There is no place like home, and also: there is no place like production. Production is the only location where your code is really put to the test by your users. This insight has helped developers to accept that we have to go to production fast and often. But we also want to do so in a responsible way. Common tactics for safely going to production often are for example blue-green deployments and the use of feature flags.

While helping to limit the impact of mistakes, the downside of these two approaches is that they both run only one version of your code. Either by using a different binary or by using a different code path, the one or the other implementation is executed. But what if we could execute both the old and the new code in parallel and compare the results? In this post I will show you how to run two algorithms in parallel and compare the results, using a library called scientist.net, which is available on NuGet.

The use case

To experiment a bit met experimentation I have created a straight forward site for calculation the nth position in the Fibonacci sequence as you can see below.

I started out with the most simple implementation that seemed to meet the requirements of the implementation. In other words, the implementation behind this calculation is done using recursion, like shown below.

While this is a very clean and to-the-point implementation code wise, the performance is -to say the least- up for improvement. So I took a few more minutes and I came up with an implementation which I believe is also correct, but also much more performant, namely the following:

However, just swapping implementations and releasing didn’t feel good to me, so I thought: how about running an experiment on this?

The experiment

With the experiment that I am going to run, I want to achieve the following goals:

  • On every user request, run both my recursive and my linear implementation.
  • To the user I want to return the recursive implementation, which I know to be correct
  • While doing this, I want to record:
    • If the linear implementation yields the same results
    • Any performance differences.

To do this, I installed the Scientist NuGet package and added the code shown below as my new implementation.

This code calls into the Scientist functionality and sets up an experiment with the name fibonacci-implementation that should return an int. The configuration of the experiment is done using the calls to Use(..) and Try(..)

Use(..): The use method is called with a lambda that should execute the known, trusted implementation of the code that you are experimenting with.

Try(..): The try method is called with another lambda, but now the one with the new, not yet verified implementation of the algorithm.

Both the Use(..) and Try(..) method accept a sync-lambda as well, but I do use async/await here on purpose. The advantage of using an async-lambda is that both implementations will be executed in parallel, thus reducing the duration of the web server call. The final thing I do with the call to the AddContext(..) method is adding a named value to the experiment. I can use this context property-bag later on to interpret the results and to pin down scenarios in which the new implementation is lacking.

Processing the runs

While the code above takes care of running two implementations of the Fibonacci sequence in parallel, I am not working with the results yet – so let’s change that. Results can be redirected to an implementation of the IResultPublisher interface that ships with Scientist by assigning an instance to the static ResultPublisher property as I do in my StartUp class.

In the ExperimentResultPublisher class, I have added the code below.

For all instances of the fibonacci-implementation experiment, I am saving the results of the last observation. Observation is the Scientist term for a single execution of the experiment. Once I have moved the results over to my own class LastResults, I am adding these last results to another class of my own OverallResults that calculated the minimum, maximum and average for each algorithm.

The LastResults and OverallResults properties are part of the IExperimentResultsGetter interface, which I later on inject in my Razor page.

Results

All of the above, combined with some HTML, will then gave me the following results after a number of experiments.

I hope you can see here how you can take this forward and extract more meaningful information from this type of experimentation. One thing that I would highly recommend is finding all observations where the existing and new implementation do not match and logging a critical error from your application.

Just imagine how you can have your users iterate and verify all your test cases, without them ever knowing.

I am a fan of private agents when working with Azure Pipelines, compared to the hosted Azure Pipelines Agents that are also available. In my experience the hosted agents can have a delay before they start and sometimes work slow in comparison with dedicated resources. Of course this is just an opinion. For this reason, I have been running my private agents on a virtual machine for years. Unfortunately, this solution is not perfect and has a big downside as well: no isolation between jobs.

No isolation between different jobs that are executed on the same agent means that all files left over from an earlier job, are available to any job currently running. It also means that it is possible to pollute NuGet caches, change files on the system, etc etc. And of course, running virtual machines in general is cumbersome due to patching, updates and all the operational risks and responsibilities.

So it was time to adopt one of the biggest revolutions in IT: containers. In this blog I will share how I created a Docker container image that hosts an Azure Pipelines agent and how to run a number of those images within Azure Container Instances. As a starting point I have taken the approach that Microsoft has laid down in its documentation, but I have made a number of tweaks and made my solution more complete. Doing so, I wanted to achieve the following:

  1. Have my container instances execute only a single job and then terminate and restart, ensuring nothing from a running job will become available to the next job.
  2. Have a flexible number of containers running that I can change frequently and with a single push on a button.
  3. Have a source code repository for my container image build, including a pipelines.yml that allows me to build and publish new container images on a weekly schedule.
  4. Have a pipeline that I can use to roll-out 1..n agents to Azure Container Instances – depending on the amount I need at that time.
  5. Do not have the PAT token that is needed for (de)registering agents available to anyone using the agent image.
  6. Automatically handle the registration of agents, as soon as a new container instance becomes available.

Besides the step-by-step instructions below I have also uploaded the complete working solution to GitHub at https://github.com/henrybeen/ContainerizedBuildAgents.

Let’s go!

Creating the container image

My journey started with reading the Microsoft documentation on creating a Windows container image with a Pipelines Agent. You can find this documentation at https://docs.microsoft.com/en-us/azure/devops/pipelines/agents/docker?view=azure-devops. I found two downsides to this approach for my use case. First, the PAT token that is used for (de)registering the agent during the complete lifetime of the container. This means that everyone executing jobs on that agent, can pick up the PAT token and abuse it. Secondly, the agent is downloaded and unpacked at runtime. This means that the actual agent is slow to spin up.

