Skip to content

Release v5.0.0

Compare
Choose a tag to compare
@System-Glitch System-Glitch released this 11 Jun 14:55
· 28 commits to master since this release
6cb4cf0

Introduction

Goyave v5 has been in the oven for over two years. With the first production applications deployed, a ton of issues and flaws revealed themselves, preventing real long-lived projects from cleanly evolving. The initial goals of the framework started to weaken the more the applications were developed and changed: the strong basis it promised to provide wasn't actually that strong. From this invaluable experience, I decided to go back to the drawing board and redesign the framework.

This new major version is an almost entire rewrite of the framework. In order to limit the frequency of releases containing breaking changes, all accumulated ideas for reworks and improvements that would introduce those were grouped in v5. This new major version not only aims at fixing the outstanding issues and design flaws, but also improves on existing features by upgrading them to the latest available technology. Expect v5 to feel modern and to neatly integrate with all the new language features such as generics, file systems, structured logging, and more.

These release notes will be organized in categories. They will explain the overall change of direction for each area of the framework as well as shortly showcase the new or improved features. The list of changes may be incomplete as many features were rewritten entirely.

Motivations

Among the many aspects that needed to be reworked, some of them stood out and fueled the initial drive to rewrite the framework.

Dependency coupling

Every layer and components of a v4 application had strong dependency coupling. The HTTP, business and database layers were all mixed up together. The framework was systematically imposing a dependency to itself, direct or indirect. Its locked-up architecture, entirely based on globals was hindering the more business-oriented applications. Those struggled to detach their domains from the rest of the application, and encountered obstacles every time they needed to handle more complex business logic.

For example: to access the database, you were forced to use the framework, which was loaded from a configuration system that also was handled by the framework. This created a long chain of dependencies that was hard to separate from the rest, even more so when it came to writing tests.

On top of that, the all-global architecture required a ton of synchronization, which were detrimental to the overall performance of the application.

Locked architecture

All components of the framework were strongly linked together and quite opaque, despite an initial effort made to make the framework flexible and hackable. In the end, many non-elegant workarounds had to be made in real-world scenarios. This made it harder to adapt an application to the constraints often encountered by companies developing their ecosystem and trying to solve real-world issues that needed deeper access to the inner-workings.

The validation system was one of the biggest, if not the biggest, culprit. It was very inconvenient to use compared to the new one brought with v5. The hacks required to make some advanced field comparison or business-logic validation were very fragile and hard to maintain. This design made it impossible to re-use code, and forced the hassle of creating a new validator for every single specific use-case.

The framework was also a bit too reliant on magic in some aspects. Many functions were using weak typing (any) and reflection, or even string identifiers, all for the sake of conciseness. But this came at a cost: no compile-time checks, hard code navigation, no completion, the need to rely on documentation all the time, etc. In the end, by trying to be concise for a better DX (developer experience), the framework sacrificed code cleanliness, reliability, readability, maintainability and actually ruined its DX this way.

Testing

All these issues accumulate and become a huge pain the moment you start trying to add tests to your project. They were very difficult to write, which is the exact opposite of what you want. Tests should be painless and as easy to read and maintain as possible. Also suffering from the locked architecture and mandatory dependency to the framework, they couldn't even be run in parallel. Almost nothing could be mocked because of the dependency coupling. This was in turn forcing you to use the database for your tests, which made tests even slower and complicated.

In short, the design of the framework prior to v5 treated tests as an after-thought despite how important they are.

Streamlining

There were many smaller and non-blocking issues as well. Together they made the entire development flow awkward by moments, like there was a missing piece to make the whole process fluid from start to finish.

The first one was the relation between the user-sent data and the internal data. Making use of the data sent by users was inconvenient and also quite unsafe, requiring a ton of type assertions and map navigation. This is very subpar compared to most Go applications which use structures. It also caused problems when interacting with the database.

Another issue was error handling and debugging. The framework was relying too much on panic, which is not a very sane nor idiomatic way to handle errors. Although it allowed for a pretty good debugging experience for developers with more precise stacktraces, a much better solution was possible. This solution wouldn't compromise on code quality for the sake of DX, again.

The last major one was the low interoperability and control, notably with the lack of access over the context API, or some missing configuration options and settings.

Philosophy

The overall philosophy of the framework stays the same. Goyave remains an opinionated framework, focused on DX (developer experience) by enabling quick and enjoyable development by being expressive, reliable and complete. The goal is to make business logic development as painless as possible by handling as many recurring things for developers as possible so they can focus on implementing what actually creates value for their company.

