Preview node drains with a read-only, honest drain plan #1584

Problem

Draining a node cordons it and evicts eligible pods. Radar currently shows one generic confirmation and immediately mutates the cluster. It does not show which pods it expects to evict or skip, why, or that it sends deleteEmptyDirData: true without an explicit acknowledgement.

Operators need a read-only estimate before committing a destructive operation. It must be honest: pods, PDB budgets, and permissions can change between the preview and the real drain.

Goal

Replace the generic confirmation with a server-backed drain plan. Opening it must not cordon, evict, or otherwise mutate the cluster. The operator reviews the expected effects and explicit options, then starts the existing drain flow, which re-lists and evaluates live state again.

Scope

1. Extract the current drain eligibility decisions into a pure, reusable classifier in pkg/k8score.

   • Return a per-pod outcome and operator-facing reason.
   • Cover terminal pods, mirror/static pods, DaemonSet pods, unmanaged pods, and pods that use emptyDir.
   • Treat a pod as managed only when it has a controller owner reference.

2. Add a read-only drain-plan endpoint beside the existing drain endpoint.

   • Use the requesting user’s cluster client and RBAC identity.
   • Accept the same relevant options: force, deleteEmptyDirData, and grace period.
   • Return evaluated options, generation time, summary counts, and per-pod evict, skip, or may-block outcomes with reasons.
   • A matching PDB with currently exhausted disruptionsAllowed is useful evidence, but never a guarantee: the Eviction API and live state are authoritative.

3. Replace the generic Drain Node confirmation with a focused plan dialog.

   • Fetch and display the plan before enabling the destructive action.
   • Explicitly label the result as an estimate that is re-evaluated on execution.
   • Start with deleteEmptyDirData disabled. Require an explicit acknowledgement before affected pods can be evicted.
   • Keep force-unmanaged-pods explicit and explain it.

4. Make post-drain results legible: actual evictions, skips, and individual failures.

Important constraints

• Previewing must make no cluster mutation.
• Do not represent missing or uncomputed information as zero.
• Do not claim that a preview proves a drain will succeed.
• This is not a generic preflight framework and does not change the MCP drain tool.

Starting points

• pkg/k8score/node_ops.go
• internal/server/node_handlers.go
• web/src/api/client.ts
• packages/k8s-ui/src/components/shared/ResourceActionsBar.tsx

There are no focused drain tests today. Please add unit coverage for the classifier and preview handler, plus frontend coverage that viewing the plan does not mutate the cluster and that emptyDir acknowledgement is required.

Acceptance criteria

• Opening Drain Node shows a read-only plan; neither cordon nor eviction happens until final confirmation.
• The plan makes expected eviction, skip, and currently PDB-constrained outcomes legible with reasons.
• The UI calls the plan an estimate and execution re-evaluates live state.
• emptyDir data is never silently included; the operator explicitly opts in.
• Managed-owner detection requires a controller owner reference.
• Execution reports actual evictions and individual skips/failures.