Coming soon Compute Forwards Learn more

Inventory Operator

The Inventory Operator continuously discovers and monitors hardware resources across your Kubernetes cluster, making them available for provider bidding decisions.

Purpose

The Inventory Operator:

  • Discovers cluster resources (CPU, GPU, memory, storage)
  • Monitors resource availability in real-time
  • Tracks resource utilization and allocation
  • Publishes inventory data to the provider service
  • Updates Kubernetes node labels with capabilities

Architecture

+---------------------------------------------+
| Inventory Operator |
| |
| +--------------------------------------+ |
| | Cluster Nodes Manager | |
| | - Watch Kubernetes nodes | |
| | - Deploy discovery pods | |
| | - Collect hardware info | |
| +--------------+-----------------------+ |
| | |
| v |
| +--------------------------------------+ |
| | Node Discovery Pods | |
| | - Run on each node | |
| | - Detect CPUs, GPUs, storage | |
| | - Read hardware capabilities | |
| +--------------+-----------------------+ |
| | |
| v |
| +--------------------------------------+ |
| | Storage Queriers | |
| | - Ceph integration | |
| | - Rancher Longhorn integration | |
| | - Storage class detection | |
| +--------------+-----------------------+ |
| | |
| v |
| +--------------------------------------+ |
| | Cluster State Aggregator | |
| | - Combine node + storage data | |
| | - Publish to event bus | |
| | - Respond to queries | |
| +--------------------------------------+ |
+---------------------------------------------+
|
v
+-------------------+
| Provider Service |
| (Bid Engine) |
+-------------------+

Discovery Process

1. Node Discovery

When the operator starts, it:

  1. Watches Kubernetes Nodes

    watch.Interface for v1.Node resources
    • Monitors node additions/removals
    • Detects node capacity changes
    • Tracks node status updates
  2. Deploys Discovery Pods

    • Creates a discovery pod on each node
    • Runs with privileged access for hardware detection
    • Uses same image as operator for consistency
  3. Collects Hardware Information

    • CPU cores and architecture
    • Memory capacity
    • GPUs (NVIDIA)
    • Storage devices (NVMe, SSD, HDD)
    • Network interfaces

2. Resource Tracking

The operator tracks:

CPU Resources

cpu:
quantity:
allocatable: 64000 # millicores
allocated: 32000 # millicores currently used
info:
- model: "AMD EPYC 7763"
vcores: 64

GPU Resources

gpu:
quantity:
allocatable: 8
allocated: 2
info:
- vendor: nvidia
name: rtx4090
modelid: "2684"
interface: pcie
memory_size: 24Gi

Memory Resources

memory:
quantity:
allocatable: 256Gi
allocated: 128Gi

Storage Resources

storage:
- class: beta3 # Storage class name (NVMe default)
size: 5000Gi # Total capacity
provisioner: ceph.rook.io

3. GPU Feature Discovery

The operator integrates with the NVIDIA Device Plugin and other GPU management tools:

Detection Process:

  1. Query PCI devices for GPUs
  2. Read GPU vendor and product IDs
  3. Match against provider-configs database
  4. Extract GPU capabilities (memory, CUDA version, features)
  5. Publish GPU info to inventory

GPU Information Collected:

  • Vendor ID (e.g., 10de for NVIDIA)
  • Product ID (e.g., 2684 for RTX 4090)
  • Model name (user-friendly)
  • Memory size
  • Interface type (PCIe, SXM)

4. Storage Discovery

The operator supports multiple storage backends:

Rook-Ceph Integration

func NewCeph(ctx context.Context) (QuerierStorage, error)

Discovers:

  • Ceph storage classes
  • Available storage capacity
  • Provisioner type
  • Storage performance class

Rancher Longhorn Integration

func NewRancher(ctx context.Context) (QuerierStorage, error)

Discovers:

  • Longhorn volumes
  • Storage pools
  • Replica counts
  • Available capacity

Real-Time Updates

The operator continuously monitors for changes:

Node Changes

case watch.Modified:
if nodeAllocatableChanged(knode, obj) {
updateNodeInfo(obj, &node)
signalLabels()
}

Triggers:

  • Node capacity changes (scale-up/down)
  • Node labels modified
  • Node conditions changed (Ready, MemoryPressure, etc.)

Pod Changes

case watch.Added, watch.Modified:
if isPodAllocated(obj.Status) {
addPodAllocatedResources(&node, obj)
}
case watch.Deleted:
subPodAllocatedResources(&node, &pod)

Tracks:

  • New pod deployments (subtract from available)
  • Pod deletions (add back to available)
  • Pod resource requests (CPU, memory, GPU)

Event Publishing

The operator publishes inventory updates to the event bus:

Topics

  • inventory.nodes - Node hardware capabilities
  • inventory.storage - Storage availability
  • inventory.cluster - Aggregated cluster state

Retained Events

bus.Pub(state, []string{topicInventoryCluster}, pubsub.WithRetain())

Events are retained so new subscribers immediately receive current state.

Kubernetes Node Labels

The operator adds labels to nodes based on discovered hardware:

Example Labels

akash.network/capabilities.gpu.vendor.nvidia: "true"
akash.network/capabilities.gpu.model.rtx4090: "true"
akash.network/capabilities.storage.class.beta3: "true"

Purpose:

  • Enable node selector constraints for scheduling
  • Support GPU model matching in deployments
  • Allow storage class requirements

Integration with Bid Engine

The Bid Engine queries the inventory to make bidding decisions:

inventory, err := cluster.Inventory(ctx)
if err != nil {
return err
}
// Check if resources are available
canBid := inventory.Has(requiredResources)

Flow:

  1. Order arrives with resource requirements
  2. Bid Engine queries inventory
  3. Inventory Operator returns available resources
  4. Bid Engine compares required vs. available
  5. Bid submitted if resources sufficient