While attributes prove useful for cross cutting concerns such as authorization or logging, they can be misused. Attributes should act as metadata, providing no direct behaviour. Failing to do so will make DI, testability and composition very difficult.
These flaws are especially true for validation. Despite all input requiring validation, the manner in which validation is performed is dependent on the entry point to the code. Context matters.
Consider order information that requires a billing address and by definition, its children to be populated. An attribute works a treat here in this simple case.
A problem arises if you only want the billing address validation to activate if the billing address and delivery address differ.
Complexity quickly starts to take over. With a more fully featured example attributes can start to overwhelm the class. This example becomes worse if the validation is required to be performed by a third party library or service. Finding a hook to integrate becomes troublesome.
Avoid attributes for validation in all but the simplest scenarios. Even simple scenarios lead to some churn if you do decide to switch. My personal preference is to now avoid attributes all together, instead opting to use a validation service.
The obvious downside to this is approach is the appearance of more code. While this is true, composed object graphs can benefit from the ability of reuse. Additionally in the case of attributes some degree of testing is required. These usually fall into the category of asserting the presence of attributes on properties which is far from ideal. The use of validation services do not suffer this problem. Internally the implementation can be switched, altered or refactored without fear of breaking any tests.
The RootValidator is a composite of zero or more actual validators. Each validator can be specific to a particular task. The only requirement being the interface must be the parent object. This is to ensure the context is not lost when making decisions. The actual interface in this case could be made to use generic types if required. The ValidationResults are a simple value type representing an aggregation of validation failures. This could be extended or modified for further enhancements.
- Composition makes it possible to provide multiple validators that all do one thing well.
- Testing is much easy as you can test each validator in isolation.
- Null validators provides easier higher level testing as you can provide a no-op validator. Removing the need to build up complex object graphs for other test cases.
- Developers can follow, debug and understand simple conditional logic more so than framework specific metadata.
- Open to extension and additions such as third party code.
- Services never lose context which allows easy runtime decisions to be made.