Virtual machines: images, sizes, quotas and the first boot
Most of what this course runs on Azure is managed: AKS nodes, App Service, Container Apps. Under all of them are virtual machines, and every platform team still owns a few directly - a jump host, a build agent, a legacy service nobody has containerised, a database vendor that only supports VMs. An administrator is expected to create one correctly the first time: the right image, a size that exists in the region and fits the quota, keys instead of passwords, and no SSH port open to the whole internet by accident.
Need to know: a VM needs an image (publisher:offer:sku:version, or an alias like Ubuntu2404), a size (Standard_D2s_v5: family D, 2 vCPUs, s = premium storage capable, version 5), a region and optionally an availability zone. A size can be unavailable in a region or zone (az vm list-skus shows the restrictions), and every subscription has vCPU quotas per region and per family (az vm list-usage). az vm create is a deployment that also makes a VNet, an NSG, a public IP, a NIC and an OS disk if you do not give it existing ones. Stopped VMs still cost compute; deallocated ones do not. Resizing to a size the current hardware cluster does not offer needs the VM deallocated first.
Images
$ az vm image list --query "[].{alias:urnAlias, urn:urn}" -o table
Alias Urn
----------------------- ------------------------------------------------------------------------------
CentOS85Gen2 OpenLogic:CentOS:8_5-gen2:latest
Debian11 Debian:debian-11:11-backports-gen2:latest
Debian12 Debian:debian-12:12-gen2:latest
FlatcarLinuxFreeGen2 kinvolk:flatcar-container-linux-free:stable-gen2:latest
OpenSuseLeap154Gen2 SUSE:openSUSE-leap-15-4:gen2:latest
RHELRaw8LVMGen2 RedHat:RHEL:8-lvm-gen2:latest
SuseSles15SP5 SUSE:sles-15-sp5:gen2:latest
Ubuntu2204 Canonical:0001-com-ubuntu-server-jammy:22_04-lts-gen2:latest
Ubuntu2404 Canonical:ubuntu-24_04-lts:server:latest
Ubuntu2404Pro Canonical:ubuntu-24_04-lts:ubuntu-pro:latest
Win2022Datacenter MicrosoftWindowsServer:WindowsServer:2022-datacenter-g2:latest
Win2022AzureEditionCore MicrosoftWindowsServer:WindowsServer:2022-datacenter-azure-edition-core:latest
Win2019Datacenter MicrosoftWindowsServer:WindowsServer:2019-Datacenter:latest
Win2016Datacenter MicrosoftWindowsServer:WindowsServer:2016-Datacenter:latest
Win2012R2Datacenter MicrosoftWindowsServer:WindowsServer:2012-R2-Datacenter:latest
az vm image list shows the popular images with their URN - publisher:offer:sku:version - and the short alias the CLI accepts instead (--image Ubuntu2404). latest as the version means "the newest published"; pin a version for builds that must be reproducible. Two details the exam likes:
- Generation 2 images (
-gen2, Hyper-V Gen2, UEFI boot) are the default for new sizes; some older sizes only run Gen1 images, and Trusted Launch (secure boot, vTPM) needs Gen2. - Marketplace images from third parties can carry a plan that must be accepted once per subscription (
az vm image terms accept) before the first deployment, or it fails withMarketplacePurchaseEligibilityFailed.
Sizes: what the name says
| part | example | meaning |
|---|---|---|
| family | D | general purpose (D), burstable (B), memory (E), compute (F), storage (L), GPU (N), very large memory (M) |
| vCPUs | D2 | number of vCPUs |
| features | D2s | s premium SSD capable, d local temp disk, a AMD, p Arm (Ampere), m more memory |
| version | v5 | hardware generation |
B sizes are burstable: they bank CPU credits when idle and spend them when busy. Perfect for a jump host, a disaster for a build agent that runs flat out (it drops to its baseline, 20-40% of a vCPU, once the credits are gone). For anything that matters you choose by memory per vCPU and whether it needs premium disks:
$ az vm list-sizes -l westeurope --query "[?starts_with(name,'Standard_D2') || starts_with(name,'Standard_E2') || starts_with(name,'Standard_B2')].{size:name, vcpu:numberOfCores, memMB:memoryInMB, disks:maxDataDiskCount}" -o table
Size Vcpu MemMB Disks
---------------- ---- ----- -----
Standard_B2s 2 4096 4
Standard_B2ms 2 8192 4
Standard_D2s_v5 2 8192 4
Standard_D2ds_v5 2 8192 4
Standard_D2as_v5 2 8192 4
Standard_E2s_v5 2 16384 4
Is it available, and may I have it?
A size can exist in a region and still be refused, for two different reasons.
