Glossary

​

1Password

Secret source for vault-scoped Fleet Manager runtime/operator credentials and service-account tokens. Best source: 1Password CLI documentation

​

Allocation

Fleet Manager grouping that associates devices with a tenant or intended customer use. Best source: Tenants and allocations

​

Argo CD

GitOps deployment UI/operator used to sync deployed Fleet Manager images and configuration. Often written ArgoCD. Best source: Argo CD documentation

​

Azure Data Explorer

See Kusto.

​

BMC

Baseboard Management Controller; out-of-band management chip used for power, console, and low-level server control. Best source: DMTF Redfish standard

​

Burn-in

Stress and validation testing before hardware is trusted for customer workloads. Best source: Build a new burn-in test

​

Cluster

A grouped compute environment, usually scheduled by SLURM or Kubernetes, where workloads run across multiple nodes. Best source: nscale compute clusters

​

Control Center

Operations console / lifecycle automation layer in the Fleet Operations product story. Best source: nscale Fleet Operations

​

Customer custody

Whether a node is under customer control, nscale control, quarantine, reclaim, or related operational custody. Best source: Device state machines

​

DCGM

NVIDIA Data Center GPU Manager; GPU health, diagnostics, and telemetry tooling. Best source: NVIDIA DCGM documentation

​

DCIM

Data Center Infrastructure Management; structured records of physical assets, locations, racks, power, and connectivity. Best source: NetBox documentation

​

Device

See Node.

​

DHCP

Protocol that assigns network configuration to a host during boot or provisioning. Best source: RFC 2131

​

Fleet

The population of hardware Fleet Manager is responsible for observing and moving through lifecycle states. Best source: Overview

​

fleetctl

Typer CLI entry point for infra, enrollment, BMC, DHCP, reports, hardware tests, and deployment standup. Best source: src/fleet_manager/cli/main.py

​

Good-put

Commercially useful GPU output after failures, idle time, and remediation overhead are removed. Best source: Fleet Manager: Product Definition

​

GPU

Graphics Processing Unit; the accelerator customers rent to train, tune, or serve AI models. Best source: NVIDIA glossary

​

Handover

Moving validated infrastructure into a customer-ready or customer-controlled state. Best source: Workflows reference

​

Harbor

nscale registry mirror for Chainguard base images; local Docker builds require login. Best source: README

​

Helm

Kubernetes packaging tool used by the hub and satellite charts. Best source: Helm documentation

​

Host

See Node.

​

HPL

High Performance Linpack; benchmark used to stress CPU/GPU compute performance. Best source: Build a new burn-in test

​

HTTP Boot

Boot patterns used by Ironic/BMC workflows to boot provisioning or cleaning images. HTTP boot fetches the image over the network; virtual media has the BMC present it as an attached device. Best source: OpenStack Ironic documentation

​

HyperDeploy

Operating model for repeatable GPU infrastructure delivery at scale. Best source: HyperDeploy lifecycle

​

IB

See InfiniBand.

​

In-band

Network path through the host operating system or service network rather than the BMC management path. Best source: Deploy Fleet Manager at a new site

​

InfiniBand

Low-latency, high-throughput fabric commonly used for large-scale GPU training clusters. Often abbreviated IB. Best source: NVIDIA InfiniBand

​

IPMI

Older management protocol used with BMCs; still encountered in server operations and validation tooling. Best source: Intel IPMI specification

​

IPv6

Internet Protocol version used in newer provisioning designs, including IPv6-only Ironic/Kea blueprints. Best source: RFC 8200

​

Ironic

OpenStack bare-metal service that enrolls, controls, cleans, and provisions physical nodes. Best source: OpenStack Ironic documentation

​

Ironic Python Agent

Agent ramdisk (abbreviated IPA) used by Ironic clean/inspection steps; Fleet Manager docs include the image update runbook. Best source: IPA image update runbook

​

Kusto

Analytics database, also known as Azure Data Explorer. The customer-facing “Matrix” inventory databases live in it, written by the matrix connector. Best source: Azure Data Explorer documentation

​

NATS

Messaging system used by Radar ingestion and connector/event work. Best source: NATS documentation

​

NCCL

NVIDIA collective communication library used by distributed GPU workloads and network/GPU validation tests. Best source: NCCL documentation

​

NetBox

Source-of-truth DCIM/IPAM system for sites, racks, devices, roles, types, interfaces, and cabling. Best source: NetBox documentation

​

Node

A physical machine in the fleet. Code usually says Device; operators often say node or host. Best source: src/fleet_manager/api/core/models.py

​

Nova

OpenStack compute service; Fleet Manager checks it when nodes are deployed or attached to instances. Best source: OpenStack Nova documentation

​

NVLink

High-bandwidth GPU-to-GPU interconnect and switching fabric used in dense NVIDIA systems. Best source: NVIDIA NVLink

​

NVSwitch

See NVLink.

​

OOB

Out-of-band management network, usually for BMC access independent of host OS networking. Best source: Deploy Fleet Manager at a new site

​

Placement

OpenStack service that tracks resource providers, inventories, and traits such as customer readiness. Best source: OpenStack Placement documentation

​

pre-commit

Local and CI hook runner for Ruff, formatting, ty, OpenAPI generation, Pyright, Bandit, and Prettier. Best source: pre-commit documentation

​

Radar API

Customer-facing visibility layer for fleet availability, repair, telemetry, and maintenance signals. Usually just called Radar. Best source: Relationship: Radar

​

RDMA

Remote Direct Memory Access; high-speed network capability critical for distributed GPU training. Best source: InfiniBand Trade Association

​

Reclaim

Taking a node back from customer use so Fleet Manager can quarantine, validate, remediate, or re-release it. Best source: Workflows reference

​

Redfish

Modern REST-style DMTF standard for managing servers through the BMC. Best source: DMTF Redfish standard

​

SLA

Service Level Agreement; the service promise made to a customer. Best source: Google SRE: service level objectives

​

SLO

Service Level Objective; the internal reliability target a team works to. Fleet Manager helps by reducing fault and repair ambiguity. Best source: Google SRE: service level objectives

​

SLURM

HPC workload manager (also written Slurm) used here for burn-in tests and managed training-style workloads. Best source: SLURM overview

​

Teleport

Infrastructure access path used to reach app hosts and dev environments. Best source: Developer guide overview

​

Temporal

Durable workflow engine used for long-running hardware operations that need retries, heartbeats, and recovery. Best source: Temporal documentation

​

Tenant

Customer or internal owner context for devices and allocations. Best source: Tenants and allocations

​

ToR

Top-of-rack switch connecting servers in a rack to the data-center fabric. Best source: NetBox data requirements

​

uv

Python dependency and execution tool used for all Python commands in this repo. Best source: uv documentation

​

Vault

Secret system used in chart/runtime integrations such as Radar ingestion product approle secrets. Best source: HashiCorp Vault documentation

​

Virtual media

See HTTP Boot.

​

VLAN

Layer-2 segmentation tag; nodes may be moved between testing, customer, quarantine, or provisioning networks. Best source: NetBox data requirements

​

VRF

Routing table isolation; used when provisioning networks must be separated from customer or management paths. Best source: Workflows reference