible for issuing identities. kleym only determines which identities should exist and which workloads they should target.

See Also

  • Read architecture for the end-to-end control flow.
  • Read examples for concrete manifests.
  • Read spec if you need the full contract.