KServe KEDA autoscaling documentation

KServe v0.21.0 added direct KEDA scaling support. Supports idleReplicaCount=0 for true scale-to-zero. KEDA can scale on custom metrics from Prometheus, making it possible to scale on LLM-specific metrics like KV cache utilization or queue depth rather than CPU. | DEAD-LINK VERIFIED 2026-10-04. THREE independent GET probes returned HTTP 404 with title 'Page Not Found | KServe' (soft-404: 18KB of rendered HTML, so this is a real missing page rather than a bot-block). The KServe site restructured its generative-inference docs. The substantive claims here (KEDA scale-to-zero, idleReplicaCount=0, custom-metric scaling) remain readable on the current KServe LLMISVC autoscaling page; a future auditor should re-point the URL at the reorganised path rather than treat the figure as void. NOTE ON METHOD: a single HEAD probe cannot distinguish this from a block - see the repo's rent-cerebras-api-token-pricing-negative lesson. | 2026-10-05 REPAIR: REPAIRED - URL REPOINTED TO A LIVE REPLACEMENT, exact-path move found and content-verified. KServe moved the LLMISVC autoscaling page DOWN ONE LEVEL: the dead URL ended .../llmisvc/llmisvc-autoscaling and the live one ends .../llmisvc/autoscaling/llmisvc-autoscaling. HTTP 200. CONTENT VERIFIED line by line against the claims in this record: (a) KEDA is still a supported actuator backend ('Two actuator backends: HPA ... or KEDA'), (b) custom-metric scaling on LLM-specific signals is explicit ('KV cache utilization, queue depth, and request saturation'), (c) idleReplicaCount is documented, with the KEDA-vs-HPA table row 'Idle scale-down | Not supported (min 1 replica) | Supported via idleReplicaCount'. ONE CORRECTION TO THE OLD CLAIM, recorded because the page is stricter than the old note: the page says idleReplicaCount 'allows scaling below minReplicas when no triggers are active. It must be less than minReplicas', and every worked example on the page uses idleReplicaCount: 1, NOT 0. So 'idleReplicaCount=0 for true scale-to-zero' is NOT what the current doc demonstrates - the doc's own examples floor at 1 idle replica. The idle-replica mechanism is real and still documented; the =0 claim is not substantiated by this page and is withdrawn. The record stays verified because every claim it still carries is on the page.