Azure Resource Graph: emit resource-type-specific fields (VM SKU, managed disks, DB tier) for real-SDK discovery/cost E2E

[NitinKumar004]

Summary

The Azure Resource Graph emulator (server/azure/resourcegraph + resourcediscovery) returns a generic cross-cloud inventory rather than Azure-type-specific ARG rows. A real azure-sdk-for-go consumer that reads resource-type-specific fields (VM SKU, disk SKU/size, standalone managed disks, DB tier, etc.) cannot drive an end-to-end discovery/cost test against cloudemu today, because those fields aren't projected into the ARG response.

This blocks a real use case: pointing a production Azure discoverer + cost engine at cloudemu to verify SKU/tier/size-sensitive logic with no real account — which is otherwise cloudemu's strongest value proposition.

Evidence (real-SDK round trip)

Driving the real armresourcegraph client against httptest.NewTLSServer(azureserver.New(...)) with a VM and a CreateVolume disk seeded:

cloudP := cloudemu.NewAzure()
cloudP.VirtualMachines.RunInstances(ctx, computedriver.InstanceConfig{InstanceType: "Standard_D2s_v3"}, 1)
cloudP.VirtualMachines.CreateVolume(ctx, computedriver.VolumeConfig{Size: 4, VolumeType: "Premium_LRS"})
// ...azureserver.New(Drivers{VirtualMachines, Disks, ResourceDiscovery, SubscriptionID})...

Querying Resources | where type in~ (...) via the real SDK returns:

Expected (real Azure ARG)	cloudemu today
VM row with sku.name = Standard_D2s_v3 (or properties.hardwareProfile.vmSize)	VM row present, but SKU field empty — consumer reads InstanceType = ""
microsoft.compute/disks row with sku.name + properties.diskSizeGB + managedBy	No disk row emitted at all (CreateVolume volume not surfaced in ARG)

Requested fidelity (fields real Azure ARG returns; needed for discovery/cost E2E)

1. Compute VMs — project into the ARG row:

• sku.name or properties.hardwareProfile.vmSize = the VM size (e.g. Standard_D2s_v3)
• properties.storageProfile.osDisk.osType (Linux/Windows)
• properties.priority = Spot (interruptible pricing)
• properties.licenseType (Windows_Server / RHEL_BYOS — Azure Hybrid Benefit)

2. Managed Disks (highest impact) — surface CreateVolume/CreateDisk volumes as first-class ARG resources of type microsoft.compute/disks with:

• sku.name = Premium_LRS / StandardSSD_LRS / Standard_LRS / *_ZRS / PremiumV2_LRS / UltraSSD_LRS
• properties.diskSizeGB = provisioned size
• top-level managedBy = owning VM id (disk→VM parent link)
• properties.diskIOPSReadWrite / properties.diskMBpsReadWrite (v2/Ultra)

3. PostgreSQL / MySQL Flexible Server — microsoft.dbforpostgresql/flexibleservers (+ mysql) with:

• sku.name = the compute SKU (e.g. Standard_D2ds_v5)
• properties.highAvailability.mode (ZoneRedundant / SameZone / Disabled)

4. Event Hubs — microsoft.eventhub/namespaces with sku.name/sku.tier (Basic / Standard / Premium).

Nice-to-have (secondary): SQL Managed Instance storage size, sqldatabase zoneRedundant, ML compute-instance vmSize + powerState.

Why it matters

With these fields, a real Azure discoverer + cost calculator can run a full offline E2E against cloudemu (discover → price → assert cost) for SKU/tier/size-sensitive logic. Today that verification stops at the SDK/protocol layer — cloudemu proves the wire round-trip and KQL, but not the resource-shape fidelity that downstream cost/rightsizing tools depend on.

Repro

Minimal armresourcegraph round-trip against azureserver.New with ResourceDiscovery wired (recipe already in server/azure/resourcegraph/sdk_test.go); seed a VM + volume as above and assert sku.name / diskSizeGB / a microsoft.compute/disks row.

[NitinKumar004]

Root cause pinpointed (source read, v1.7.3):

resourcediscovery.Resource carries only:

type Resource struct { Provider, Service, Type, ID, ARN, Region string; Tags map[string]string; CreatedAt time.Time }

There is no SKU / Size / type-specific Properties field, so the ARG projection can only ever emit id/name/type/region/tags — VM SKU, diskSizeGB, managedBy, DB tier, etc. have nowhere to live. Additionally, walkCompute (resourcediscovery/walkers.go) only calls Compute.DescribeInstances — volumes/disks are never walked, so managed disks can't appear in ARG at all.

So this needs (1) a way for the inventory Resource to carry type-specific attributes (e.g. a Attributes map[string]any / Properties + SKU), (2) walkers that populate them from the drivers (instance type, volume type+size+managedBy), and (3) a volume walker so disks surface as microsoft.compute/disks. Then the ARG projection can render the fields listed above.

