แก้ไข

แชร์ผ่าน

How to remove a Service Fabric node type

  • บทความ

This article describes how to scale an Azure Service Fabric cluster by removing an existing node type from a cluster. A Service Fabric cluster is a network-connected set of virtual or physical machines into which your microservices are deployed and managed. A machine or VM that's part of a cluster is called a node. Virtual machine scale sets are an Azure compute resource that you use to deploy and manage a collection of virtual machines as a set. Every node type that is defined in an Azure cluster is set up as a separate scale set. Each node type can then be managed separately. After creating a Service Fabric cluster, you can scale a cluster horizontally by removing a node type (virtual machine scale set) and all of its nodes. You can scale the cluster at any time, even when workloads are running on the cluster. As the cluster scales, your applications automatically scale as well.

Warning

Using this approach to remove a node type from a production cluster is not recommended to be used on a frequent basis. It is a dangerous command as it deletes the virtual machine scale set resource behind the node type.

Durability characteristics

Safety is prioritized over speed when using Remove-AzServiceFabricNodeType. The node type must be Silver or Gold durability level, because:

  • Bronze does not give you any guarantees about saving state information.
  • Silver and Gold durability trap any changes to the scale set.
  • Gold also gives you control over the Azure updates underneath scale set.

Service Fabric "orchestrates" underlying changes and updates so that data is not lost. However, when you remove a node type with Bronze durability, you may lose state information. If you're removing a primary node type and your application is stateless, Bronze is acceptable. When you run stateful workloads in production, the minimum configuration should be Silver. Similarly, for production scenarios the primary node type should always be Silver or Gold.

More about Bronze durability

When removing a node type that is Bronze, all the nodes in the node type go down immediately. Service Fabric doesn't trap any Bronze nodes scale set updates, thus all the VMs go down immediately. If you had anything stateful on those nodes, the data is lost. Now, even if you were stateless, all the nodes in the Service Fabric participate in the ring, so an entire neighborhood may be lost, which might destabilize the cluster itself.

Remove a node type

  1. Take care of these pre-requisites before you start the process.

    • The cluster is healthy.
    • There will still be sufficient capacity after the node type is removed, for example, number of nodes to place required replica count.

  2. Move all services that have placement constraints to use node type off the node type.

    • Modify Application / Service Manifest to no longer reference the node type.
    • Deploy the change.

    Then validate that:

    • All the services modified above are no longer running on the Node belonging to the node type.
    • All the services are healthy.

  3. Unmark the node-type as non-primary (Skip for non-primary node types)

    • Locate the Azure Resource Manager template used for deployment.
    • Find the section related to the node type in the Service Fabric section.
    • Change isPrimary property to false. ** Do not remove the section related to the node type in this task.
    • Deploy the modified Azure Resource Manager template. ** Depending on the cluster configuration this step may take a while.

    Then validate that:

    • Service Fabric Section in Portal indicates cluster is ready.
    • Cluster is healthy.
    • None of the nodes belonging to the node type are marked as Seed Node.

  4. Disable each node in the node type.

    Connect to the cluster using PowerShell and then run the following step.

    $nodeType = "" # specify the name of node type
$nodes = Get-ServiceFabricNode

foreach($node in $nodes)
{
  if ($node.NodeType -eq $nodeType)
  {
    $node.NodeName

    Disable-ServiceFabricNode -Intent RemoveNode -NodeName $node.NodeName -Force
  }
}
    • For bronze durability, wait for all nodes to get to disabled state
    • For silver and gold durability, some nodes go in to disabled and the rest will be in disabling state. Check the details tab of the nodes in disabling state, if they are all stuck on ensuring quorum for Infrastructure service partitions, then it is safe to continue.

  5. Stop data for the node type.

    Connect to the cluster using PowerShell and then run the following step.

    foreach($node in $nodes)
{
  if ($node.NodeType -eq $nodeType)
  {
    $node.NodeName

    Start-ServiceFabricNodeTransition -Stop -OperationId (New-Guid) -NodeInstanceId $node.NodeInstanceId -NodeName $node.NodeName -StopDurationInSeconds 10000
  }
}

    Wait till all the nodes for node type are marked Down.