SKU restrictions - capacity or policy keeps a size out of a region or zone for your subscription:
$ az vm list-skus -l westeurope --size Standard_D -o table
ResourceType Locations Name Zones Restrictions
--------------- ---------- ---------------- ----- ------------------------------------------------------------------------------
virtualMachines westeurope Standard_DS1_v2 1,2,3 None
virtualMachines westeurope Standard_DS2_v2 1,2,3 None
virtualMachines westeurope Standard_D2s_v5 1,2 ['NotAvailableForSubscription, type: Zone, locations: westeurope, zones: 3']
virtualMachines westeurope Standard_D4s_v5 1,2,3 None
virtualMachines westeurope Standard_D8s_v5 1 ['NotAvailableForSubscription, type: Zone, locations: westeurope, zones: 2,3']
virtualMachines westeurope Standard_D2ds_v5 1,2,3 None
virtualMachines westeurope Standard_D4ds_v5 1,2,3 None
virtualMachines westeurope Standard_D2as_v5 1,2,3 None
virtualMachines westeurope Standard_D4as_v5 1,2,3 None
Standard_D8s_v5 is offered in zone 1 only here; Standard_D2s_v5 not in zone 3. Ask for it there and the deployment fails with SkuNotAvailable. The fix is another zone, another size of the same family, or another region - not a support ticket, unless the business really needs that exact size there.
Quotas - each subscription has a limit of vCPUs per region, and per VM family in the region:
$ az vm list-usage -l westeurope -o table
Name CurrentValue Limit
------------------------------- ------------ -----
Availability Sets 0 2500
Total Regional vCPUs 3 20
Virtual Machines 2 25000
Virtual Machine Scale Sets 0 2500
Standard BS Family vCPUs 3 10
Standard DSv2 Family vCPUs 0 4
Standard DSv5 Family vCPUs 0 10
Standard DDSv5 Family vCPUs 0 10
Standard DASv5 Family vCPUs 0 10
Standard ESv5 Family vCPUs 0 0
Standard FSv2 Family vCPUs 0 4
Standard NCASv3_T4 Family vCPUs 0 0
Two limits apply at once: Total Regional vCPUs and the family's. Here the E family (memory optimised) has a limit of 0 - nobody asked for it yet. A VM that would exceed either limit fails with OperationNotAllowed ... exceeding approved standardESv5Family Cores quota, and the error contains the exact numbers and a link to request more. Quota requests for most families are approved automatically within minutes; GPU families usually need a business case.
Creating a VM
$ az vm create -g rg-az5-vm -n vm-tools --image Ubuntu2404 --size Standard_D2s_v5 --zone 1 --admin-username azureuser --generate-ssh-keys
{
"fqdns": "",
"id": "/subscriptions/00000000-1111-2222-3333-444444444444/resourceGroups/rg-az5-vm/providers/Microsoft.Compute/virtualMachines/vm-tools",
"location": "westeurope",
"macAddress": "00-0D-3A-F0-A4-2B",
"powerState": "VM running",
"privateIpAddress": "10.0.0.4",
"publicIpAddress": "20.117.5.113",
"resourceGroup": "rg-az5-vm",
"zones": "1"
}
One command, and look at what it actually made:
$ az resource list -g rg-az5-vm --query "[].{kind:type, name:name}" -o table
Kind Name
--------------------------------------- --------------------------------------------------
Microsoft.Network/virtualNetworks vm-toolsVNET
Microsoft.Network/networkSecurityGroups vm-toolsNSG
Microsoft.Network/publicIPAddresses vm-toolsPublicIP
Microsoft.Network/networkInterfaces vm-toolsVMNic
Microsoft.Compute/virtualMachines vm-tools
Microsoft.Compute/disks vm-tools_OsDisk_1_367a73756ec243108116b7c2c982de47
A VNet with a subnet, an NSG, a public IP, a NIC, the VM and its OS disk - six resources, all named after the VM. That is convenient in a lab and wrong for production, where the VM goes into an existing subnet (--vnet-name/--subnet, or --subnet <id> in another group), with no public IP (--public-ip-address "") and no per-NIC NSG (--nsg "") because the subnet already has one. The NSG az vm create makes for a Linux image allows SSH from any source - the default every security scanner flags.
Authentication: --generate-ssh-keys uses ~/.ssh/id_rsa.pub (creating the pair if it does not exist), --ssh-key-values takes a public key you have, and --admin-password turns on password login (Windows always needs one). The admin user cannot be root, admin or administrator.
$ az vm show -g rg-az5-vm -n vm-tools --query "{size:hardwareProfile.vmSize, zone:zones[0], osDisk:storageProfile.osDisk.managedDisk.storageAccountType, sizeGb:storageProfile.osDisk.diskSizeGb, user:osProfile.adminUsername}" -o json
{
"osDisk": "Premium_LRS",
"size": "Standard_D2s_v5",
"sizeGb": 30,
"user": "azureuser",
"zone": "1"
}
$ ssh -o StrictHostKeyChecking=accept-new azureuser@$(az vm show -d -g rg-az5-vm -n vm-tools --query publicIps -o tsv) 'hostname; nproc; free -h | head -2'
Warning: Permanently added '20.117.5.113' (ED25519) to the list of known hosts.
vm-tools
2
total used free shared buff/cache available
Mem: 7.4Gi 373Mi 6.5Gi 4.1Mi 598Mi 7.0Gi
az vm show -d adds the details that live outside the VM resource: power state, public and private IPs, FQDN.
