Note: Canonical Services are supported automatically in Anthos Service Mesh version 1.6.8 and higher.
Canonical Services allow you to navigate many different configurations. For the best experience with Anthos Service Mesh Service Dashboards, consider the following standard practices when setting up your services:
- Reserve a unique service [namespace, name] across the whole mesh.
- Define one software application per Canonical Service.
- Do not group Canonical Services across environments (for example, prod/stage/dev).
- Use Cloud Monitoring dashboards for higher-level views of multiple services.
- Plan for Canonical Services to be long-lived in production.
Reserve a unique service [namespace, name] across the whole mesh
If a Canonical Service deployed in one cluster or region has the same Kubernetes namespace and Canonical Service name as one deployed in another cluster or region, Anthos Service Mesh assumes that it is the same logical service.
This behavior is consistent with Environ's principle of "sameness", which says that a namespace should have the same meaning and represent the same entity across the entire Environ.
One software application per Canonical Service
Canonical Services are meant to represent a single logical service or microservice. They are meant to span homogenous binaries/workloads that represent the same software application and business function.
While you could define a Canonical Service to group several conceptually different microservices together, the Service Dashboards would not provide their full value.The Service Dashboards would display an aggregation of dissimilar components which may individually be performing and configured very differently. It would be difficult or even impossible to understand the health, performance, and configuration of the whole.
The following are not necessarily bad practices, but your Canonical Service may be too big if:
- There is network traffic between different workloads within a single Canonical Service.
- A Canonical Service comprises multiple workloads that are deployed on different release schedules.
- Different teams within your organization are responsible for operating different pieces of a single Canonical Service.
Do not group a Canonical Service across environments
Many technology organizations employ multiple deployment environments to ensure software quality and limit risk. These environments most often include dev, test, staging, prod, or some subset.
Even if you deploy the same conceptual service across each of your various environments, it is bad practice to make them a single Canonical Service. These services are not the same and do not represent the same level of operational concern or focus for your organization.
For instance, a failure on a critical production service may cause 3AM pages and firefights. You do not want to alert anybody if the "dev" deployment fails in the middle of the night. The same goes for understanding performance, capacity, and release safety.
From the easiest but least rigorous, to the highest-effort but most powerful, there are three ways to separate services into different envirionments:
- Separate using multiple service names, for example,
- Separate using multiple namespaces, for example
- Separate using multiple Environs, one for each environment.
Prefer Cloud Monitoring custom dashboards for arbitrary aggregations
Rather than artificially bloating Canonical Services into larger scopes for aggregate data, use Cloud Monitoring dashboards to create higher-level views of multiple logical services at once.
Canonical Services are meant to be long-lived
Outside of development, exploration, and testing use cases, Canonical Services
should represent global, high-level logical services. These services are
slow-changing and tend to be long-lived. This longevity includes not changing
service names. While you can change their names, doing so impacts metrics, SLOs,
and logs. However, you can freely adjust the
Display name field without
A new Canonical Service often represents new or updated software while a Canonical Service going away often represents a service deprecation. Your ability to see the historical performance of your service, plan, and project capacity all depend on maintaining a single notion of that service in Anthos Service Mesh for the duration of its life.
Note that this contrasts to lower-level resources like VM instances, Kubernetes Pods/Deployments, and even whole clusters, which often come and go as part of updating and maintaining production systems.