# Bounded Kubernetes (k3s on NVIDIA DGX Spark) — Fourth Cloud Assessment

*Fourth Cloud Control Plane Assessment - Kubernetes, bounded by a workload*

**Version:** v1.0  
**Date:** July 4, 2026  
**Status:** complete  
**Evolution model:** continuous

**Source:** Scored on the Fourth Cloud methodology (v2.4) against the assembled stack a single production workload requires. Each score was earned by a validator that executed during a live, cloud-free deployment on the node. Fixed function taxonomy; gap ownership on every score below 4; DAPM per function.


## Full Authority, Fully Assembled

This assessment scores a Kubernetes control plane assembled from open-source components on a single node and bounded by one production workload, which supplies the product boundary that makes each function testable. k3s provides the orchestration kernel; the data plane (CloudNativePG with pgvector, MinIO), model serving (vLLM), local embeddings, identity (Keycloak), and ingress (Traefik) are assembled and operated by the enterprise.

The gap portfolio is dominated by Closeable gaps at the catalog and integration layers: application catalog, developer self-service, event fabric, workflow orchestration, and API management are absent and must be acquired and operated as additional open-source components. The context fabric is provided at retrieval, where the vector store and local inference are self-hosted and drilled, while the estate-scale data functions — gravity awareness, governance metadata, lineage — are absent. FC-0 substrate is Ceded to the hardware vendor. FC-2C reasoning is Structural and absent, consistent with on-prem.

Identity continuity is partial: the self-hosted Keycloak plane reaches the workload runtime but not the remaining layers, and the enterprise owns extending it.

The buyer's trade is authority for operational responsibility. Under DAPM the assembly is Retained above the substrate: every capability is open-source and swappable without rebuilding, so no vendor holds the enterprise's accumulated opinions. The cost is that the enterprise operates every layer it retains. Each provided function carries a lifecycle the enterprise owns, and each absent function is a component it must acquire, deploy, and maintain. The substrate is the single Ceded exception, bound to the hardware vendor.


## Scoping note

This row scores a Kubernetes control plane bounded by a single production workload. Kubernetes has no product boundary to assess in the abstract because it is an assembly; the workload supplies that boundary, and each function is scored against what the assembled, deployed stack provides, earned by a validator that executed during the deployment. The assessed assembly is k3s with CloudNativePG and pgvector, MinIO, Keycloak, vLLM, a local embedding model, and Traefik, on a single NVIDIA DGX Spark. Every open gap names the component the enterprise would acquire, deploy, and operate to close it.


## Identity Plane Continuity

**Score:** 2  
**Classification:** partial  
**Gap ownership:** closeable  
**Layers in plane:** fc2b  
**Layers siloed:** fc0, fc1, fc2a, fc2c, fc3, fc4


A federated identity plane is present as a self-hosted Keycloak instance running on the platform's own Postgres, issuing OIDC tokens with RS256 signing. It reaches the workload runtime: the application validates realm tokens and authenticates against them, rejecting unauthenticated requests. It does not yet span orchestration, catalog, or integration, so continuity is partial. The enterprise owns the identity provider and must extend it — additional clients and single sign-on — to reach the remaining layers.


**Buyer implication:** Can FC-0 identity control FC-2B runtime execution? Partially. The workload authenticates against the self-hosted Keycloak plane, but identity is not yet enforced across orchestration, catalog, and integration. The enterprise owns extending it, and the cost is configuration of additional clients and single sign-on across services it already runs, not a new acquisition.


## Layer-by-layer scoring

| Layer | Avg score | Status |
|---|---|---|
| FC-0 · Substrate — Physical & Virtual Substrate | 1.33 | Gap |
| FC-1 · Context — Distributed Data & Context Fabric | 1.75 | Gap |
| FC-2A · Orchestration — Infrastructure Orchestration | 2.00 | Moderate |
| FC-2B · Runtime — Execution & Runtime | 2.25 | Moderate |
| FC-2C · Reasoning — The Reasoning Plane | 0.00 | Absent |
| FC-3 · Catalog — Application Distribution and Governance | 1.25 | Gap |
| FC-4 · Integration — Integration Fabric | 1.20 | Gap |

**DAPM profile:** Retained 24 · Delegated 0 · Ceded 2


### FC-0 · Substrate — Physical & Virtual Substrate

*The physical foundation the control plane lifecycles.*


#### Hardware lifecycle management *(universal)*
**Score:** 1 · **Gap ownership:** closeable · **DAPM:** Ceded

Provided: none. k3s runs on the host operating system but does not lifecycle node hardware, firmware, or the OS. The substrate is a single vendor appliance whose accelerator and driver are vendor-maintained. The enterprise provisions and updates the node manually; a managed model requires fleet tooling (Cluster API, Metal3) the platform does not include.


