0 Comments Posted in:

Containers Containers Containers

One of the big themes from this year's Build conference was containers. Service Fabric Mesh will make it easier to run containers on Service Fabric. Web Apps for Containers now supports multi-container web apps and Windows containers. Azure Container Registry (ACR) now has the ability to build container images. And Dev Spaces is a really exciting feature making it really easy for a development team to use Azure Kubernetes Service (AKS) to run and debug their containers in the cloud.

Azure Container Instances

One of my favourite new container related services in Azure is Azure Container Instances (ACI). This is essentially serverless containers and is designed to be the quickest and easiest way to get a container up and running in the cloud.

I've already written about some of the things you could do with ACI, such as building and deploying an ASP.NET Core app, running a media processing task, and running multiple ACI containers together in a "container group".

But the real power of ACI comes not from using them on their own, but combining them with other Azure services. So today I want to highlight some services that integrate really well with Azure Container Instances: Azure Functions, Azure Logic Apps, and AKS. For each one I'll point you to a great demo app that shows off the capabilities of using ACI in conjunction with that service.

Managing ACI Containers with Azure Functions

Azure Container Instances are great for batch jobs, or handling bursts of load. As work becomes available, you can rapidly spin up new instances of your ACI container groups to meet demand or work through the backlog, and then delete them once the demand subsides.

But of course you need something to monitor the load and manage the creation and deletion of those container groups, and Azure Functions is a great fit. If you've created a service principal, you can use that to create or delete new ACI container groups with the Azure .NET SDK or REST APIs.

Brian Peek and Dimitris Gkanatsios have created a superb demo scenario that shows how Azure Container Instances can be managed from Azure Functions. In their example, they monitor how many players are connected in an online game, and if there are too many players for one server to handle, they spin up a new server instance with ACI. They also shut them down when they're no longer needed. You can check out the video and code.

Using ACI Containers as activities in a Durable Functions workflow

I'm a huge fan of Durable Functions for building serverless workflows, but sometimes you need to perform an activity that isn't a good fit for a function. Maybe it needs to run for a long time (the consumption plan limits your function duration to 10 minutes), access an Azure File Share (you can't mount Azure File Shares to an app service plan), or use custom software that is difficult to install directly onto an App Service instance.

One current drawback of using an ACI Container Group to implement a step in a Durable Functions workflow is how to determine when it's finished. I'd love to see Event Grid integration so the container group itself publishes an event when it finishes running. But until then, you'd need to either poll it for status, or get it to report back with a webhook when it's done. Fortunately, the Durable Functions features of waiting for external events, and supporting timeouts for long-running tasks make this kind of pattern possible to implement.

I've not yet got a demo of ACI + Durable Functions to share with you yet, but I've built out most of the necessary pieces, so I'll blog again once I've connected them together. Basically I want to take my media processing Durable Functions workflow sample, but use ACI instances to implement the actual FFMPEG transcodes rather than doing them directly in the Function App.

Using ACI containers in Logic Apps workflows

I've got less experience with Azure Logic Apps. They in many ways are a similar to Durable Functions in that they offer a serverless way of running workflows in the cloud. The development experience is using a visual designer, rather than a code based approach. One strength they have is over 100 connectors to other services, meaning that if your workflow is mostly about connecting different services together, it can be quick to get something up and running.

Logic Apps includes an "ACI Connector" which supports creating and deleting container groups, getting the status or logs of a container group, and listing container groups. This means you can quite easily create a workflow that creates a container group and uses a Logic Apps "until" block to watch its status until a certain condition is met, and then delete the group.

ACI Logic Apps Integration

For a sample application that shows this off, check out this sentiment analysis demo which shows a Logic App triggered off incoming emails. For each email, it extracts the text and then creates an ACI container group that uses Azure Cognitive Services to perform sentiment analysis on that text. It's a slightly contrived example as Logic Apps could easily have just called Cognitive Services directly, but it shows the principle of how you can use a container inside a Logic Apps workflow. The project contains detailed instructions for how to try this out for yourself.

Using ACI for elastic bursting with Azure Kubernetes Service (AKS)

The final integration possibility is to use ACI to provide additional elastic scale to a Kubernetes cluster. Thanks to the "Virtual Kubelet" ACI Connector project you can make ACI appear as a "virtual" node in a Kubernetes cluster, essentially providing infinite scale. The az aks install-connector Azure CLI command makes this especially easy to set up if your Kubernetes cluster is an instance of AKS.

