Skip to main content
Autopilot’s recommendations are usually right, but sometimes you know something the metrics don’t — a workload needs a minimum reservation for a cold cache, or must never exceed a hard budget. Resource Overrides are per-workload annotations that constrain what Autopilot may recommend. They’re the operator’s guardrails, and they always win over the algorithm.

​
Floors and ceilings

Set annotations on a Deployment or StatefulSet. A floor clamps the recommendation upward (never go below); a ceiling clamps it downward (never go above). Behavior: floor clamps up, ceiling clamps down, and ceiling is applied after floor. A ceiling also overrides the CPU downscale-delay hold — an operator ceiling is an explicit directive. The API validates that floor ≤ ceiling for each resource pair.

​
Setting them

Three ways, all equivalent:
  • Workload detail — the edit form has Min/Max fields per resource row, with a “set to recommended” helper and an over-recommendation warning.
  • Resource Overrides page — a namespace-grouped find-and-clean tool: groups sorted by locked vCPU, multi-select with a bulk-remove bar, and “wants Xm” stale-floor badges when a floor now sits above what the algorithm wants.
  • API / bulk — PATCH /api/v1/overrides/{ns}/{name}/floor for one workload, or POST /api/v1/overrides/bulk to apply set/remove across up to 500 workloads at once.
A system-managed eviction memory floor is separate from these operator floors: when a workload is evicted under node memory pressure, Autopilot ratchets a memory floor from the observed usage (decaying after 7 days) so it isn’t repeatedly evicted. It’s shown as an amber “eviction floor” badge and can’t be lowered below by a recommendation.

​
Alert suppression

Related per-workload annotations let you silence a noisy health alert on a specific workload — useful when a tenant won’t fix a known issue and you don’t want it dragging cluster health: A suppressed alert stays visible in the health detail panel (and on the Resource Overrides page’s suppressions list) but no longer affects the cluster’s overall health status. Set and remove them from the health panel’s per-workload “Suppress” button.