Add an api convention doc (#385)

Signed-off-by: Jian Qiu <[email protected]>

Author: qiujian16

docs/api-conventions.md

@@ -0,0 +1,103 @@
+# OCM API Conventions
+
+OCM APIs follow the [Kubernetes API conventions](https://github.com/kubernetes/community/blob/master/contributors/devel/sig-architecture/api-conventions.md)
+with some additional guidlines inspired from [OpenShift API Convention](https://github.com/openshift/enhancements/blob/master/dev-guide/api-conventions.md).
+
+## API Guidance
+
+### Configuration vs Workload APIs
+
+A configuration API is one that is typically cluster-scoped, a singleton within the cluster, and managed by a cluster
+administrator only. Operator APIs such as `ClusterManager`, `Klusterlet` are configuration APIs.
+
+A workload API typically is namespaced and is used by end users of the cluster. Most of the APIs, such as
+`ManagedCluster`, `ManifestWork`, and `Placement` are workload APIs.
+
+#### Defaulting
+
+In workload APIs, we typically default fields on create (or update) when the field isn't set.
+This has the effect that changing the default value in the API does not change the value for objects that have
+previously been created.
+
+This has the implication that you cannot change the behaviour of a default value once the API is defined as that would
+cause the same object to result in different behavior for different versions of the API, which would surprise users and
+compromise portability.
+
+To change the default behaviour could constitute a breaking change and disrupt the end user's workload;
+the behaviour must remain consistent through the lifetime of the resource.
+This also means that defaults cannot be changed without a breaking change to the API.
+If a user were to delete their workload API resource and recreate it, the behaviour should remain the same.
+
+With configuration APIs, we typically default fields within the controller and not within the API.
+
+Typically, optional fields on configuration APIs contain a statement within their Godoc to describe
+the default behaviour when they are omitted, along with a note that this is subject to change over time.
+
+#### Pointers
+
+In configuration APIs specifically, we advise to avoid making fields pointers unless there is an absolute need to do so.
+An absolute need being the need to distinguish between the zero value and a nil value.
+
+Using pointers makes writing code to interact with the API harder and more error prone, and it also harms the
+discoverability of the API.
+
+##### Pointers to structs
+
+An exception to this rule is when using a pointer to a struct in combination with an API validation that requires the
+field to be unset.
+
+The JSON tag `omitempty` is not compatible with struct references. Meaning any struct will always, when empty, be
+marshalled to `{}`. If the struct **must** genuinely be omitted, it must be a pointer.
+
+### Discriminated Unions
+
+A discriminated union is a structure within the API that closely resembles a union type.
+A number of fields exist within the structure and we are expecting the user to configure precisely one of
+the fields.
+
+In particular, for a discriminated union, an extra field exists which allows the user to declaratively
+state which of particular fields they are configuring.
+
+We use discriminated unions in Kubernetes APIs so that we force the user to make a choice.
+We do not want our code to guess what the user meant, they should tell us which of the choices they made
+using the discriminant.
+
+Important to note:
+* All structs within the union **MUST** be pointers
+* All structs within the union **MUST** be optional
+* The discriminant should be required
+* The discriminant **MUST** be a string (or string alias) type
+* Discriminant values should be PascalCase and should be equivalent to the camelCase field name (json tag) of one member of the union
+* Empty union members (discriminant values without a paired union member) are also permitted
+
+### No annotation-based APIs
+
+Do not use annotations for extending an API. Annotations may seem as a good candidate for introducing experimental/new
+API. Nevertheless, migration from annotations to formal schema usually never happens as it requires breaking changes
+in customer deployments.
+
+1. Validation does not always come with definition. User set values can be too broad and hard to limit later on.
+2. Lack of discoverability. There's no pre-existing schema that can be published.
+3. Validation is limited. Certain kinds of validators aren't allowed on annotations, so hooks are more frequently used instead.
+4. Hard to extend. An annotation value (a string) can not be extended with additional fields under a parent.
+5. Unclear versioning. Annotation keys can omit versioning. Or, there are multiple ways to specify a version.
+6. Users can "squat" on annotations by adding an unvalidated annotation value for a key that is used in a future version.
+
+### Use JSON Field Names in Godoc
+
+Ensure that the godoc for a field name matches the JSON name, not the Go name,
+in Go definitions for API objects.  In particular, this means that the godoc for
+field names should use an initial lower-case letter.
+
+### Use Resource Name Rather Than Kind in Object References
+
+Use resource names rather than kinds for object references.
+
+### Do not use Boolean fields
+
+Many ideas start as a Boolean value, e.g. `FooEnabled: true|false`, but often evolve into needing 3, 4, or even more
+states at some point during the API's lifetime.
+As a Boolean value can only ever have 2 or in some cases 3 values (`true`, `false`, `omitted` when a pointer), we have
+seen examples in which API authors have later added additional fields, paired with a Boolean field, that are only
+meaningful when the original field has a certain state. This makes it confusing for an end user as they have to be
+aware that the field they are trying to use only has an effect in certain circumstances.