CONCEPT Cited by 1 source
Reserved vs. shared capacity¶
Reserved capacity guarantees exclusive resources to a tenant; shared capacity provides a common pool any tenant can burst into when demand exceeds its reservation.
Definition¶
Most multi-tenant compute platforms offer two resource pools:
- Reserved — pre-allocated resources dedicated to a specific tenant or subtree. Provides throughput guarantees and isolation. Tradeoff: unused reserved capacity is wasted unless the system supports lending.
- Shared — a global surplus pool accessible to all tenants. Provides elasticity and higher utilization. Tradeoff: no throughput guarantee under contention.
At Netflix¶
In CMB, reserved capacity on internal tenants was fair-shared across the subtree; on leaf tenants it was exclusively partitioned. Shared capacity was fair-shared at admission only — once admitted, jobs ran to completion.
With Kueue, the semantics evolved: reserved resources are lent to other Cohort members when idle and reclaimed via preemption when needed (reclaimWithinCohort: Any). This eliminates the waste of idle reservations while preserving guarantees.
(Source: sources/2026-06-22-netflix-how-netflix-simplified-batch-compute-with-kueue)
Seen in¶
- sources/2026-06-22-netflix-how-netflix-simplified-batch-compute-with-kueue — Netflix's dual-capacity model and how Kueue's lending/reclaim semantics improve upon admission-only sharing.