This document summarizes the areas of focus for the next release (v2). These cover improvements and new features that introduce non-backwards compatible changes.
The generated context data structure contains both the request and response state. While this makes it convenient to write controller code it does not make sense to pass these objects all the way down in all layers. It also creates hidden dependencies.
The proposal is to break up the generated context into two data structures, one that contains the request state and the other the response state. Concretely today the following design:
Action("update", func() {
Routing(
PATCH("/:bottleID"),
)
Params(func() {
Param("bottleID", Integer)
})
Payload(BottlePayload)
Response(NoContent)
Response(NotFound)
})
produces the following context data structure:
// UpdateBottleContext provides the bottle update action context.
type UpdateBottleContext struct {
context.Context
*goa.ResponseData
*goa.RequestData
AccountID int
BottleID int
Payload *UpdateBottlePayload
service *goa.Service
}
This proposal would break it down into:
// UpdateBottleRequest provides the bottle update action request context.
type UpdateBottleContext struct {
*goa.RequestData
AccountID int
BottleID int
Payload *UpdateBottlePayload
}
// UpdateBottleResponse provides the bottle update action response context.
type UpdateBottleResponse struct {
*goa.ResponseData
service *goa.Service
}
The response functions would move from the UpdateBottleContext
struct to the
UpdateBottleResponse
struct.
Both the response and request state structs would be given to the controller action function. Today the design above generates:
Update(*UpdateBottleContext) error
Under this proposal the generated code would be:
Update(ctx context.Context, resp *UpdateBottleResponse, req *UpdateBottleRequest) error
Move from a model using structs to implement goa errors and check for them to a model using checking for behavior.
Consider changing the goa request handler signature from:
func (context.Context, http.ResponseWriter, *http.Request) error
to the standard:
func (context.Context, http.ResponseWriter, *http.Request)
And keep the error in the response struct instead. This means tweaking how error handling is done so that the error handler middleware knows where to look for errors.
The first planned improvement made to code generation is the ability for arbitrary plugins to hook
into the generated code of another generator. The main scenario that this should enable is the
ability to inject code in the built-in generator outputs (in particular gen_app
). This would make
it possible to move the security and CORS generation in plugins and help keep the main generator
streamlined. Details on how this is enabled TBD. One possibility is to cater specifically to the use
case above and have the gen_app
generator expose explicit hooks. Would be nice to come up with a
more generic approach though.
Make it easier to invoke plugins on the command line. One possibility could be to have a hosted plugin
registry that goagen could look up where plugins could register with a simple moniker. This registry
could also provide discoverability of plugins (via something like goagen plugins
). The registry
would only contain metadata (moniker, description, author, last update data etc.) and point to the
actual plugin package.
Generators start from the design data structures then massage that data into a shape suitable for
code generation. Today this is done in a ad-hoc way - each generator having its own strategy. For
example the gen_app
generator uses Template
data structures except for test code generation that
use a different kind of structs. Other generators rely on functions called at generation time to
perform the transformations which makes it hard to follow.
The proposal is to implement a standard strategy of first creating explicit structs representing the intermediary representation used by the code generation templates. This would help clarify what the templates expect and keep that up-to-date as the definitions evolve. This should be done in a way that doesn't force generators to use it (some simple generators may not have a need for such a IR) but it should encourage it. For example there could be an additional optional interface that generators could implement that would get called by the generation engine.
Move DateTime
and UUID
to formats on the String
type. This is to keep the set of basic types
small and remain consistent with how Integer
support formats for various underlying Go data types
(see below). This also will allow removing a lot of cruft that supporting these types introduced.
From the need to "alias" them to string in code generators to the special cases they introduce with
gopherjs support.
Look into generating .proto files from the design and invoke protoc
on them during code
generation. Integrate the generated code with the generated data structures.
This also requires the ability to specify the bit length of numerical values. The proposal here
would be to add support for the Format
DSL to the Integer
and Number
primitive types. The
format would indicate bitness and whether the generated integer should be signed.
Try to remove the signers from the client package and instead make it possible to integrate with 3rd party client packages. Make the client function generate http.Request to enable that integration.