Strata

A multi-tenant control plane and resource manager for managed database and inference workloads.

Status

Early development. Core resource modeling, node allocation, first-fit scheduling, and tenant admission control are implemented.

Build

cmake -S . -B build
cmake --build build
ctest --test-dir build --output-on-failure

Current Progress

Strata currently has the first version of its core resource-management model.

Implemented so far:

• ResourceVector

  • CPU, memory, and GPU resource representation
  • Validation preventing negative resource quantities
  • Resource addition and subtraction
  • Fit checks between requested and available capacity

• Workload

  • Workload ID
  • Tenant ownership
  • Requested resource vector

• Node

  • Total and available capacity tracking
  • Resource fit checks
  • Resource allocation with capacity updates

• Scheduler

  • Maintains a collection of nodes
  • First-fit workload placement
  • Allocates resources on the first node capable of running a workload
  • Returns no placement when cluster capacity is insufficient

• Tenant

  • Tenant ID
  • Resource quota
  • Current resource usage
  • Admission checks against tenant quota

• GoogleTest coverage for resource arithmetic, workload state, node allocation, scheduling behavior, and tenant admission control

Example Scheduling Flow

strata::Workload workload(
    "job-1",
    "team-a",
    strata::ResourceVector(4, 8, 1));

std::vector<strata::Node> nodes;
nodes.emplace_back("node-1", strata::ResourceVector(2, 4, 1));
nodes.emplace_back("node-2", strata::ResourceVector(8, 16, 2));

strata::Scheduler scheduler(nodes);

strata::Node* node = scheduler.schedule(workload);

if (node != nullptr) {
    // workload was allocated to the selected node
}

Roadmap

Next:

• Update tenant usage when workloads are admitted
• Release resources when workloads complete
• Tenant accounting and quota enforcement
• Scheduling policies beyond first-fit
• Multi-tenant fairness and DRF-inspired admission
• Placement scoring and fragmentation awareness
• Large-scale workload simulator
• Autoscaling and capacity policies
• gRPC policy-engine interface
• Kubernetes control-plane integration
• Managed PostgreSQL and inference-provider integrations