To [Decorate] or Not to [Decorate]
Attributes are a very useful feature of many languages but I personally feel that they are used too often and in wrong circumstances. They’re primary shortcoming is the fact that attributes are set up at compile time. During code execution they can only be read but not modified.
How I use attributes
In Argolis I created a number of attributes to decorate classes which I want to expose in my Hydra API. I also reuse some attributes from Newtonsoft.Json and core libraries.
1 2 3 4 5 6 7 8 9 10 11
This is used to produce a machine-readable description of a supported class on the server, which ensures that the client is always using the latest API.
When I don’t use attributes
The above usage seems legit. After all this information is static: the identifiers don’t change at runtime, neither does the general contract.
The next piece of my API Documentation puzzle are operations on a resource. In the simplest form they can define whether
given a resource
/users/tpluscode it is possible to perform
However, unlike the contract, resource’s behaviour is likely to change depending on its state (eg. it shouldn’t be possible to delete an Issue in progress) or the current authorized user (eg. only administrators and owners should be allowed to delete an issue). With this requirement in mind I decided that attributes are not a good choice, because I would have to somehow bind them with contextual behaviour. That behaviour would live in another type, be passed to the attribute instance and somehow have to be combined at runtime. (Forgive the hideous class name )
1 2 3 4
I don’t like the code above, because it pollutes the static nature of attributes. I fear this solution would also lead to coupling with a higher layer framework - Nancy in my case. Instead I created a basic interface, which is used to gather information about the available operations and can work nicely with dependency injection inside Nancy or Web API. In addition I prepared a simple base class, where operation can be wired up and some per-request context injected by the DI container.
1 2 3 4 5 6 7 8 9 10 11 12 13 14
[Decorate] or not to
I must say I feel quite dirty nonetheless, because by mixing to solution to complementary features of my library I caused some sort of duality - part of the API is to use attributes and part is to implement a base class, which the infrastructure discovers at runtime.
I’m on edge here. It’s like mixing NHibernate attributes and Fluent mappings in the same project. I’d like to hear if people do also find it a smell or is it acceptable.