Skip to content

BLAST Submit — subscription-wide cluster picker

Motivation

The dashboard's ClusterCard was migrated to a subscription-wide AKS list (see 2026-05-25-aks-subscription-wide-list.md) so researchers can operate the typical multi-cluster ELB fleet (heavy / light / gpu / general, each in its own AKS-managed RG) from a single screen. The Submit page was still pinned to the workspace anchor RG (config.workloadResourceGroup), so:

  • The cluster <select> only listed clusters in the anchor RG, even when the operator wanted to dispatch a job to a different tier in a different RG.
  • The submit + preflight payloads hard-coded resource_group: workloadRg, which would have caused list_aks_clusters(cred, sub, rg) lookups to fail with "cluster not found in ''" once the picker was sub-wide.
  • The empty-state copy ("No AKS clusters in ") implied the anchor RG is the only valid scope — at odds with the Dashboard story.

User-facing change

  • Submit → Compute picker now lists every ELB-managed cluster the caller can see in the subscription (same filter as the ClusterCard).
  • <option> label now reads aks-elb-heavy — koreacentral · heavy · rg-blast-heavy (Running) — region · tier · RG joined by middots, so multi-cluster fleets are unambiguous without expanding the row.
  • Auto-selection still prefers the first Running cluster, falling back to the first available.
  • Empty-state copy is now scope-honest: No ELB-managed AKS clusters found in this subscription.
  • Preflight + submit payloads now carry the selected cluster's RG, so cross-RG dispatch works end-to-end without changing the backend contract.

API / IaC diff

Backend: no change. /api/blast/pre-flight and /api/blast/submit already accept any resource_group; they only use it to locate the cluster via list_aks_clusters(cred, sub, rg). Sending the cluster's actual RG (which the SPA now does) keeps the existing contract intact.

Frontend:

  • web/src/pages/blastSubmit/useClusterSelection.ts — query swapped from RG-scoped monitoringApi.aks(subId, workloadRg) to sub-wide monitoringApi.aks(subId); workloadRg field removed from the args interface (caller no longer passes it).
  • web/src/pages/blastSubmit/ComputeSection.tsx — <option> label enriched with tier + resource_group; option key qualified by RG to remain unique when names collide across RGs (defensive, unlikely in practice); empty-state copy reworded.
  • web/src/pages/BlastSubmit.tsx — useClusterSelection call drops the workloadRg arg; preflight payload uses selectedCluster?.resource_group || workloadRg; handleSubmit passes selectedCluster.resource_group || workloadRg as the workloadRg arg to buildSubmitRequest so the submit payload's resource_group follows the selected cluster.
  • web/src/pages/blastSubmit/taxonomyFilter.test.ts — added two regression tests (emits the selected cluster's RG in the submit payload, uses the cluster's region in the submit payload) under a new blast submit cross-RG cluster picker describe block.

Infra: no change.

Validation evidence

  • cd web && npx tsc --noEmit → clean.
  • cd web && npx eslint src/pages/BlastSubmit.tsx src/pages/blastSubmit/ComputeSection.tsx src/pages/blastSubmit/useClusterSelection.ts → 0 problems.
  • cd web && npx vitest run → 314 passed (40 files), including the two new cross-RG regression tests.
  • uv run ruff check api → all checks passed.
  • uv run pytest -q api/tests/test_smoke.py api/tests/test_blast_database_availability.py api/tests/test_monitoring_aks_subwide.py → 91 passed in 6.11s (smoke + preflight DB availability + sub-wide AKS contract).

Risk notes

  • Load-bearing contract: the picker now relies on selectedCluster.resource_group always being populated for ELB-managed clusters. _serialise_cluster derives this from the ARM resource ID (/resourceGroups/<rg>/...), so an absent value implies a malformed ID and is covered by test_subwide_returns_empty_when_resource_group_missing in api/tests/test_monitoring_aks_subwide.py. If this contract ever drifts, the picker would silently fall back to the workspace anchor RG via the || workloadRg guard — visible as a preflight "cluster not found" check on submit, not a silent misroute.
  • No backend change means existing operators on older SPA bundles continue to work unchanged — the route still accepts the anchor RG when the old client sends it (provided the cluster happens to be in that RG).
  • region was already cluster-driven (region: selectedCluster.region || region in useSubmitMutation.ts); the new RG path matches that pattern.