Power states and what you pay
$ az vm stop -g rg-az5-vm -n vm-tools
Virtual machine is being stopped (powered off). You continue to be charged for the compute resources it reserves: use `az vm deallocate` to stop paying for them.
$ az vm list -d -g rg-az5-vm --query "[].{name:name, power:powerState}" -o table
Name Power
-------- ----------
vm-tools VM stopped
$ az vm deallocate -g rg-az5-vm -n vm-tools
$ az vm list -d -g rg-az5-vm --query "[].{name:name, power:powerState}" -o table
Name Power
-------- --------------
vm-tools VM deallocated
az vm stop powers the OS off but keeps the hardware reserved: you keep paying for compute, and the CLI says so. az vm deallocate releases the hardware: compute billing stops, disks and a static public IP are still billed, the temporary disk is wiped, and a dynamic public IP would be given back. Shutting a VM down from inside the OS (sudo poweroff) leaves it stopped, not deallocated
- the classic surprise on the bill. Auto-shutdown schedules deallocate.
Resizing
Sizes run on hardware clusters, and a running VM can only move to a size its current cluster offers:
$ az vm start -g rg-az5-vm -n vm-tools
$ az vm list-vm-resize-options -g rg-az5-vm -n vm-tools --query "[].name" -o tsv
Standard_D2s_v5
Standard_D4s_v5
Standard_D8s_v5
$ az vm resize -g rg-az5-vm -n vm-tools --size Standard_D4s_v5 --query hardwareProfile.vmSize -o tsv
Standard_D4s_v5
A resize reboots the VM, even when it stays on the same cluster - plan it in a maintenance window. A size the cluster does not offer is refused while the VM runs (... is not available in the current hardware cluster); deallocate first and every size in the region becomes possible, then start it. Quotas apply to resizes too:
$ az vm resize -g rg-az5-vm -n vm-tools --size Standard_E4s_v5
ERROR: (OperationNotAllowed) Unable to resize the VM 'vm-tools' because the requested size Standard_E4s_v5 is not available in the current hardware cluster. The available sizes in this cluster are: Standard_D2s_v5,Standard_D4s_v5,Standard_D8s_v5. Read more on VM resizing strategy at https://aka.ms/azure-resizevm.
Code: OperationNotAllowed
Message: Unable to resize the VM 'vm-tools' because the requested size Standard_E4s_v5 is not available in the current hardware cluster. The available sizes in this cluster are: Standard_D2s_v5,Standard_D4s_v5,Standard_D8s_v5. Read more on VM resizing strategy at https://aka.ms/azure-resizevm.
Things a resize cannot do: move between a size with a local temp disk and one without (D2ds <-> D2s) on older generations, move a Gen1 image to a Gen2-only size, or keep the temporary disk's data. Anything on /mnt (Linux) or D: (Windows) is gone after a deallocation or a resize.
Getting in when SSH does not work
$ az vm run-command invoke -g rg-az5-vm -n vm-tools --command-id RunShellScript --scripts "id -un; uptime" --query "value[0].message" -o tsv
Enable succeeded:
[stdout]
root
20:00:32 up 0 min, 1 user, load average: 0.08, 0.08, 0.04
[stderr]
Run Command sends a script through the VM agent, runs it as root and returns its output. It needs no open port and no SSH key - which makes it the way in when the NSG, the SSH daemon or the key is broken, and also why Microsoft.Compute/virtualMachines/runCommand/action is a permission to hand out carefully (it is in Virtual Machine Contributor). It takes up to 90 minutes per script and one script at a time per VM. The other ways in: Bastion (the networking part of this chapter), the serial console (needs boot diagnostics; a login prompt even without a network), and the VM access extension to reset a key or a password.
On the exam (AZ-104 3.2.1, 3.2.4): create a VM with the CLI or PowerShell (
New-AzVM), pick a size from requirements (memory per vCPU, premium disks, GPU), and know that resizing outside the current cluster requires deallocation, that a resize restarts the VM, and thatSkuNotAvailableand quota errors are fixed by another size/zone/region or a quota increase.
In an interview: "A VM deployment fails with SkuNotAvailable. What do you do?" - "Check
az vm list-skusfor that size in the region and zone: it shows the restriction. Pick another zone, a similar size, or another region. If it is a quota error instead,az vm list-usageshows the family and regional limits and I request an increase."