New direction

However, there are some important evolutions in the general direction of the framework:

  • v5 takes architecture and utilities one step further. At first glance, a typical application will be less simple to grasp. Therefore, the "progessive" nature of the framework is no more.
  • Despite its opinionated nature, the new design tries to be as open as possible. The framework should not be a black box and developers should have no trouble making it their own effortlessly.
  • The design should focus its efforts on creating tools that simplify the reduction or elimination of the dependencies between application layers and business layers.
  • The design and development of the framework now assumes full focus on catering to businesses and medium to large projects.
  • The development of the framework will now consider real-world use-cases and scenarios with more importance. No "quick and easy" solutions, tools or fancy gadgets will be developed from now on.
  • The framework and documentation will expand even more on architecture recommendations and good practices to cover more questions developers could have when working on their project. The goal of an opinionated framework is to save time to its users, while giving them all the tools they need to do produce the best possible quality software.
  • The open-source process has been revised to make contributions easier. Gathering and taking care of a strong community has always been very important. More efforts will be made in this direction.

Architecture

The framework goes even further than before when it comes to architecture recommendations for your projects. The goal is to answer more of the questions any team will inevitably encounter when setting up their project. By providing tools that will works seemlessly with this architecture, the frameworks aims at saving you a lot of time and make the development process as fluid as possible.

A project build with Goyave v5 will be split in three distinct layers:

  • Presentation: HTTP/REST layer, it's your application's facade
  • Domain/Business: contains services
  • Data: interacts with the database with repositories and contains the models

Each layer doesn't directly depend on the others because they define interfaces representing their own needs. The following chart describes the usual flow of a request into a Goyave application.

Architecture overview

This architecture has several advantages:

  • Good separation of concerns and no direct dependency
  • Easily testable
  • The data layer doesn't leak into the business layer even if there are transactions involved
  • Lowers the risk of exposing information that is not meant to be public
  • Easily readable, explorable and maintainable

In Goyave v5, nothing is global anymore. The costly need for goroutine synchronization is eliminated. The overall design is now interface-focused. On top of that, the entire framework now takes full advantage of the standard context API and encourages it use.

Components

Because nothing is global, a mechanism is necessary so the server's essential resources (such as the configuration, logger, etc) can be distributed to every component of the server. This mechanism is actually called Components, and described by the interface goyave.Composable. Most structures in the presentation layer actually are Goyave components: controllers, middleware, validators, etc.

Server

Changing all global resources to non-global required a central element that would hold them all. A server contains all the base resources for an application: the HTTP server, the router, configuration, language, the database pool, services, etc. It is the parent of all components. Previous versions were a bit too focused on how short the code could be. It was detrimental to its flexibility. A typical main() function for v5 will be longer than before but will offer more ways for developers to tweak and configure their application as they see fit, without adding a lot of complexity.

  • A server is created with goyave.New(options). The server doesn't start listening on the network right away, which is different from the previous version's goyave.Start().
  • A server starts listening and serving requests only when server.Start() is called.
  • Now that nothing is global anymore, multiple servers can be created simultaneously. This use-case, outside of tests, should be marginal.
  • goyave.Options is a structure containing new settings allowing to use manually loaded config, a custom logger, and more. It also allows to set options on the underlying net/http.Server that were previously not available, such as ConnState(), BaseContext(), ConnContext() and MaxHeaderBytes.
  • The server's concurrency safety and synchronization has been improved.
  • The underlying net/http.Server's logger is now unified with the Goyave logger.
  • The server instance is always injected as a context value in the server's base context and can be retrieved with goyave.ServerFromContext(ctx).
  • Global
  • The server has a service container that holds services and allows controllers to easily inject them as dependencies. Goyave doesn't use anything complex for dependency injection, only native Go and no code generation.
  • Services are initialized and registered in main.go.
  • server.Host() and server.Port() new methods return the hostname and port the server is running on.
  • BaseURL() and ProxyBaseURL() are now methods of *Server.
  • Startup hooks and shutdown hooks now take a *Server as parameter.
  • Startup hooks are not execued if the server fails to start before their goroutine started.
  • Startup hooks now all share the same goroutine and are executed in the order of registration. Each startup hooks was using their own goroutine before.
  • The shutdown signal hook is now optional. It can be added with server.RegisterSignalHook().
  • Routes are registered using server.RegisterRoutes() instead of passing the main route registrer on Start().
  • The main router can be retrieved with server.Router().
  • server.Stop() doesn't attempt to stop the server a second time if it was already stopped.
  • server.Stop() won't attempt to close the signal channel again if the server has already been stopped. This method can this be called several times safely now.

File systems

Go 1.16 introduced the io/fs package as well as support for embedded files and directories. These additions add a number of important benefits ranging from distribution packaging to storage abstraction. As v5 was already going to contain many breaking changes, the opportunity to integrate these new systems into the framework was taken. Every system or feature interacting with files in any way should now use file system interfaces. This allows developer maximum flexibility when working with files and makes it easier to write tests.

Thanks to file systems, static resources can be embedded into the compiled executable and external resources or storage can be seemlessly used.

