Skip to content

2026-05-15 — AKS provision UI: system / user pool split + grouped SKU dropdown

Motivation

The sibling repo (elastic-blast-azure commit a2d2f0a — feat(azure): add default system VM size and configuration for AKS node pools) standardised the cluster on a two-pool layout:

  • systempool (mode=System, taint CriticalAddonsOnly=true:NoSchedule, default VM size Standard_D2s_v3) — hosts CoreDNS / metrics-server / csi-azuredisk-node etc.
  • blastpool (mode=User, taint workload=blast:NoSchedule, label workload=blast) — runs every ElasticBLAST workload pod.

The dashboard's Celery task and /api/aks/provision route were already wired for this split (see api/tasks/azure.py::provision_aks and api/services/aks_skus.py), but the wizard UI had no controls for the system pool — so every cluster was created with the static defaults and the user could not deviate without hand-editing JSON in the network tab.

Separately, the SKU dropdown was a single flat <select> with 31 options. Picking the right SKU for memory-vs-storage workloads required knowing the naming convention; there was no visual grouping by series.

User-facing change

  • Create AKS Cluster modal now has two clearly labelled sections:
  • Workload pool · blastpool — Node SKU + Node Count (existing).
  • System pool · systempool · CriticalAddonsOnly — System VM size
    • System node count (1–3, capped to match AKS minimums).
  • The cost summary line now breaks down blastpool and systempool separately and sums them, e.g. ~$20.26/hr · blastpool: 10 × Standard_E32s_v5 (~$20.16/hr) · systempool: 1 × Standard_D2s_v3 (~$0.10/hr).
  • Both SKU dropdowns are now grouped by series with <optgroup> separators. Order: System pool · HPC (HB/HC) · Memory-optimised E v5 · Memory + NVMe E bs v5 · Memory E v3 · General D v3 · Storage L v3 · Storage L as v3. Each option also shows the hourly USD price inline: Standard_E32s_v5 (32 vCPUs, 256 GB · $2.02/hr).
  • The blast pool dropdown only lists SKUs flagged role=blast|both; the system pool dropdown only lists role=system|both. The user can no longer accidentally pick a 96-vCPU HPC SKU for the system pool.
  • The provisioning banner mirrors the same split: blastpool 10 × Standard_E32s_v5 / systempool 1 × Standard_D2s_v3 · Est. 5–10 min.

API / IaC diff summary

Backend: * api/services/aks_skus.py — SkuListResponse now carries group_labels: dict[str,str] and group_order: list[str]. sku_list_response() builds them from the catalog so the SPA never hardcodes the per-series labels. * api/tests/test_aks_skus.py — test_allowed_skus_match_sibling_azure_hpc_machines now asserts that the blast-only subset of SKU_BY_NAME matches sibling AZURE_HPC_MACHINES, and that every system-only SKU is allowed and contains DEFAULT_SYSTEM_SKU. The route test additionally verifies default_system_sku, every SKU's role ∈ {system,blast,both}, and that group_order covers every used group exactly once.

Frontend: * web/src/api/aks.ts — AksSkuListResponse gains optional group_labels + group_order. * web/src/hooks/useAksSkus.ts — FALLBACK_AKS_SKUS now contains both a system-pool entry (Standard_D2s_v3) and the blast default with the required role/group fields. New exports DEFAULT_AKS_SYSTEM_SKU, groupAksSkus(skus, pool, order, labels), defaultSystemSku, groupLabels, groupOrder. The fallback group labels/order mirror SKU_GROUP_LABELS / SKU_GROUP_ORDER in the backend so the SPA still renders correctly when the API response is cached without the new fields. formatAksSkuOption now appends the hourly USD price. * web/src/components/cards/ClusterCard.tsx — adds systemVmSize / systemNodeCount state, a second SKU <select> for the system pool, an effect that adopts the backend's default_system_sku once it loads, and forwards both new fields to aksApi.provision. Both <select>s render <optgroup>s sourced from groupAksSkus().

No infra changes — the Celery task and /api/aks/provision route already accept system_vm_size / system_node_count; this PR just wires them through the UI.

Validation evidence

  • uv run pytest -q api/tests — 123 passed.
  • cd web && npm run build — tsc -b && vite build clean (671 kB JS gzipped 183 kB; pre-existing chunk-size warning unchanged).
  • Manual smoke (next): user will exercise the + Add Cluster modal in the dashboard and create a real cluster with the new system / user pool fields; the resulting AKS will show two agent pools (systempool D2s_v3 ×1 + blastpool E32s_v5 ×N) per the sibling constants.py::ELB_AZURE_*_POOL_NAME mirror.

Cross-repo consistency note

If sibling repo bumps ELB_DFLT_AZURE_SYSTEM_VM_SIZE, ELB_AZURE_SYSTEM_POOL_NAME, ELB_AZURE_BLAST_POOL_NAME, or the workload=blast label/taint values, update both of these in lockstep:

  • api/services/aks_skus.py (DEFAULT_SYSTEM_SKU, SKU_GROUP_LABELS),
  • api/tasks/azure.py::provision_aks (the SYSTEM_POOL_NAME / BLAST_POOL_NAME / BLAST_TAINT / SYSTEM_TAINT constants).