Everything multi-cell is proven by the two-apiserver test harness, not by hardware, because the lab has one GPU. Specifically unvalidated against real devices:
- capacity aggregation across cells, and the heterogeneous case (
Homogeneous=false, where the pool-wide compute aggregate is deliberately omitted)
cell.spreadPolicy: Spread actually spreading launcher pods across hosts- automatic scale-down picking the right victim — membership must remove the cells the autoscaler named, not the highest index, or it destroys work on cell 2 while cell 0 sits empty
- index reuse after a cell is removed
A second GPU converts this, #6 and part of #5 into validated behaviour in an afternoon.
Everything multi-cell is proven by the two-apiserver test harness, not by hardware, because the lab has one GPU. Specifically unvalidated against real devices:
Homogeneous=false, where the pool-wide compute aggregate is deliberately omitted)cell.spreadPolicy: Spreadactually spreading launcher pods across hostsA second GPU converts this, #6 and part of #5 into validated behaviour in an afternoon.