Scheduler

Scheduler

The Scheduler analyzes an ExecutionPlan’s topology and distributes nodes across available workers.

Rules

1. Sequential phases → single worker (avoids data transfer)
2. Parallel branches → round-robin across workers by capability
3. Differentiable connected nodes → same worker (gradient flow must be preserved)
4. Cache hits → resolved at runtime on the assigned worker (the plan itself carries no cached nodes)
5. Loop/Branch bodies → same worker as controller

DistributionPlan

The scheduler produces a DistributionPlan containing:

Field	Description
assignments	Node → worker mapping with reason
phases	Sequential/parallel groupings
data_transfers	S3 transfers needed between workers
warnings	Capacity issues, single-worker fallback

Example

Graph: [Load] → [Normalize] → [Train SVM]
                                → [Train KNN]

Scheduler with 2 workers:
  Worker A (GPU): Load → Normalize → Train SVM  (sequential, data locality)
  Worker B (CPU): Train KNN                       (parallel branch)

  Data Transfer: Normalize → Train KNN (via S3)

Usage

use somatize_compiler::{schedule, WorkerInfo};

let workers = vec![
    WorkerInfo { id: "gpu-1", gpu: true, cpu_cores: 16, ... },
    WorkerInfo { id: "cpu-1", gpu: false, cpu_cores: 64, ... },
];

let plan = schedule(&execution_plan, &workers, &differentiable_nodes);
// plan.assignments, plan.data_transfers, plan.warnings