What problem are you trying to solve?
Karpenter already does the right thing for selected ODCRs: it turns AvailableInstanceCount into finite, near-zero-price reserved offerings and tries them before Spot/On-Demand.
A Zonal RI has the same useful shape: exact instance type, AZ, count, reserved capacity, and zero marginal cost while unused. Today Karpenter still prices a matching launch as ordinary On-Demand, so a GPU NodeClaim can launch elsewhere while paid-for GPU capacity sits idle.
Can we discover active Zonal RIs, subtract matching running On-Demand usage, and model the remainder as finite preferred offerings whenever they satisfy the NodeClaim requirements? Once exhausted, fall back normally.
#8717 is already exploring this; #2259 tracks the broader RI/Savings Plans problem. This ask is narrower because Zonal RIs are exact and capacity-backed.
How important is this feature to you?
High for scarce GPU pools: launching another GPU while matching reserved capacity is idle wastes both guaranteed capacity and committed spend.
What problem are you trying to solve?
Karpenter already does the right thing for selected ODCRs: it turns
AvailableInstanceCountinto finite, near-zero-pricereservedofferings and tries them before Spot/On-Demand.A Zonal RI has the same useful shape: exact instance type, AZ, count, reserved capacity, and zero marginal cost while unused. Today Karpenter still prices a matching launch as ordinary On-Demand, so a GPU NodeClaim can launch elsewhere while paid-for GPU capacity sits idle.
Can we discover active Zonal RIs, subtract matching running On-Demand usage, and model the remainder as finite preferred offerings whenever they satisfy the NodeClaim requirements? Once exhausted, fall back normally.
#8717 is already exploring this; #2259 tracks the broader RI/Savings Plans problem. This ask is narrower because Zonal RIs are exact and capacity-backed.
How important is this feature to you?
High for scarce GPU pools: launching another GPU while matching reserved capacity is idle wastes both guaranteed capacity and committed spend.