This diagram was shown in several presentations at this year's Build conference to show how the virtual Kubelet ACI connector integrates with an AKS cluster:

AKS with ACI connector

There's a great demo application that shows the ACI connector in action available here. The scenario is a facial recognition application. By default, the one node AKS cluster only has a single instance of the "image recognizer" container running, resulting in a backlog of tasks building up. But with a single command you can scale up to multiple instances running in ACI, enabling you to rapidly work through the backlog. You can see a video of this demo app in action in this Azure Friday episode with Ria Bhatia


Azure Container Instances really shines when used in conjunction with other Azure services, bringing the benefits of containers, per-second billing, and elastic scale to your existing Functions Apps, Logic Apps and Kubernetes deployments.

Want to learn more about how easy it is to get up and running with Durable Functions? Be sure to check out my Pluralsight course Azure Durable Functions Fundamentals.

0 Comments Posted in:

I'm really pleased to announce that my latest Pluralsight course has been released. In Durable Functions Fundamentals I show you everything you need to get started with developing and debugging durable workflows locally, how to implement patterns such as fan-out fan-in and waiting for human interaction, and how to deploy and monitor your Durable Functions online.

To celebrate the launch here's a quick rundown of my top 10 reasons why you should use Durable Functions for your serverless workflows:

1. Express your workflows in code

With Durable Functions, you get to define your workflows in code. This means one place to look to see the big picture of what the whole workflow does, rather than it being spread across multiple functions. Currently Durable Functions orchestrators have to be written in C#, but JavaScript support is in beta so you can define your workflows in the language you're most comfortable with.

2. Retry activities

Another great feature of Durable Functions is the support for retrying activities with back-off. That used to be awkward to implement in regular Azure Functions, but with Durable Functions it's trivially easy to add retries to both individual activity functions and to sub-orchestrations, giving your workflows much greater resilience to transient errors.

3. Run activities in parallel

In a typical multi-step workflow there are probably some activities that can be performed in parallel, but without a framework like Durable Functions, implementing the "fan-in" part of a "fan-out fan-in" pattern is complex and risks introducing race conditions.

With Durable Functions, running activities (or whole sub-orchestrations) in parallel is easy to accomplish, and combined with the power of Azure Functions to scale out, brings the potential for dramatic speedup in the end-to-end time your workflows take to complete.

4. Timeout workflows

Sometimes in a workflow, you're waiting for some kind of external event - maybe for a human to respond, or for an external system to send you a message, but you want to time out if you don't receive the event within a certain time period. Durable Functions makes this pattern straightforward to implement, allowing you to detect that an orchestration has got stuck, and take some kind of mitigating action to get it moving again, or to alert a system administrator.

5. State management for free

Workflows inherently have state associated with them - you need to know where you've got to in the workflow in order to decide what the next step is whenever an activity completes. Durable Functions transparently manages workflow state for you, meaning you can implement complex workflows without needing your own database at all. Of course, if you do have a database, you may still will want to update it during your workflows, to track the state of your business entities, but you don't need to manage the state for the orchestrations themselves.

6. Check on workflow progress with REST API

If you've ever built out a workflow with regular Azure Functions, you'll know it can be a real pain to work out where in the pipeline you currently are. This can be especially important for trouble-shooting if a workflow has got stuck. How far through did it get before failing?

With Durable Functions you can use the query status API to find out if an orchestration is still running or not. The query status API includes a showHistory flag to request to see the history of the workflow, allowing you to see exactly where it got up to before it got stuck or failed.

Even better, there is now a SetCustomStatus API allowing you to store an arbitrary JSON object at any point in your workflow representing its current status. This is a great tool for diagnosing why an orchestration is taking longer than expected to complete.

7. Cancel workflows

If you've built a workflow out of regular Azure Functions chained together with queue messages, then cancelling it is not going to be easy. But with Durable Functions, the REST API includes a cancellation method making it really straightforward to abandon an in-progress workflow.

8. Serverless pricing model

Just because your workflows run for days at a time, doesn't mean you need to pay for days of compute. In fact, in many long-running workflows, most of the time is spent just waiting around. Because Durable Functions is built on top of Azure Functions, you get all the benefits of a serverless pricing model. You only pay for the time your functions are actually running and your orchestrator function invocations will all be extremely quick as they simply wake up, decide what the next step in the workflow is, and go straight back to sleep.