To work around these downsides I started with splitting the Microsoft provided script into two parts, starting with a file called Build.ps1 as shown below.

The script downloads the latest agent, if agent.zip is not existing yet, and unzips that file. Once the agent is in place, the docker image is build using the call to docker build. Once completed, the unpacked agent.zip folder is removed – just to keep things tidy. This clean-up also allows for rerunning the script from the same directory multiple times without warnings or errors. A fast feedback loop is also the reason I test for the existence of agent.zip before downloading it.

The next file to create is the Dockerfile. The changes here are minimal. As you can see I also copy over the agent binaries, so I do not have to download these anymore when the container runs.

First we ensure that the directory c:\azdo\work exists by setting it as the working directory. Next we move to the directory that will contain the agent files and copy those over. Finally, we move one directory up and copy the Start-Up script over. To run the container image, a call into that script is made. So, let’s explore Start-Up.ps1 next.

The script first checks for the existence of four, mandatory, environment variables. I will provide these later on from Azure Container Instances, where we are going to run the image. Since the PAT token is still in this environment variable, we are setting another variable that will ensure that this environment variable is not listed or exposed by the agent, even though we will unset it later on. From here on, the configuration of the agent is started with a series of command-line arguments that allow for a head-less registration of the agent with the correct agent pool.

It is good to know that I am using a non-random name on purpose. This allows me to re-use the same agent name -per Azure Container Instances instance- which prevents an ever increasing list of offline agents in my pool. This is also the reason I have to add the –replace argument. Omitting this would cause the registration to fail.

Finally, we run the agent, specifying the –once argument. This argument will make that the agent will pick-up only a single job and terminate once that job is complete. Since this is the final command in the PowerShell script, this will also terminate the script. And since this is the only CMD specified in the Dockerfile, this will also terminate the container.

This ensures that my container image will execute only one job ever, ensuring that the side-effects of any job cannot propagate to the next job.

Once these files exist, it is time to execute the following from the PowerShell command-line.

and…

This shows that the container image can be build and that running it allows me to execute jobs on the agent. Let’s move on to creating the infrastructure within Azure that is needed for running the images.

Creating the Azure container registry

As I usually do, I created a quick ARM templates for provisioning the Azure Container Registry. Below is the resource part.

Creating a container registry is fairly straightforward, specifying a sku and enabling the creation of an administrative account is enough. Once this is done, we roll this template out to a resource group called Tools, tag the image against the new registry, authenticate and push the image.

Now that all the scripts and templates are proven, let’s automate this through a pipeline. The following YAML is enough to build and publish the container.

To make the above work, two more changes are needed:

  1. The Build.ps1 needs to be changed a bit, just remove steps 4 and 5
  2. A service connection to the ACR has to be made with the name BuildAgentsAcr

With the container image available in an ACR and a repeatable process to get updates out, it is time to create the Azure Container Instances that are going to run the image.

Creating the Azure container instance(s)

Again, I am using an ARM template for creating the Azure Container Instances.

At the start of the resource definition, we specify that we want multiple copies of this resource, not just one. The actual number of resources is specified using a template parameter. This construct allows us to specify the number of agents that we want to run in parallel in ACI instances every time this template is redeployed. If we combine this with a complete deployment mode, the result will be that agents in excess of that number get removed automatically as well. When providing the name of the ACI instance, we are concatenating the name with the outcome of the function copyIndex(). This function will return a integer, specifying in which iteration of the copy loop the template currently is. This way unique names for all the resources are being generated.

As with most modern resources, the container instance takes the default resource properties in the root object and contains another object called properties that contains all the resource specific configuration. Here we first have to specify the imageRegistryCredentials. These are the details needed to connect ACI to the ACR, for pulling the images. I am using ARM template syntax to automatically fetch and insert the values, without ever looking at them.

The next interesting property is the RestartPolicy. The value of Always instructs ACI to automatically restart my image whenever it completes running, no matter if that is from an error or successfully. This way, whenever the agent has run a single job and the container completes, it gets restarted within seconds.

In the second properties object, a reference to the container image, the environment variables and the container resources are specified. The values here should be self-explanatory, and of course the complete ARM template with all parameters and other plumbing is available on the GitHub repository.

With the ARM template done and ready to go, let’s create another pipeline – now using the following YAML.

Nothing new in here really. This pipeline just executes the ARM template against another Azure resource group. Running this pipeline succeeds and a few minutes later, I have the following ACI instances.

And that means I have following agents in my Pool.

All-in-all: It works! Here you can see that when scaling the number of agents up or down, the names of the agents stay the same and that I will get a number of offline agents that is at most the number of agents to which I have scaled up at a certain point in time.

And with this, I have created a flexible, easy-to-scale, setup for running Azure Pipelines jobs in isolation, in containers!

Downsides

Of course, this solution is not perfect and a number of downsides remain. I have found the following:

  1. More work
  2. No caching

More work. The largest drawback that I have found with this approach is that it is more work to set up. Creating a virtual machine that hosts one or more private agents can be done in a few hours, while the above has taken me well over a day to figure out (granted, I was new to containers).

No caching. So slower again. With every job running in isolation, many of the caching benefits that come with using a private agent in a virtual machine are gone again. Now these can be overcome by building a more elaborate image, including more of the different SDK’s and tools by default – but still, complete sources have to be downloaded every time and there will be no intermediate results being cached for free.

If you find more downsides, have identified a flaw, or like this approach, please do let me know!

But still, this approach works for me and I will continue to use it for a while. Also, I have learned a thing or two about working with containers, yay!