UPDATE: 2020-01-30 export with https://resources.azure.com/

If you found your way to this post and your first reaction is “well duh”, please leave a comment, because I’d love to know how you came upon your knowledge which caused you to respond this way. This was hard-won knowledge for me, but I suspect I’m missing out on some fundamental bit of training in the art of efficient ARM template development.

An ARM template encodes instructions that the Azure Resource Manager executes which result in some stuff being deployed to Azure. As such, an ARM template is a DSL for the domain of Azure. Unfortunately, as DSLs go, it’s awful. It’s a bigass JSON file, with all of the pitfalls that entails, such as:

  • It’s a challenge to manage complexity. Things can spiral out of control very quickly unless you are very disciplined while authoring.

  • The instructions in an ARM template execute over time, in a sequence, imperatively, but JSON is absolutely not an imperative syntax.

  • It’s very easy to make stupid mistakes at authoring time. You can opt-in to using some tooling to catch some stupid mistakes at authoring time, but a proper DSL would do a better job of it.

  • You can’t even put comments in there. UPDATE: 2020-04-21: // comments in JSON work now.

As a DSL for Azure, I think Terraform may be a better fit. Indeed Terraform is well supported by Azure, but you can’t use it when authoring for the Azure marketplace. Accept JSON and move on. Here are some tips to ease your acceptance.

Accept JSON

This blog post shares some tips I’ve developed to get faster development cycle times when iteratively developing ARM templates. The cycle looks like this:

  1. Get your mainTemplate.json somehow. Write it yourself, export it from the Azure portal, start with a quickstart, etc.

  2. Deploy it to Azure with the Azure CLI.

  3. Look at what happened.

I want to minimize the time between steps 2 and 3.

Tips for step 1

Use the “export template” feature as a jump start

Create the system you want to have in the portal, then use the export template feature of the portal to download the ARM template JSON file. You may want to do this for each big chunk of functionality in your system, then stitch them together. Iterate at the command line using the technique in the next tip.

Use the resources.azure.com

I have observed that the “export template” feature does not always give you the full JSON. For example, in the case of Application Gateway, the backendAddresses was empty, even though it had a pool and targets. A colleague pointed out this enormously useful resource https://resources.azure.com/.

Here is a quick sketch of how I used it to get the full JSON of a deployed Application Gateway. Note, this is not an ARM template, but it can help a lot.

  1. Visit https://resources.azure.com/

  2. Type in the resource group in which the resource is deployed: ejb012803c

  3. Maybe make a choice.

    a. If offered with an autosuggest, pick the resource and you are done.

    b. If not offered with an autosuggest. In the left pane, expand nodes like this

    subscriptions > Your subscription > resourceGroups > ejb012803c > Providers >
      Microsoft.Network > applicationGateways > ejb012803c

Start with one of the templates from azure-quickstart-templates

My extremely helpful colleague Brian Moore leads a team that curates this vast resource of templates, and tests for templates. The templates are organized by the kind of thing you want to deploy to Azure. For example, here is one for Azure Application Gateway.

Tips for step 2

Always use the --debug argument to commands for which you need to iterate. Like this

az group deployment create --debug --resource-group ejb012405c --parameters @20200124-dd-1052162-gateway-parameters.json --template-file arm-oraclelinux-wls-cluster/arm-oraclelinux-wls-cluster/src/main/arm/mainTemplate.json > azure.out 2>&1 &

This also causes the output to be redirected to a file, in the background, capturing stdout and stderr.

A tale of two tails

Use some tails to observe the output as the command runs.

  1. Grep through the outfile for some low frequency string.

    tail -f azure.out | grep "INFO: Accepted"
  2. In a separate terminal window, just tail the output for everything, so you can see if the command simply failed.

    tail -f azure.out

Extract the {"properties": {"template": JSON object from azure.out, so you can interpret column offsets correctly

The template you pass to az group deployment create will have all formatting removed before Azure sees it, so when Azure gives you error messages, they are usually expressed like “error on line 1 column 60299”. You need to translate that to “line N” in your human readable JSON. If you save aside the {"properties": {"template": JSON object from the azure.out file, you can seek to that column, and then derive where the error is in your human readable JSON file.

Even though you ultimately need to produce mainTemplate.json iterate with smaller, disjoint files

If you are building a template that ultimately deploys an entire datacenter, such as with an Azure Application you should consider breaking the template into smaller nested templates, as described in Using linked and nested templates when deploying Azure resources


Developing ARM templates can be very frustrating, especilaly if you are used to the rapid compile-edit-debug cycle that many actual programming languages offer.