#### Substrate heterogeneity *(universal)*
**Score:** 1 · **Gap ownership:** structural · **DAPM:** Ceded

Provided: none. The substrate is one aarch64 node, so there is no heterogeneity to manage until additional or dissimilar nodes exist. Adding nodes is a separate deployment, not a capability the platform closes here.


#### Substrate portability *(universal)*
**Score:** 2 · **Gap ownership:** structural · **DAPM:** Retained

Provided: the assembled workload is portable — containers, a served model, and an embedder that move to any conformant Kubernetes. The accelerator path is bound to the vendor appliance and its driver. The enterprise keeps the software's portability; the GPU binding is a constraint of the chosen hardware.


*Notes: The substrate is a single vendor appliance. The control plane does not lifecycle hardware; the accelerator and driver are vendor-maintained, and substrate authority is Ceded. Only the portability of the assembled software above it is Retained.*


### FC-1 · Context — Distributed Data & Context Fabric

*The data fabric the reasoning plane queries — full enterprise data estate.*


#### Data location and gravity awareness *(universal)*
**Score:** 1 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: none. The platform has no data-location or gravity awareness; all data is local to the single node. The enterprise owns any multi-location placement, and closing this at estate scale requires a data-fabric layer the platform does not include.


#### Governance and compliance metadata *(universal)*
**Score:** 1 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: none. No governance or compliance metadata layer is present. The enterprise must acquire and operate a data catalog and policy tooling (for example OpenMetadata) to close it.


#### Retrieval and context services *(ai-workload)*
**Score:** 3 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: a self-hosted document and vector store on the CloudNativePG operator with pgvector, serving grounded retrieval to the workload and surviving a force-kill restart with full vector recovery. The enterprise owns the operator lifecycle, the index build, and the recovery-point policy. No managed vector service is included; the capability is assembled from open-source components the enterprise runs itself.


#### Data pipeline and lineage *(universal)*
**Score:** 2 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: an ingest path that chunks, embeds, and loads the corpus. No lineage or provenance tracking is present. The enterprise owns pipeline orchestration and must acquire lineage tooling to close the gap.


*Notes: The context fabric is provided at retrieval, where a self-hosted vector store and local inference are assembled and drilled. The estate-scale data functions — location awareness, governance metadata, lineage — are absent and Closeable.*


### FC-2A · Orchestration — Infrastructure Orchestration

*Unified orchestration of the full enterprise workload portfolio.*


#### Workload universality *(universal)*
**Score:** 3 · **Gap ownership:** structural · **DAPM:** Retained

Provided: the Kubernetes API surface schedules containers of any shape across the node. VM and bare-metal workload types are not addressed by this assembly, so coverage is container-scoped. The enterprise holds full control of workload placement.


#### Resource lifecycle automation *(universal)*
**Score:** 3 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: operator-driven resource lifecycle for the data tenant — the CloudNativePG operator handles provisioning, failover, backup, and replica creation without operator action, demonstrated by an automatic standby provision on scale-out. The enterprise owns operator selection and upgrades; lifecycle automation for other tenants requires deploying their operators.


#### Policy and quota enforcement *(universal)*
**Score:** 1 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: Kubernetes-native RBAC and resource limits. No admission-control policy engine and no GPU quota are present. The enterprise must acquire and operate a policy engine (Kyverno, OPA Gatekeeper) and, for GPU quota, extend the open-source device plugin.


#### Substrate lifecycle integration *(universal)*
**Score:** 1 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: none. The control plane does not integrate substrate lifecycle. The gap is high-effort to close on a single node with limited return, and the enterprise owns it.


#### Accelerator and GPU management *(ai-workload)*
**Score:** 2 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: GPU access to scheduled pods through runtime-class injection. GPU accounting and quota are not provided: the driver cannot report unified memory, so the standard device plugin schedules no GPU capacity. The enterprise can close accounting by extending the open-source device plugin; the platform ships no supported path.


*Notes: Orchestration is container-scoped. Operator-driven resource lifecycle is provided for the data tenant; policy, quota, and substrate integration are absent and must be assembled. Accelerator management provides access without accounting.*


### FC-2B · Runtime — Execution & Runtime

*The execution plane where workloads actually run.*


#### Runtime universality *(universal)*
**Score:** 3 · **Gap ownership:** structural · **DAPM:** Retained

Provided: a container runtime for any workload the enterprise packages, over the Kubernetes API. Non-container runtimes are out of scope for this assembly. The enterprise holds runtime control.


