CONCEPT Cited by 1 source
Fair sharing¶
Fair sharing is a resource allocation policy where compute capacity is distributed among tenants proportionally to their configured weights, ensuring no single tenant monopolizes shared resources while idle capacity is utilized efficiently.
Definition¶
In a multi-tenant batch compute system, fair sharing governs how resources are distributed when demand exceeds a single tenant's reservation or when shared pools are contested. Weight-based fair sharing assigns each tenant a proportional share of the available resources.
Admission-only vs. preemption-based fair sharing¶
Netflix's CMB system applied fair sharing only at admission time — once a job was admitted, it ran to completion regardless of shifts in fair-share demand. This meant a tenant that submitted a burst of low-priority work early could monopolize capacity even when higher-priority tenants arrived later.
Kueue introduces preemption-based fair sharing: the system continuously enforces fair-share allocations by preempting workloads when demand patterns shift. Idle reserved capacity is lent to borrowing tenants and reclaimed on demand via reclaimWithinCohort: Any.
(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 — comparison of admission-only (CMB) vs. preemption-based (Kueue) fair sharing at Netflix.