  • New interfaces:
    • fsutil.FS: combines fs.ReadDirFS and fs.StatFS. This is used to require versatile read-only file systems.
    • fsutil.WorkingDirFS: a file system that has a working directory for relative paths, get by the Getwd() method.
    • fsutil.MkdirFS: a writable file system able to create directories.
    • fsutil.WritableFS: a writable file system with an OpenFile() method.
    • fsutil.RemoveFS: a writable file system supporting file or directory deletion.
  • osfs.FS is a complete implementation of the above interfaces for the local OS file system.
  • fsutil.Embed is a wrapper around fs.ReadDirFS used to enrich the embed.FS file system so they also implement fs.StatFS and Sub(). This is useful for serving embedded static resources, loading embedded language files, etc.
  • fsutil.GetMIMEType() now takes a fs.FS as parameter and returns an error instead of panicking.
  • fsutil.FileExists() and fsutil.IsDirectory() now take a fs.StatFS as parameter.
  • fsutil.File.Data was removed. You now have to open the File.Header yourself.
  • fsutil.File.Save() now takes a fsutil.WritableFS as parameter and returns an error instead of panicking.
  • fsutil.ParseMultipartFiles() now takes a []*multipart.FileHeader instead of a *http.Request.
  • By default, configuration, language files and JWT keys are loaded with the osfs.FS, but options are available in all these features if you want to use a different file system.
  • fsutil.Delete(path) was removed. Use osfs.FS.Remove() instead.

Configuration

Apart from moving from global to scoped in a structure, the configuration system didn't receive a lot of changes. Many configuration entries and defaults changed though.

  • Configuration is now a struct: *config.Config.
  • Configuration loading functions now take a file system as parameter. Configuration can be embedded if needed.
  • Configuration is not protected for race conditions anymore for better performance. Don't modify the config after the server started. config.Set() should only be used right after the config was loaded or inside tests.
  • config.LoadDefault() new function loads the default configuration without.
  • A manually loaded configuration can be used in the server Options with the options.Config field.
  • server.protocol, server.httpsPort and server.tls were removed: protocol is only http as TLS/HTTPS support has been removed because Goyave applications are most of the time deployed behind a proxy.
  • server.timeout has been split: server.writeTimeout, server.readTimeout, server.idleTimeout, server.readHeaderTimeout, server.websocketCloseTimeout.
  • server.maintenance was removed.
  • database entries do not have a default value anymore. They were previously using default values for MySQL.
  • New entries database.defaultReadQueryTimeout and database.defaultWriteQueryTimeout add a timeout mechanism to your database operations. If you have long queries, increase their values. Set to 0 to disable the timeouts.
  • auth.jwt.rsa.password was removed.
  • Setting server.port to 0 will now assign an automatically chosen port that is available. The actual port used can be retrieved with server.Host() or server.Port().
  • Slices of int, float64, bool and strings are now correctly supported by the configuration system.

Routing

Without dramatically changing how routing works, the simple addition of metadata assigned to routers and routes opened up a lot of possibilities and helped smooth out the route definition. It is now much easier to control middleware settings with the greatest granularity. Without this feature in previous versions, developers were limited in their API design and sometimes had to either sacrifice the consistency of their route tree, or duplicate a ton of code for fine-grained control.

  • Routes and routers now have Meta. Metadata are string-identified and can be inherited from parents thanks to the LookupMeta() method.
    • Metadata can be defined by directly accessing the Meta map, or using SetMeta() / RemoveMeta().
    • Metadata is now used by many middleware to store CORS configuration, tell if a route should require authentication, etc.
    • There are two options both having their use-case:
      • Define the middleware settings in the component structure as fields. This is used for validation for example.
      • Define the middleware settings in the router or route metadata. This is more suited for global middleware and allows fine-grained control over settings for each individual route or group, while using the same middleware instance.
  • CORS options can now be applied on individual routes. It was previously only possible to define CORS settings at the router level.
  • The CORS middleware is now global so it can be used before a global auth middleware for example.
  • route.BuildProxyURL() new method builds a full URL pointing to this route using the proxy base URL.
  • route.GetParent() and router.GetParent() new methods returns the parent router.
  • route.GetValidationRules() was removed. Validation rules are now stored in struct fields of the validation middleware.
  • As body and query data were split, route.Validate() was replaced by route.ValidateBody() and route.ValidateQuery().
    • These two methods take a goyave.RuleSetFunc as parameter. Rule sets are now generated per-request.
  • Changed the subrouter matching so it doesn't work with prefix only. This will remove conflicts between two subrouters having a prefix starting with the same characters (e.g.: /test and /test-2 won't conflict anymore)
  • The router won't turn back and explore other branches if the subrouter matches but none of its routes do. This will fix some false matches.
  • Fixed / route defined at the main router being matched if a subrouter matches but none of its routes do and a trailing slash is present in the request URI.
  • The special "not found" and "method not allowed" routes are now named so you can identify them from middleware and status handlers.
  • Status handlers are now components instead of simple functions. They must implement the goyave.StatusHandler interface.
  • goyave.GetRoute() was removed. Use the new method router.GetRoute() instead. Route naming is global to a main router. If you can GetRoute() from a subrouter, it will be able to return routes registered in a parent router.
  • Routes can be registered by controllers if they implement the goyave.Registrer interface, using the new method router.Controller(). This is encouraged so route definition for a feature will be located in the same package as the feature itself. This makes things cleaner and helps group related code together.
  • router.Route() now takes a slice of methods instead of a pipe-separated string.
  • Static file serving now uses file systems. router.Static() doesn't take a variadic slice of middleware as last parameter and now returns a *Route. You can then apply middleware, a name, meta, etc to that route as usual.
  • request.Params becomes request.RouteParams.
  • request.Route() becomes request.Route.
  • The main route registrer now takes server as parameter: func Register(server *goyave.Server, router *goyave.Router).
  • Routes are registered using server.RegisterRoutes() instead of passing the main route registrer on server start.