  6. Deallocate nodes in the original Virtual Machine Scale Set

    Login to the Azure subscription where the scale set was deployed and remove the Virtual Machine Scale Set.

    $scaleSetName="myscaleset"
$scaleSetResourceType="Microsoft.Compute/virtualMachineScaleSets"

Remove-AzResource -ResourceName $scaleSetName -ResourceType $scaleSetResourceType -ResourceGroupName $resourceGroupName -Force

  7. Remove data for the node type.

    Connect to the cluster using PowerShell and then run the following step.

    foreach($node in $nodes)
{
  if ($node.NodeType -eq $nodeType)
  {
    $node.NodeName

    Remove-ServiceFabricNodeState -NodeName $node.NodeName -Force
  }
}

    Wait till all the nodes are removed from the cluster. The nodes should not be displayed in SFX.

  8. Remove node type from Service Fabric section.

    • Locate the Azure Resource Manager template used for deployment.
    • Find the section related to the node type in the Service Fabric section.
    • Remove the section corresponding to the node type.
    • Only for Silver and higher durability clusters, update the cluster resource in the template and configure health policies to ignore fabric:/System application health by adding applicationDeltaHealthPolicies under cluster resource properties as given below. The below policy should ignore existing errors but not allow new health errors.
    "upgradeDescription":  
{ 
 "forceRestart": false, 
 "upgradeReplicaSetCheckTimeout": "10675199.02:48:05.4775807", 
 "healthCheckWaitDuration": "00:05:00", 
 "healthCheckStableDuration": "00:05:00", 
 "healthCheckRetryTimeout": "00:45:00", 
 "upgradeTimeout": "12:00:00", 
 "upgradeDomainTimeout": "02:00:00", 
 "healthPolicy": { 
   "maxPercentUnhealthyNodes": 100, 
   "maxPercentUnhealthyApplications": 100 
 }, 
 "deltaHealthPolicy":  
 { 
   "maxPercentDeltaUnhealthyNodes": 0, 
   "maxPercentUpgradeDomainDeltaUnhealthyNodes": 0, 
   "maxPercentDeltaUnhealthyApplications": 0, 
   "applicationDeltaHealthPolicies":  
   { 
       "fabric:/System":  
       { 
           "defaultServiceTypeDeltaHealthPolicy":  
           { 
                   "maxPercentDeltaUnhealthyServices": 0 
           } 
       } 
   } 
 } 
},
    • Deploy the modified Azure Resource Manager template. ** This step takes a while, usually up to two hours. This upgrade change settings to the InfrastructureService, therefore a node restart is needed. In this case forceRestart is ignored. The parameter upgradeReplicaSetCheckTimeout specifies the maximum time that Service Fabric waits for a partition to be in a safe state, if not already in a safe state. Once safety checks pass for all partitions on a node, Service Fabric proceeds with the upgrade on that node. The value for the parameter upgradeTimeout can be reduced to 6 hours, but for maximal safety 12 hours should be used.

    Then validate that:

    • Service Fabric Resource in portal shows ready.

  9. Remove all reference to the resources relating to the node type from the ARM template.

    • Locate the Azure Resource Manager template used for deployment.
    • Remove the virtual machine scale set and other resources related to the node type from the template.
    • Deploy the changes.

    Then:

    • Wait for deployment to complete.

  10. Remove resources relating to the node type that are no longer in use. Example Load Balancer, and Public IP.

    • To remove these resources, you can use the same PowerShell command as used in step 6 specifying the specific resource type and API version.
    • For silver and gold durability any repair task left in the cluster, which is targeting any of the nodes that were present in the nodetype that was removed, should be completed with the command:
       Complete-ServiceFabricRepairTask -TaskId <repair task name>

Note

This step is optional if same Load Balancer, and IP is reused between node types.

Next steps