[NitinKumar004]

Full spec — everything needed to make cloudemu drive a real Azure discovery + cost E2E

Consolidating the exact gaps found by source-read (v1.7.3) so this is actionable end-to-end.

1. Current ARG row is fixed and minimal

server/azure/resourcegraph/handler.go:215 builds every row as exactly:

map[string]any{
    "id":             r.ARN,                       // (see shape issue A)
    "name":           r.ID,                        // (see shape issue A)
    "type":           portableToAzureType(...),
    "location":       r.Region,
    "resourceGroup":  "default",                   // (see shape issue B)
    "subscriptionId": extractSubscription(r.ARN),
    "tags":           tagsOrEmpty(r.Tags),
}

No sku, properties, managedBy, kind, or zones. And server/azure/resourcegraph/kql.go silently ignores project / summarize / join, so even a client that projects those columns gets the fixed shape above.

2. Root cause (model, not projection)

resourcediscovery.Resource = {Provider, Service, Type, ID, ARN, Region, Tags, CreatedAt} — no SKU/Size/Properties, and walkCompute only calls Compute.DescribeInstances (no volume walk), so managed disks never enter the inventory.

3. Required changes

1. Inventory model — give resourcediscovery.Resource type-specific attributes, e.g. SKU string + Properties map[string]any (+ ManagedBy string, Kind string, Zones []string).
2. Walkers — populate them from the drivers (instance.InstanceType, volume Type/Size/ManagedBy, DB tier, …), and add a volume/disk walker so disks surface as their own resources.
3. ARG row builder — render sku ({"name": …}), properties, managedBy, kind, zones when present.
4. KQL — at minimum keep project a no-op (fine), but the fields above must exist in the full row.

4. Per-type fields real Azure ARG returns (ranked by impact)

Azure type	Fields needed
microsoft.compute/virtualmachines	sku.name or properties.hardwareProfile.vmSize; properties.storageProfile.osDisk.osType; properties.priority (Spot); properties.licenseType (Windows_Server/RHEL_BYOS); zones
microsoft.compute/disks (not emitted today)	sku.name (Premium_LRS/StandardSSD_LRS/Standard_LRS/*_ZRS/PremiumV2_LRS/UltraSSD_LRS); properties.diskSizeGB; top-level managedBy; properties.diskIOPSReadWrite + properties.diskMBpsReadWrite; properties.diskState
microsoft.compute/virtualmachinescalesets	sku.name (VM size) + sku.capacity (instance count); OS/priority/license as VMs
microsoft.dbforpostgresql/flexibleservers (+ mysql)	sku.name (compute SKU, e.g. Standard_D2ds_v5); properties.highAvailability.mode; properties.storage.storageSizeGB
microsoft.eventhub/namespaces	sku.name/sku.tier (Basic/Standard/Premium); properties.isAutoInflateEnabled
microsoft.sql/managedinstances	sku; properties.vCores; properties.storageSizeInGB
microsoft.sql/servers/databases	sku.tier; properties.maxSizeBytes; properties.zoneRedundant
microsoft.machinelearningservices/…/computeinstance	properties.vmSize; properties.state (Running/Stopped)
AI Foundry managed-compute deployment	model / accelerator VM SKU

5. Additional shape issues (would trip a real Azure discoverer)

• A. id/name inversion — id = r.ARN, name = r.ID. Real Azure ARG id is the ARM resource ID (/subscriptions/…/providers/Microsoft.Compute/virtualMachines/<name>) and name is the short resource name. Consumers that parse the id (resource group, provider, name) get the wrong shape.
• B. resourceGroup hardcoded "default" — real rows carry the actual RG (and consumers derive it from the id).
• C. project ignored — acceptable, but worth documenting so users don't expect column selection.

6. Acceptance criteria

Seed a VM (InstanceType: Standard_D2s_v3) + a CreateVolume(Size: 4, VolumeType: Premium_LRS) attached to it, then via the real armresourcegraph SDK run
Resources | where type in~ ('microsoft.compute/virtualmachines','microsoft.compute/disks') | project id,name,type,location,resourceGroup,managedBy,properties,sku,tags,zones
and assert:

• VM row: sku.name == "Standard_D2s_v3" (or properties.hardwareProfile.vmSize)
• disk row present with type == "microsoft.compute/disks", sku.name == "Premium_LRS", properties.diskSizeGB == 4, managedBy == the VM id
• id is an ARM-shaped resource id; name is the short name.

7. Why (motivating use case)

A production Azure discoverer + cost engine (10 cost fixes: disk tiers, VM SKU/Spot/AHB, flexible-DB managed rate, Event Hubs tier, v2/Ultra capacity, …) can then run a fully offline discover → price → assert-cost E2E against cloudemu — no account, no network. Today that E2E stops at the SDK/protocol layer (round-trip + KQL verified) but can't reach the resource-shape fidelity those SKU/tier/size-sensitive consumers depend on.