What drift looks like
Over time a workload’s live resources can diverge from its template for legitimate reasons (an in-place resize, an OOM bump, an operator floor) or accidental ones (a stale annotation, a half-applied change). The audit makes each visible:- template vs recommendation — the workload is over- or under-provisioned relative to what usage says it needs.
- template vs pod — the running pod no longer matches what the Deployment declares.
- QoS drift — the template is Guaranteed but the live pod is Burstable (or vice-versa), which affects eviction priority.
truly-guaranteed, drift) isolate the
workloads where QoS class matters.
Remediation
Each row offers targeted, idempotent actions — and any of them can be applied in bulk to up to 500 workloads at once (concurrency-limited, with per-item error reporting):Every action is blocked on the controller’s own namespace (
ops-ai) — Autopilot never resets
or restarts itself. “Reset to deploy” also clears LimitExceptions so a stale exception’s reject
can’t later revert the workload back toward its pre-reset values.When to use it
- After a migration or a period of manual tuning, to find workloads that drifted and reconcile them to a known-good baseline.
- To undo Autopilot on a specific workload — “reset to deploy” is the clean off-switch for one workload’s sizing.
- To spot QoS drift before it causes a surprise eviction ordering.