Glossary¶
Short definitions of terms used throughout these docs, in alphabetical order. For the full explanation of the main ideas, see Key concepts.
- Ad-hoc inference
- Sending a task straight to a single block, rather than through a vDAG workflow. Ad-hoc inference goes through an inference server.
- AIGr.id (AIGrid)
- An open, plural and polycentric network of AI components and compute, run as a digital commons. See What is AIGr.id?
- AIOS / OpenOS.AI
- The open-source, distributed AI operating system that powers AIGr.id. "AIOSv1" refers to this release.
- App layout
- The logical design of an AI application, expressed as a vDAG, with no reference to hardware. Compare scale layout. See App layout and scale layout.
- Assets registry
- A distributed store for model weights, policy code archives, images, videos and other files that blocks can reference. See Assets registry.
- Assignment policy
- A policy that picks the most suitable running block for a node of a vDAG when the vDAG is created. See vDAG spec.
- Auto-scaler
- A block service that watches instance metrics and runs the auto-scaling policy to add or remove instances.
- Batcher
- An Instance SDK helper that groups a fixed number of packets before processing. See Instance SDK.
- Block
- A running deployment of a component that serves AI or general computation. It comes with its own executor (load balancer), auto-scaler, health checker and metrics. A block lives in one cluster. See Block.
- Block ID (
blockId) - The unique identifier of a block. You use it to send tasks to the block and to manage it.
- Cluster
- A Kubernetes cluster that has joined a network and contributes nodes to it. See Cluster.
- Cluster allocator policy
- A block policy that shortlists and selects the cluster where a block will be placed. See Block spec.
- Cluster Controller
- The service that manages one cluster: it places blocks, monitors nodes and enforces the cluster's pre-check policies. See Cluster Controller.
- Cluster Controller Gateway
- The network-level service that manages all clusters, schedules blocks and enforces the network's pre-check policies. See Cluster Controller Gateway.
- Component
- A packaged, versioned piece of software (container images plus metadata) from which blocks are created. See Component.
- Component URI (
componentURI) - The unique identifier of a component, generated from its name, version and release tag. Block specifications refer to it.
- Executor
- The gateway service of a block. It receives tasks and load-balances them across instances using the load-balancer policy.
- Feasibility score
- A number from 0 to 1, returned by a resource allocation policy in dry-run mode, that says how well a cluster can host a block.
- FrameDB
- A storage system for large payloads. Instead of sending a large file with a task, you store it in FrameDB and send a pointer (
frame_ptr). - Governance policy
- See pre-check policy.
- Grid
- Another name for a network.
- Health checker
- A block (or vDAG controller) service that periodically checks instance health and applies a health-check policy.
- Inference server
- A gateway that accepts tasks for blocks over gRPC. Anyone can deploy one and list it in the inference server registry. See Inference server.
- Init container
- A container that AIOS runs when a block is created, and again when it is removed, to install and uninstall a third-party system. See Third-party systems.
- Instance
- One running copy (pod) of a block's container. A block scales by adding or removing instances.
- Instance SDK
- The Python SDK used to write the code that runs inside a block's instances. See Instance SDK.
- Management cluster
- The Kubernetes cluster that hosts a network's management services and registries. There is one per network.
- Management command
- An instruction sent to a running block or cluster service, for example to change a setting or query a policy. See Management commands.
- Master DAG
- The combined, network-wide graph formed by overlaying every running vDAG.
- Model splitting
- Cutting a large model into parts that run as separate blocks, connected by a vDAG. See Model splitting.
- Muxer
- An Instance SDK helper that merges logically related packets before processing.
- Network
- A group of clusters governed by one management cluster and one set of rules. See Network.
- Node
- A single machine in a cluster. See Node.
- Parser
- The management service that accepts JSON specifications and creates components, blocks, clusters and vDAGs. See How the Parser works.
- Policy
- A piece of Python code that the system loads and runs at a decision point to customize behavior. See Policy and Policies.
- Policy executor
- A service that runs policies on behalf of other services (Type 2), or runs them as jobs on its cluster (Type 3).
- Policy rule URI
- The unique identifier of a registered policy, for example
aiosv1.policies.cluster-controller-gateway.add-node:v0.0.12-stable. - Pre-check policy
- A governance policy that runs before an action, such as adding a node or creating a block, and can allow, reject or modify it. Networks and clusters each have their own.
- Pre- and post-processing policies
- Policies that transform data before and after a vDAG node runs its inference.
- Quality checker policy
- A vDAG controller policy that periodically samples inputs and outputs to audit quality.
- Quota checker policy
- A vDAG controller policy that limits how much a session, or everyone, may use a vDAG.
- Resource allocation policy
- A block policy that picks nodes and GPUs for the block's instances, and scores clusters in dry-run mode.
- Resource Allocator
- The management service that runs resource allocation policies. See Resource Allocator.
- Scale layout
- How an app layout is physically deployed: which clusters, nodes and GPUs, and how many replicas. Compare app layout.
- Sidecar
- An extra container deployed with a block that receives the instances' outputs and does extra work, such as saving results or sending alerts.
- Spec (specification)
- A JSON document that describes something to create, such as a component, block, cluster or vDAG. See Specifications.
- Spec store
- A registry of reusable specifications. See Spec store.
- Stability checker policy
- A policy that judges whether a cluster, or a block's instances, are stable and healthy, based on collected data.
- Template store
- A registry of custom parser templates that define alternative specification formats. See Template store.
- Third-party system
- An external system, such as a vLLM server, that a block delegates computation to. See Third-party system.
- vDAG (virtual Directed Acyclic Graph)
- A workflow defined over running blocks, possibly across clusters. See vDAG.
- vDAG controller
- The gateway for a vDAG. It accepts tasks and enforces quotas, health checks and quality checks. See vDAG controller.
- vDAG URI
- The unique identifier of a vDAG, built from its name and version, for example
pose-estimator:v0.0.1-stable.