Reservations and capacity borrowing

A nominal quota is an accounting entitlement under a particular queue policy. A time-bounded resource reservation can instead identify resources and an interval for selected users or jobs. Neither should be confused with a completed pod placement or an application-ready guarantee. Resource flavors, topology domains, and other dimensions still constrain what can run. State whether a promise concerns a quantity of compatible resources, particular nodes, or an admission priority when those resources become available.

Borrowing uses another eligible queue’s unused share under configured rules. Limits on borrowing and lending can restrict how much is exposed. The borrowed amount remains opportunistic unless the policy explicitly converts it to another entitlement. A tenant should know whether long non-checkpointable jobs are allowed to consume it and how much notice accompanies reclamation. Avoid selling the same nominal quantity as an unconditional immediate-start promise to two tenants when the physical resources can serve only one.

When the owner submits work, the policy may prioritize its new admission or evict borrowers to reclaim enough quota. Eligibility and priority rules determine which work can be interrupted. The resource does not become physically free until victims exit and cleanup completes. A borrower may need to checkpoint, transfer state, or restart later. Track intent to reclaim, eviction, resource release, quota update, and the owner’s actual useful start as separate observations; otherwise a successful accounting decision can hide a long service delay.

Kueue separates job admission and quota management from Kubernetes pod-to-node scheduling. This distinction prevents an incorrect assumption that sufficient borrowed quota automatically implies a compatible node assignment. Also verify reclaim semantics for the installed policy rather than assuming every queue can preempt every other queue. A protected workload or topology mismatch may delay reclamation or require another candidate set. The owner’s entitlement must be explained within these conditions and the availability of the underlying cluster.

Time-bounded reservations can help plan large training runs or maintenance, but backfilling around them depends on runtime estimates and enforcement. A borrower whose execution cannot end in time can conflict with the reserved interval. Define what happens on overruns, failures, or changed reservations. Keep physical inventory and accounting synchronized during node replacement and autoscaling. Test a simultaneous return of several nominal owners rather than evaluating only the easy case where one owner reclaims a small amount from idle borrowers.

The queue lab illustrates the drain component only. Reset, set arrival rate to 0 and service rate to 4, predict when the initial four-unit backlog empties, then Step twice. For a separate run, Reset and explicitly set arrival rate to 2 and service rate to 4, predict the one-second count, then Step twice. These are independent fluid runs from four units. There are no tenants, quotas, indivisible jobs, checkpoint delays, or physical reservations, so an empty backlog here cannot certify a real borrowing-revocation promise.