#### Persona abstraction at execution *(universal)*
**Score:** 1 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: none. There is no persona or tenant abstraction at the execution layer. The enterprise must build or acquire it.


#### Execution lifecycle and observability *(universal)*
**Score:** 2 · **Gap ownership:** mixed · **DAPM:** Retained

Provided: pod lifecycle and self-healing — a force-killed pod is recreated without operator action. Observability is not included: no metrics, logging, or tracing stack is deployed. The enterprise must acquire and operate an observability stack (Prometheus, Grafana, Loki) to close the second half of the function.


#### AI inference and agent execution *(ai-workload)*
**Score:** 3 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: model serving on the cluster via vLLM, delivering the workload's grounded, cited generation end to end over the ingress with no external dependency. The enterprise owns the model, the memory budget, and serving uptime. No managed inference is included; the capability is self-hosted.


*Notes: Runtime is container-scoped and self-healing, with model serving provided and drilled. Observability and persona abstraction are absent and Closeable.*


### FC-2C · Reasoning — The Reasoning Plane

*Autonomous, policy-driven placement deriving from live data governance metadata.*


#### Autonomous placement reasoning *(universal)*
**Score:** 0 · **Gap ownership:** structural · **DAPM:** Retained

Provided: none. No policy-driven placement reasoning derives from live metadata, and no such reasoning plane exists in this assembly. Placement decisions rest with the enterprise and its operators. The gap is structural: the reasoning plane is unavailable on-prem.


*Notes: No reasoning plane is present. Placement is operator-driven, not derived from live metadata. Structural and absent, consistent with on-prem.*


### FC-3 · Catalog — Application Distribution and Governance

*The governed surface through which enterprise applications are published, versioned, discovered, and consumed.*


#### Application catalog and distribution *(universal)*
**Score:** 1 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: none. Deployments are applied as manifests directly. The enterprise must acquire and operate a GitOps or catalog layer (Argo CD, a Helm-based catalog) to close it.


#### Application lifecycle governance *(universal)*
**Score:** 2 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: operator-level day-2 governance for the data tenant (backup, restore, and replica lifecycle). No application-wide governance layer spans the estate. The enterprise owns broader lifecycle governance and must assemble it.


#### Developer experience and self-service *(universal)*
**Score:** 1 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: none. There is no self-service portal or golden-path tooling. The enterprise must acquire and operate a developer portal (Backstage) to close it.


#### AI application and agent distribution *(ai-workload)*
**Score:** 1 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: none. There is no packaging or distribution mechanism for AI applications or agents. The enterprise owns it.


*Notes: No application catalog, developer self-service, or AI-application distribution is provided. Application lifecycle governance exists only at the operator level for the data tenant. The layer is Closeable in full.*


### FC-4 · Integration — Integration Fabric

*Event bus, API management, workflow orchestration, and system connectors.*


#### Event fabric and messaging *(universal)*
**Score:** 1 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: none. There is no managed event or messaging fabric. The enterprise must acquire and operate a broker (NATS, Kafka) if the workload requires one.


#### API management and gateway *(universal)*
**Score:** 2 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: ingress and TLS termination via Traefik, routing external traffic to the workload over HTTPS. API management is not included: no rate limiting, authentication, or lifecycle at the edge. The enterprise must acquire an API management layer (Kong, an equivalent gateway) to close it.


#### Workflow and process orchestration *(universal)*
**Score:** 1 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: none. There is no workflow or process orchestration. The enterprise must acquire and operate a workflow engine (Argo Workflows, Temporal).


#### SaaS and enterprise system integration *(universal)*
**Score:** 1 · **Gap ownership:** opinion · **DAPM:** Retained

Provided: none, by design. The deployment carries no external SaaS or enterprise-system connectors; external integration was deliberately excluded. The runtime can host connectors the enterprise applies if integration is later required.


#### AI-native integration *(ai-workload)*
**Score:** 1 · **Gap ownership:** closeable · **DAPM:** Retained

Provided: the workload integrates its own retrieval and model internally. No platform-level AI-native integration fabric is present. The enterprise owns cross-application AI integration.


*Notes: Ingress and TLS are provided as a gateway. API management, event fabric, workflow orchestration, and external integration are absent; external integration is excluded by design. The layer is Closeable, some of it by choice.*