9. Versioning made easier

One of the hardest problems of implementing workflows is how to deal with versioning. If I make a breaking change to the workflow, what happens to in-flight orchestrations when I perform an upgrade? Durable Functions doesn't have a magic bullet to solve the problem, but it does provide several workable strategies for dealing with this issue. Currently I'm leaning towards just making a V2 version of my orchestrator functions and retiring the V1 orchestrator later once all old workflows have finished, but you can pick the versioning strategy that works best for you.

10. Develop and test locally

Finally, it's possible to develop and test your Durable workflows locally. You can get the full local debugging experience of stepping through orchestrator and activity functions, as well as examining the contents of your "task hub" (which can also be local if you are using the Azure Storage Emulator) using Storage Explorer.

When you do publish your workflows to Azure, then the Application Insights integration gives you access to rich and powerful querying capabilities on your Function App telemetry and logs.


If you're currently building workflows out of a series of Azure Functions triggering each other, then Durable Functions is a no-brainer. It really is a game-changer that makes development and management of your serverless workflows much easier. Do give it a try and if you're a Pluralsight subscriber then my new Durable Functions Fundamentals course will teach you the key concepts and provide lots of examples of the sorts of workflows you can build with Durable Functions.

Want to learn more about how easy it is to get up and running with Durable Functions? Be sure to check out my Pluralsight course Azure Durable Functions Fundamentals.

0 Comments Posted in:

I was really pleased to see Durable Functions went GA yesterday, and continues to pick up some great new features, such as the ability to write orchestrator functions in JavaScript (still in preview). If you've not tried Durable Functions yet, it really is a game-changer, giving you a much better way to manage multiple functions that form a workflow, and greatly simplifies the implementation of complex workflows such as fan-out fan-in (map-reduce) and waiting for human interaction.


In this post I want to highlight an interesting feature of Durable Functions called "sub-orchestrations". In Durable Functions an "orchestrator" function describes the order of the steps in your workflow, and "activity" functions are used to implement each of those steps.

With sub-orchestrations, an orchestrator function calls into another orchestrator function, allowing you to make workflows that are themselves built up of other workflows.

Why sub-orchestrations?

But why would you want to do this? When I first read about sub-orchestrations, I didn't initially think they would be a particularly important feature, but the more workflows I have built, the more benefits I can see for using them.

So here's a quick run-through of some of the reasons why I think you should consider using them once an orchestrator function grows to call more than about four or five activities.

1. Clean code

Real-world workflows consist of multiple steps, and tend to grow in complexity over time. If you trying to ensure that each of your activity functions has a "single responsibility" (which you should be), then you're likely to end up with a lot of them, resulting in a long and complex orchestrator function.

Also the strict "orchestrator function constraints" in Durable Functions, which stipulate that your orchestrator functions must be deterministic, have a tendency to increase the number of activity functions in use, as you need to create an activity function each time you need to perform any non-deterministic task such as fetching a value from a database or config.

Using sub-orchestrations allows you to logically group together smaller sections of your workflow, which makes for much easier to read and understand code than one giant function consisting of numerous activities.

Here's a very simple code example showing how an orchestrator function might run three sub-orchestrations in sequence:

public static async Task MultiStageOrchestrator(
    [OrchestrationTrigger] DurableOrchestrationContext ctx)
    string input = ctx.GetInput<string>();

    var output1 = await ctx.CallSubOrchestratorAsync<string>("Stage1", input);
    var output2 = await ctx.CallSubOrchestratorAsync<string>("Stage2", output1);
    await ctx.CallSubOrchestratorAsync("Stage3", output2);

2. Error-handling and retrying

I wrote recently about how great Durable Functions is for handling errors. It allows you to handle errors for the workflow as a whole, or for individual functions. But if you have a large and complex workflow made up of a few smaller workflows implemented as sub-orchestrations, then it becomes possible for each sub-orchestration to manage its own exception handling with any clean-up that is appropriate just for that part of the overall workflow.

Sub-orchestrations can be retried with back-offs, in exactly the same way that activity functions can. This is very powerful, as retrying a series of activities without the use of sub-orchestrations would be complex to implement.

