7
votes

I have configured a CI build for a Service Fabric application, in Visual Studio Team Services, according to this documentation: https://azure.microsoft.com/en-us/documentation/articles/service-fabric-set-up-continuous-integration

But instead of having my CI build do the publishing, I only perform the Build and Package tasks, and include all Service Fabric related output, such as pkg folder, scripts, publish profiles and application parameters, in the drop. This way I can pass it along to the new Release pipeline (agent-based releases) to do the actual deployment of my service fabric application.

In my release definition I have a single Azure Powershell task, that uses an ARM endpoint (with proper service principals configured).

When I deploy my app to an existing service fabric cluster, I use the default Deploy-FabricApplication cmdlet passing along the pkg folder and a publish profile that is configured with a connection to the existing cluster.

The release fails with an error message "Cluster connection instance is null". And I cannot understand why?

Doing some debugging I have found that: The Deploy-FabricApplication cmdlet executes the Connect-ServiceFabricCluster cmdlet just fine, but as soon as the Publish-NewServiceFabricApplication cmdlet takes over execution, then the cluster connection is lost.

I would expect that this scenario is possible using the service fabric cmdlets, but I cannot figure out how to keep the cluster connection open during depoyment.

UPDATE: The link to the documentation no longer refers to the Service Fabric powershell scripts, so the pre-condition for this question is no longer documented. The article now refers to the VSTS build and release tasks, which can be prefered over the powershell cmdlets I tried to use.

4
Can you share the error logs? Did it always occurs? And which agent are you using? Usually, the cluster connection shouldn't lost when run Publish-NewServiceFabricApplication .Eddie Chen - MSFT

4 Answers

31
votes

When the Connect-ServiceFabricCluster function is called (from Deploy-FabricApplication.ps1) a local $clusterConnection variable is set after the call to Connect-ServiceFabricCluster. You can see that using Get-Variable.

Unfortunately there is logic in some of the SDK scripts that expect that variable to be set but because they run in a different scope, that local variable isn't available.

It works in Visual Studio because the Deploy-FabricApplication.ps1 script is called using dot source notation, which puts the $clusterConnection variable in the current scope.

I'm not sure if there is a way to use dot sourcing when running a script though the release pipeline but you could, as a workaround, make the $clusterConnection variable global right after it's been set via the Connect-ServiceFabricCluster call. Edit your Deploy-FabricApplication.ps1 script and add the following line after the connection logic (~line 169):

$global:clusterConnection = $clusterConnection  

By the way, you might want to consider setting up custom build/release tasks that deploy a Service Fabric application, rather than using the various Deploy-FabricApplication.ps1 scripts.

2
votes

There now exists a built-in VSTS task for deploying a Service Fabric app so you no longer need to bother with executing the PowerShell script on your own. Task documentation page is at https://www.visualstudio.com/docs/build/steps/deploy/service-fabric-deploy. The original CI article has also been updated which provides details on how to set everything up: https://azure.microsoft.com/en-us/documentation/articles/service-fabric-set-up-continuous-integration/.

1
votes

Try to use "PowerShell" task instead of "Azure PowerShell" task.

0
votes

I hit the same bug today and opened a GitHub issue here

On a side note, VS generated script Deploy-FabricApplication.ps1 uses module

"$((Get-ItemProperty -Path "HKLM:\SOFTWARE\Microsoft\Service Fabric SDK" -Name "FabricSDKPSModulePath").FabricSDKPSModulePath)\ServiceFabricSDK.psm1"

That's where Publish-NewServiceFabricApplication comes from. You can check the deployment logic and rewrite it in more sane way using lower-level ServiceFabric SDK cmdlets (potentially getting connection using Get-ServiceFabricClusterConnection instead of global-ling it)