Setting up a production-ready GCP organization is often a time-consuming process. Fabric FAST aims to speed up this process via two complementary goals. On the one hand, FAST provides a design of a GCP organization that includes the typical elements required by enterprise customers. Secondly, we provide a reference implementation of the FAST design using Terraform.
Note that while our implementation is necessarily influenced (and constrained) by the way Terraform works, the design we put forward only refers to GCP constructs and features. In other words, while we use Terraform for our reference implementation, in theory, the FAST design can be implemented using any other tool (e.g., Pulumi, bash scripts, or even calling the relevant APIs directly).
Fabric FAST comes from engineers in Google Cloud's Professional Services Organization, with a combined experience of decades solving the typical technical problems faced by GCP customers. While every GCP user has specific requirements, many common issues arise repeatedly. Solving those issues correctly from the beginning is key to a robust and scalable GCP setup. It's those common issues and their solutions that Fabric FAST aims to collect and present coherently.
Fabric FAST was initially conceived to help enterprises quickly set up a GCP organization following battle-tested and widely-used patterns. Despite its origin in enterprise environments, FAST includes many customization points making it an ideal blueprint for organizations of all sizes, ranging from startups to the largest companies.
FAST uses the concept of stages, which individually perform precise tasks but, taken together, build a functional, ready-to-use GCP organization. More importantly, stages are modeled around the security boundaries that typically appear in mature organizations. This arrangement allows delegating ownership of each stage to the team responsible for the types of resources it manages. For example, as its name suggests, the networking stage sets up all the networking elements and is usually the responsibility of a dedicated networking team within the organization.
From the perspective of FAST's overall design, stages also work as contacts or interfaces, defining a set of pre-requisites and inputs required to perform their designed task and generating outputs needed by other stages lower in the chain. The diagram below shows the relationships between stages.
Security was, from the beginning, one of the most critical elements in the design of Fabric FAST. Many of FAST's design decisions aim to build the foundations of a secure organization. In fact, the first two stages deal mainly with the organization-wide security setup.
FAST also aims to minimize the number of permissions granted to principals according to the security-first approach previously mentioned. We achieve this through the meticulous use of groups, service accounts, custom roles, and Cloud IAM Conditions, among other things.
A resource factory consumes a simple representation of a resource (e.g., in YAML) and deploys it (e.g., using Terraform). Used correctly, factories can help decrease the management overhead of large-scale infrastructure deployments. See "Resource Factories: A descriptive approach to Terraform" for more details and the rationale behind factories.
FAST uses YAML-based factories to deploy subnets and firewall rules and, as its name suggests, in the project factory stage.
As mentioned before, fast relies on multiple stages to progressively bring up your GCP organization(s). Please refer to the stages section for further details.
There are many decisions and tasks required to convert an empty GCP organization to one that can host production environments safely. Arguably, FAST could expose those decisions as configuration options to allow for different outcomes. However, supporting all the possible combinations is almost impossible and leads to code which is hard to maintain efficiently.
Instead, FAST aims to leverage different reference architectures as “pluggable modules”, and then have a small set of variables covering only the essential options of each stage. While we could expose every option of the underlying resources as stage-level variables, we prefer to provide the basic implementation and encourage users to modify the codebase if additional (or different) behavior is needed.
Since we expect users to customize FAST to their specific needs, we strive to make its code easy to understand and modify. Root-level modules (i.e., stages) should be low in complexity, which among other things, means:
- Code should avoid magic and be as explicit as possible.
- We hide advanced features and complexity behind modules.
- We prefer as little indirection as possible.
- We favor flat over nested.
We also recognize that FAST users don't need all of its features. Therefore, you don't need to use our project factory or our GKE implementation if you don't want to. Instead, remove those stages or pieces of code and keep what suits you.
Those familiar with Python will note that FAST follows many of the maxims in the Zen of Python.
Besides the features already described, FAST roadmap includes:
- Stage to deploy environment-specific multitenant GKE clusters following Google's best practices
- Stage to deploy a fully featured data platform
- Reference implementation to use FAST in CI/CD pipelines
- Static policy enforcement