So here we can see how the second stage in our example above could be set up to retry up to four times, with a back-off of five seconds:

var output2 = await ctx.CallSubOrchestratorWithRetryAsync<string>("Stage2", 
                         new RetryOptions(TimeSpan.FromSeconds(5),4), output1);

3. Run sub-orchestrations in parallel

Another thing you might notice if you take the trouble to break a long and complex workflow up into sub-orchestrations is that some of them could be run in parallel as they don't depend on each other's output. Just like you can run activities in parallel, implementing a fan-out fan-in pattern, you can do exactly the same with sub-orchestrations, kicking off several different sub-orchestrations (or several instances of the same sub-orchestration) and then waiting for them all to complete with Task.WhenAll.

Running orchestrations in parallel opens the door for performance boosts that would be too much of a pain to implement without the benefit of sub-orchestrations.

In this example, two sub-orchestrations ("stage1" and "stage2") are started in parallel, then we wait for both to complete, and use their outputs in a call to a third stage.

public static async Task ParallelSubOrchestrations(
    [OrchestrationTrigger] DurableOrchestrationContext ctx)
    string input = ctx.GetInput<string>();

    var stage1Task = ctx.CallSubOrchestratorAsync<string>("Stage1", input);
    var stage2Task = ctx.CallSubOrchestratorAsync<string>("Stage2", input);

    await Task.WhenAll(stage1Task, stage2Task);

    await ctx.CallSubOrchestratorAsync("Stage3", Tuple.Create(stage1Task.Result, stage2Task.Result));

4. Reuse across workflows

If you break complex workflows up into sub-orchestrations, you make find that the same sub-orchestration can be re-used by multiple different orchestrators. This eliminates duplication, or the need to make orchestrators that contain complex branching code. If there is some shared logic used by two different workflows, put it into a sub-orchestration that they can both make use of.

In this simple example, "workflow1" and "workflow2" share a common "SharedStage" orchestrator but also perform different tasks before or after.

public static async Task Workflow1(
    [OrchestrationTrigger] DurableOrchestrationContext ctx)
    string input = ctx.GetInput<string>();
    var output1 = await ctx.CallSubOrchestratorAsync<string>("StageA", input);
    var output2 = await ctx.CallSubOrchestratorAsync<string>("SharedStage", output1);
    await ctx.CallSubOrchestratorAsync("StageB", output2);

public static async Task Workflow2(
    [OrchestrationTrigger] DurableOrchestrationContext ctx)
    string input = ctx.GetInput<string>();
    var output1 = await ctx.CallSubOrchestratorAsync<string>("StageD", input);
    var output2 = await ctx.CallSubOrchestratorAsync<string>("SharedStage", output1);
    await ctx.CallSubOrchestratorAsync("StageE", output2);

5. Simplified event sourcing history

Behind the scenes, Durable Functions uses an "event sourcing" approach to storing the history of orchestrations. Every time an activity completes, the orchestrator wakes up and must "replay" through all the prior events that have happened in this orchestration to reconstruct the current state of the workflow and work out what to do next.

The longer and more complex an orchestrator is, the more event sourcing steps must be stored and replayed, making debugging a pain (if you are using breakpoints in your orchestrators), and possibly impacting performance.

However, if sub-orchestrations are used, the event sourcing history for the parent orchestration can replace the entire call to a sub-orchestration with its serialized JSON output, greatly reducing the number of overall events stored against the parent orchestration. This leads to the final benefit I want to mention.

6. Safer upgrades

One possible gotcha with Durable Functions is what happens when you upgrade your code while orchestrations are in progress. You need to be very careful if you do this as any breaking changes to your orchestrator (such as re-ordering functions, or changing the input or output format of activities) will result in things going wrong when your event sourcing history that was generated by a previous version of the orchestrator function is replayed against the new one.

Sub-orchestrations can't save you from this, but can actually provide some protection, if breaking changes can be isolated to a single sub-orchestration, which could allow a larger workflow to recover even if one of its sub-orchestration steps fails.


Durable Functions sub-orchestrations allow you to break large and complex workflows into more granular pieces, which open the door to retries, better error handling, reuse, and parallel execution. They also make for easier to read code, and could help make upgrades more reliable.

Want to learn more about how easy it is to get up and running with Durable Functions? Be sure to check out my Pluralsight course Azure Durable Functions Fundamentals.