| Access profile |
The administrator-managed record for a subject (UserAccess): role, assigned services and projects, compute and storage grants, and whether the subject is disabled. The RBAC middleware loads it on every request. |
| Access request |
A request from a signed-in but unprovisioned user for services, reviewed by an administrator. |
| Agent DNS name |
A collision-safe, hash-suffixed DNS label derived from an agent's immutable ID; the name of its Deployment and Service on Kubernetes. |
| Audit event |
An immutable record of a mutation: action, resource, resource ID, actor, metadata, time. Written in the same transaction as the change. |
| Break-glass account |
A local account enabled next to single sign-on (KIONGA_LOCAL_LOGIN=true) for access when the identity provider is unavailable. |
| Building block |
arc42 term for a component in the static structure; see section 5. |
| Catalog |
The live list of connections, models, features, agents and tools, derived from stored resources rather than fixed data. |
| Connection |
A registered endpoint of an external or bundled system, with a health check run from inside the gateway. |
| Control plane |
The Go services that own desired state and orchestration: the gateway, and on Kubernetes the operator and integration worker. |
| Data plane |
The stores that hold durable state and carry events (PostgreSQL, Redis, Kafka, RustFS, MLflow), plus the feature gateway and storage proxy. |
| Desired state |
What a user asked for, as persisted by the gateway. It is not evidence that an engine has done the work. |
| Digest |
The SSE payload from /api/v1/events: a per-caller summary of visible state. The console reloads a view only when it changes. |
| Editorial membership |
A per-person role in the blog workspace (author, editor, editorial_admin), separate from platform roles. |
| Function DAG |
A pipeline whose steps call project-owned OpenFaaS functions instead of containers. |
| Grant |
An administrator's allocation to a user: services, projects, CPU, memory, GPU, storage, and run or function quotas. |
| IAM policy |
A versioned set of allow and deny statements over actions and kionga: resources, attached to users or groups. |
| In-stack hostname |
A Compose service name such as objectstore or mlflow, reachable only from inside the Compose network. |
| Integration worker |
The Kubernetes process that consumes lifecycle commands from Kafka and creates or updates Kionga custom resources. |
| Internal service token |
The shared bearer token (MLAIOPS_INTERNAL_TOKEN) that in-platform services use to report to the gateway as the service role. |
| Kionga custom resource |
One of KiongaAgent, KiongaPipelineRun, KiongaTool, KiongaConnection, KiongaModelPromotion, KiongaWorkspace in API group mlaiops.io/v1alpha1. |
| Lease |
A row-locked document that lets exactly one gateway replica run a singleton loop, such as the scheduler or the editorial publisher. |
| Lifecycle command |
An outbox event on a mlaiops.<kind>.commands topic that describes a change the Kubernetes path must reconcile. |
| Outbox |
The outbox_events table. Rows are written in the mutation's transaction and published to Kafka by a background worker. |
| Personal API key |
A user-created bearer credential with the nxs_ prefix, scoped to a subset of the creator's services and projects. |
| Preflight |
The server-side check before a manual run: parameters, image pins and allowlist, quotas, concurrency, and resources. |
| Principal |
The resolved caller of a request: subject, roles, groups, services, projects, and the credential type. |
| Project template |
A versioned starter (production ML, distributed training, production agent, full-stack AI, blank) that fixes framework, accelerator, profile and scaffold command. |
| Ready-set scheduling |
Running every pipeline node whose own dependencies have succeeded, up to a parallelism limit. |
| Resource profile |
A named size for compute and storage that a project or workspace requests and a grant must cover. |
| Role baseline |
A built-in managed policy that reproduces a role's permissions; evaluated together with attached policies. |
| Scale path |
The Kubernetes deployment with the operator, CRDs and per-user workloads, as opposed to single-host Compose. |
| Serving image |
The container image used for a deployed model; the model's serving_image, or SERVE_IMAGE as a fallback. |
| Service role |
The machine role (service) for in-platform reporting; it may read and post to reporting paths only. |
| Sidecar (workspace) |
The loopback helper inside the Jupyter and IDE containers that runs structured scaffold jobs for the gateway. |
| Tenant |
The isolation unit for stored data, one per gateway process (MLAIOPS_TENANT), enforced with PostgreSQL row-level security. |
| Trace proxy |
The OpenAI-compatible HTTP proxy that forwards LLM calls and publishes one event per call. |
| Trusted Compose profile |
The single-operator local or single-host stack in which the Docker socket is mounted into engine-control services. |
| Two DNS worlds |
Services reach each other by Compose service name inside the stack and by localhost ports from the host; health checks run inside the gateway. |
| Workspace |
A user's development environment: Jupyter workbench and optional IDE sharing one persistent volume. |