Put four promises in the compute contract.

Video caption

Put four promises in the compute contract.

Borrowing improves utilization until two contracts collide.

A policy table and a controlled contention run.

My rule: Treat sharing rules as part of the product.

#EricFieldNotes

Full written post / accessibility read

A neocloud quote should say when a job can start, which connectivity class it receives, whether it can be interrupted and how accepted progress is billed. A GPU-hour rate without those terms leaves the scheduler making product decisions in the dark.

Reserved Tenant A is idle, so Burst Tenant B borrows its allocation. A returns. The operator must reclaim a valid placement, B may need to checkpoint, and the invoice must explain the lost or resumed work. A named queue alone does not decide liability.

For each service class, specify reservation window, topology, borrowing eligibility, preemption notice, checkpoint responsibility and restart billing. Rehearse contention and failure in a test pool. Keep the scheduler trace, application progress and invoice as one evidence packet.

Compare providers on kept start promises, topology-valid placements and cost per accepted result under contention. Do this because the cheap GPU-hour that disappears at the wrong time is not the capacity the buyer thought they purchased.

#EricFieldNotes

Four-beat scene transcript

1. Put four promises in the compute contract.

A neocloud quote should say when a job can start, which connectivity class it receives, whether it can be interrupted and how accepted progress is billed. A GPU-hour rate without those terms leaves the scheduler making product decisions in the dark.

Visual: Start, topology, preemption and completed-run cost.

2. Shared capacity exposes the omissions.

Reserved Tenant A is idle, so Burst Tenant B borrows its allocation. A returns. The operator must reclaim a valid placement, B may need to checkpoint, and the invoice must explain the lost or resumed work. A named queue alone does not decide liability.

Visual: Borrowing improves utilization until two contracts collide.

3. Make the contract replayable.

For each service class, specify reservation window, topology, borrowing eligibility, preemption notice, checkpoint responsibility and restart billing. Rehearse contention and failure in a test pool. Keep the scheduler trace, application progress and invoice as one evidence packet.

Visual: A policy table and a controlled contention run.

4. Price the delivered run.

Compare providers on kept start promises, topology-valid placements and cost per accepted result under contention. Do this because the cheap GPU-hour that disappears at the wrong time is not the capacity the buyer thought they purchased.

Visual: Treat sharing rules as part of the product.

Research and claim limits

• Kueue overview (S124)
• Slurm scheduling configuration (S125)

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.