Skip to content

2026-05-27 — Multi-cluster Cluster Plane hardening

Motivation

User review of the dashboard "Cluster Plane" surfaced 25 issues, most notably that the per-cluster API p95 / Errors 15m cells were actually rendering the dashboard backend's own /api/blast ring-buffer metric. The same value was therefore being repeated on every cluster row and was driving misleading per-cluster "degraded" status lines. Several related ergonomics gaps (no fleet KPI, no sort, no Portal link, weak Delete confirmation, mobile users could not Start/Stop a cluster) were addressed in the same pass.

User-facing change

  • Correctness
  • /api/blast request metrics are no longer rendered inside every cluster row. They moved to a single card-header strip labelled Dashboard p95 N · M 5xx / 15m with a tooltip making clear that the value is the dashboard backend's own metric, not the K8s API server / per-cluster signal.
  • useClusterHealth no longer mixes that global metric into the per-cluster verdict or its status line — the dot tone and the substatus line now always agree (a green dot can no longer be paired with a yellow substatus text).

  • The Pulse meta grid hides CPU peak, Mem peak and DBs when the cluster is stopped / transitioning / provisioning, instead of rendering empty — cells that look like "0%" / "0 visible".

  • Multi-cluster ergonomics

  • Cluster rows are now sorted issues-first (failed → provisioning → stopped → not-ready → healthy, alphabetical within each bucket).
  • A fleet KPI strip above the row list reads N clusters · X running · Y stopped · … plus the dashboard p95 summary.

  • Clusters missing the elb-tier ARM tag now show a dashed untagged chip instead of nothing, so the tagging gap is visible.

  • Actions / safety

  • Every row carries an Open in Portal button that deep-links to the AKS resource overview in the Azure Portal.
  • Start / Stop / Delete are now reachable on mobile (the dashboard-hide-mobile class was removed).
  • The Delete dialog requires the user to type the cluster name exactly before the Delete button enables (extends ConfirmDialog with a generic typeToConfirm prop).

  • Start/Stop/Delete completion + error messages are now mirrored into a visually-hidden aria-live="polite" region so screen reader users hear them, and the visible banners also carry role="status".

  • Copy / cleanup

  • The third pulse stat is now labelled Load (max CPU/Mem peak) instead of the previously-undefined Pressure.
  • The "warming · ready · failed" Databases legend is now only rendered when there is actually at least one warm chip, so the "No warmed databases yet" empty state stays focused.

API / IaC diff summary

  • No backend, OpenAPI or Bicep change.
  • Frontend only:
  • web/src/components/cards/ClusterPulse/useClusterHealth.ts — drop apiP95/apiErrors from inputs, drop the soft-degraded substatus branch.
  • web/src/components/cards/ClusterPulse/usePulseSignals.ts — drop the per-cluster monitoringApi.requestMetrics query; reuse topQuery.isError as the per-cluster metricsDegraded signal.
  • web/src/components/cards/ClusterPulse/PulseMetaGrid.tsx — drop API p95 / Errors 15m cells; new operational prop gates CPU peak / Mem peak / DBs cells.
  • web/src/components/cards/ClusterPulse/ClusterPulse.tsx — pass subscriptionId + operational through, drop the lifted props.
  • web/src/components/cards/ClusterPulse/PulseActions.tsx — add Open in Portal link, drop dashboard-hide-mobile.
  • web/src/components/cards/ClusterPulse/PulseRowSummary.tsx — rename Pressure → Load; render untagged tier badge when the tag is missing.
  • web/src/components/cards/ClusterCard/ClusterCard.tsx — sort clusters issues-first, fleet KPI strip, lifted dashboard-metrics fetch, ARIA live region, name-typed Delete.
  • web/src/components/ConfirmDialog.tsx — new optional typeToConfirm / typeToConfirmLabel props with disabled-until-match guard.
  • web/src/components/ClusterItem/DatabaseChipStrip.tsx — hide the warming/ready/failed legend when warmChips.length === 0.

Validation

  • cd web && npm test -- --run → 363 / 363 passing.
  • cd web && npx tsc --noEmit → clean.
  • cd web && npm run build → clean (warns only on pre-existing chunk-size limit).
  • cd web && npx eslint src → clean.

Deferred (intentionally not in this PR)

  • Pre-Start cost / time estimate in the row (#16): needs a backend cost-projection endpoint that does not exist yet.
  • Provision queueing instead of strict single-in-flight (#14): architectural change to useClusterProvisioning; out of scope for a UX hardening pass.
  • Per-row "last refreshed" timestamp (#20): card-footer already shows freshness; per-row would clutter the line without a clear win until rows can stale independently.
  • Delta-only Region / K8s row (#13): cosmetic, debatable in a fleet where regions or K8s versions are heterogeneous.