FIELD NOTE / TIKTOK
A burst tenant borrows spare GPUs.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved TikTok edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
A burst tenant borrows spare GPUs.
Video caption
A burst tenant borrows spare GPUs. A hard kill may protect A but waste B's run. The service class must price interruption honestly. #EricFieldNotes
Full written post / accessibility read
Tenant B borrows idle capacity. Good utilization. Now Tenant A, who paid for a start promise, submits early. Does B get notice? Can B checkpoint? Does A get a topology-valid slot on time? The queue name does not answer those questions.
If B's job has no reliable checkpoint, evicting it may throw away hours. If we wait for B to finish, A may miss its deadline. Both outcomes are policy costs. The correct answer depends on contracts, workload recoverability and explicit priority.
Use a disposable queue with the real job shapes. Let B borrow, then activate A's reservation. Capture the admission decision, notice, checkpoint success, actual A start, B resume and invoice. Repeat with B's checkpoint failing.
Sell B an interruptible class only if it can tolerate the tested recovery behavior, and sell A a deadline supported by eligible capacity. Do this because utilization gains are not free when another tenant owns the clock.
#EricFieldNotes
Four-beat scene transcript
1. A burst tenant borrows spare GPUs.
Tenant B borrows idle capacity. Good utilization. Now Tenant A, who paid for a start promise, submits early. Does B get notice? Can B checkpoint? Does A get a topology-valid slot on time? The queue name does not answer those questions.
Visual: Then the reserved tenant's deadline arrives.
2. A naive preemption can destroy work.
If B's job has no reliable checkpoint, evicting it may throw away hours. If we wait for B to finish, A may miss its deadline. Both outcomes are policy costs. The correct answer depends on contracts, workload recoverability and explicit priority.
Visual: A hard kill may protect A but waste B's run.
3. Rehearse the deadline collision.
Use a disposable queue with the real job shapes. Let B borrow, then activate A's reservation. Capture the admission decision, notice, checkpoint success, actual A start, B resume and invoice. Repeat with B's checkpoint failing.
Visual: Trace quota, topology, preemption and billed progress.
4. Promise only what you can reclaim.
Sell B an interruptible class only if it can tolerate the tested recovery behavior, and sell A a deadline supported by eligible capacity. Do this because utilization gains are not free when another tenant owns the clock.
Visual: The service class must price interruption honestly.
Research and claim limits
- Kueue overview (S124)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.