A housing authority colo cage wants dual spines so maintenance does not halt overnight training. Which evaluation outcome supports that design?

Throughput from bonded uplinks sustaining long-running training jobs during maintenance fails to account for long-running training jobs that cannot be paused or gracefully rescheduled mid-epoch, since higher uplink bandwidth does not keep a training collective alive when its spine is taken down. Dual spines provide an alternate path while one spine is upgraded or replaced, which matters for multi-hour or multi-day civic training runs that cannot pause for every maintenance window. Continuity under planned change is as important as continuity under failure for colo AI cages.

Throughput from bonded uplinks sustaining long-running training jobs during maintenance is wrong because bonded uplinks move more bits but cannot keep a multi-GPU collective running through a spine maintenance window, so training stalls or aborts rather than surviving on a peer path. Cabling from single-homed leaves keeping long-running training jobs during maintenance fails because single-homed leaves remove the alternate spine path, so there is no peer path to prevent a full stop when the single spine is upgraded or fails. Checkpointing to object storage sustaining long-running training jobs through maintenance does not sustain the job through maintenance, it only resumes training after a full stop, which is exactly what overnight training that cannot tolerate interruptions must avoid.

Exam caveat: Redundant spines are evaluated as an operational continuity property for long-running AI jobs—not as optional cabling luxury. Operational check: prove that draining one spine leaves GPU training paths intact on the peer, schedule spine maintenance against that runbook, and refuse single-spine colo designs for overnight housing-authority training that cannot tolerate full stops.