JournalDAY 29 / X

FIELD NOTE / X

A spot GPU hour needs a failure budget.

The complete written thought and the evidence behind it. The video edition will follow its public release.

Journal September 25, 2026 · X target October 26, 2026

A spot GPU hour needs a failure budget.

Video caption

A spot GPU hour needs a failure budget. Progress and the invoice follow different rules. Buy the completed run, including its failure path. #EricFieldNotes

Full written post / accessibility read

Consider a fictional training run on a preemptible GPU pool. The lower hourly rate buys capacity the provider may reclaim. A green reschedule event says nothing about checkpoint age, queue delay, or when the customer gets an accepted model.

The fictional job is reclaimed just before a checkpoint. It resumes from an older snapshot and repeats work. Whether both allocations are charged, excluded, or credited depends on the provider's contract. Record lost progress and net billed usage separately.

In a disposable test, interrupt one worker through an approved fault route or local mock. Measure the last valid checkpoint, replacement wait, replayed steps, final result, and metered versus net billed usage. Tie the job, scheduler, and invoice records to one run ID.

Ask for reclaim notice, restart priority, recoverable progress, credits, and p ninety-five cost per accepted run. Do this because a low spot rate is useful only when the recovery and billing terms still deliver the customer's job on budget.

#EricFieldNotes

Four-beat scene transcript

1. A spot GPU hour needs a failure budget.

Consider a fictional training run on a preemptible GPU pool. The lower hourly rate buys capacity the provider may reclaim. A green reschedule event says nothing about checkpoint age, queue delay, or when the customer gets an accepted model.

Visual: A cheap allocation can lose expensive progress.

2. The replay can consume another allocation.

The fictional job is reclaimed just before a checkpoint. It resumes from an older snapshot and repeats work. Whether both allocations are charged, excluded, or credited depends on the provider's contract. Record lost progress and net billed usage separately.

Visual: Progress and the invoice follow different rules.

3. Rehearse the recovery packet.

In a disposable test, interrupt one worker through an approved fault route or local mock. Measure the last valid checkpoint, replacement wait, replayed steps, final result, and metered versus net billed usage. Tie the job, scheduler, and invoice records to one run ID.

Visual: Detect, restore, finish, and reconcile.

4. Contract for recovered work.

Ask for reclaim notice, restart priority, recoverable progress, credits, and p ninety-five cost per accepted run. Do this because a low spot rate is useful only when the recovery and billing terms still deliver the customer's job on budget.

Visual: Buy the completed run, including its failure path.

Research and claim limits

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

More notes from the work ↗