StackSets are a very powerful way to deploy software at scale across multiple AWS accounts and also to multiple regions within a single account.
AutoSpotting supports being deployed using StackSets across multiple accounts, and also leverages them internally to deploy some of its components across multiple regions within each account.
This document explains the way the current CloudFormation deployment method of AutoSpotting uses CloudFormation StackSets internally, what consequences it has on existing StackSet environments or when AutoSpotting is installed repeatedly within an AWS account, and a few workarounds on how to deploy AutoSpotting on such environments.
- The current AutoSpotting architecture deploys a few central resources in a main AWS region but requires additional resources to be deployed in other regions to enable certain advanced behaviors such as handling of spot instance termination or rebalancing events, immediate replacement of on-demand instances with spot and startup lifecycle hook emulation. Without these regional resources AutoSpotting will only run in the basic/legacy Cron mode, which is suboptimal.
- These regional resources have been historically deployed using a second regional CloudFormation template, deployed by a custom Lambda-backed CloudFormation resource across all available AWS regions. This was error-prone in particular when having multiple installations of AutoSpotting side by side, especially when some of them were uninstalled.
- There was also another installation mode in which the same main CloudFormation template could itself be deployed using a StackSet, deploying only the main resources in the main region and only the regional resources in the other regions based on some parameters. This duplicated a lot of infrastructure code between the two CloudFormation templates and many conditionals that overcomplicated the implementation of the main template enough that it became almost unmanageable, so we decided to simplify it.
- The current implementation uses a StackSet to deploy the regional resources instead of the custom Lambda-backed resource, and has been simplified greatly by removing all the conditionals and duplicated code that enabled it to use a StackSet for deploying the regional resources with the same template code.
- To keep the user-friendly single-click installation support, the main AutoSpotting CloudFormation template currently also deploys out of the box the required IAM resources needed for self-managed StackSet permissions that enable it to deploy the regional template as a StackSet, and that's why it's conflicting with self-managed StackSet permissions you may already have in your account or if they were created by another installation of AutoSpotting. If you run into any such installation issues, you can see the Workarounds section below.
- Grant permissions for using StackSets at AWS Organization level using these instructions. See Note 1 below for further information on this topic.
- Use the same CloudFormation template and parameters used for deploying AutoSpotting in stand-alone mode within a single AWS account.
- Create the StackSet only in the "Main" region. For the official binaries you
will need to use
us-east-1
, otherwise installation fails. See Note 2 below in case you run a custom build hosted in another region. - Set the OrganizationUnit where you want to deploy AutoSpotting and complete the installation wizard. You can also use the Organization root to install AutoSpotting across an entire AWS Organization.
- For faster installation you can allow a 50% failure percentage, otherwise the StackSet will deploy AutoSpotting only one account at a time which can be slow on large organizations.
- Self-managed StackSet permissions are not supported out of the box and will break the default installation of AutoSpotting if you have them configured in the account. See below the workaround for this issue if you run into it.
- If you run a custom build use the region where you created the S3 bucket hosting the code of the customized AutoSpotting Lambda functions.
- In case the installation fails because of the conflict with pre-existing IAM
resources created for self-managed StackSet permissions or by an existing
AutoSpotting installation, you'll need to configure the AutoSpotting Stack
parameters to not deploy any regional resources. You can do it by setting the
DeployRegionalResourcesStackSet
parameter tofalse
. - This which will render the current AutoSpotting installation to run in the legacy cron mode, also lacking termination event handling and lifecycle hooks emulation. In order to re-enable the event-based execution mode and other advanced features, you will then need to deploy those regional resources yourself with a second AutoSpotting regional StackSet deployed to the regions you want to run AutoSpotting against. For this you will need to use the regional CloudFormation template.
- This regional StackSet will need a couple of parameters that enable it to send events to the main Lambda function: the ARN of the main Lambda function and the ARN of the regional execution IAM role created by the main AutoSpotting Stack.
- These would be set automatically when installing the main AutoSpotting
CloudFormation template with the default parameters, but they need to be
manually set if the regional stack is installed manually. You can get these
values from the Outputs of the main AutoSpotting CloudFormation Stack that
corresponds to the regional Stack you want to install, in particular the
AutoSpottingLambdaARN
andLambdaRegionalStackExecutionRoleARN
output values.
- As mentioned above, the current StackSet implementation requires manual workarounds inc certain situations, such as when the StackSet self-managed StackSet permissions already exist in the account or when AutoSpotting is installed multiple times within an account. The above Workarounds may help in such situations.
- Parallel installations using StackSets will require the installation in legacy
mode by setting the
DeployRegionalResourcesStackSet
parameter tofalse
on all but the first StackSet, and performing the same workarounds on each target account where the subsequest StackSets are deployed.
As always, if you need Enterprise support for more exotic or large configurations, you can get in touch on gitter.