Requests

  • request.ToStruct() was removed, use typeutil.Convert() instead.
  • request.Data is now any instead of map[string]any. You should use safe type assertions before use.
  • Query data is not in request.Data anymore, it is now split in request.Query.
  • The request's context is now accessible through request.Context(). The context can be replaced with request.WithContext().
  • request.URI() was renamed request.URL().
  • Request accessors such as Has(), String(), Numeric(), etc were all removed.
  • request.CORSOptions() was removed. You can access CORS options via the route meta: request.Route.Meta[goyave.MetaCORS].(*cors.Options).
  • request.Lang is now of type *lang.Language instead of string. See the localization section below for more details.
  • request.Now is a new field indicating the time at which the request was received (after routing).
  • request.Extra is now a map[any]any. You are encouraged to use empty struct keys instead of literals.
  • request.Params was renamed request.RouteParams.
  • request.Route is now exported. It is not a method anymore.
  • request.Body() new method returns the request's body reader.

Responses

  • goyave.NewResponse() is now exported to make testing easier.
  • response.HandleDatabaseError(db) was removed and replaced by response.WriteDBError(err)
    • WriteDBError() returns true if there is an error and that you should return. This is the opposite of HandleDatabaseError.
  • response.Error(), response.JSON(), response.String() etc do not return an error anymore.
  • response.Redirect() was removed. You can replace by http.Redirect(response, request.Request(), url, http.StatusPermanentRedirect).
  • Template rendering was removed. response.Render() and response.RenderHTML() are not available anymore.
  • response.GetError() now returns *errors.Error instead of any. See the error handling section below for more details.
  • response.GetStacktrace() was removed. You can now access the stacktrace from the error itself.
  • response.File() and response.Download() now take a file system as first parameter.

Validation

Validation is a complex and very broad topic. That's why every single major release since since the birth of the framework changed the system in one way or another. Previous changes were mostly incremental, but this time, v5's validation system is very different. This new system learned from all the mistakes of the previous one.

  • "Rules" are now called "Validators". They're not a simple function anymore, but an implementation of the validation.Validator interface. Validators are components. This new approach alos allows for completion, compile-time checks and better discoverability.
  • Rule sets are now slices instead of maps. This allows developers to control the order of validation as they want. The only exception is arrays, which are still validated recursively.
  • Before, the order was now guaranteed and field comparisons required a hacky ComparesFields option in order to work. This option doesn't exist anymore. Rules comparing fields now use a *walk.Path.
  • The root element of the data being validated can now be anything, not necessarily an object.
  • Some of the structures were renamed to make more sense considering the fact that the root element is not always an object anymore:
    • validation.Errors is now validation.FieldsErrors.
    • validation.FieldErrors is now validation.Errors.
  • validation.CurrentElement now works everywhere, not only on composed rule sets.
  • Rule sets are now generated per-request. They are not meant to be re-used or used concurrently.
  • Validated routes now expect a function returning a rule set. This function has access to the request, allowing it to dynamically return different rule sets.

Here is an example of a new rule set function:

func SomeRequest(_ *goyave.Request) v.RuleSet {
	return v.RuleSet{
		{Path: v.CurrentElement, Rules: v.List{v.Required(), v.Object()}},
		{Path: "string", Rules: v.List{v.Required(), v.String()}},
		{Path: "number", Rules: v.List{v.Required(), v.Float64(), v.Min(10)}},
	}
}
  • Query and body validation are now separated.
  • The validation error response body is slightly different:
    • Instead of validationError, the key is now error to be consistent with the rest of the error handlers.
    • The errors are now separated in two fields: "body" and "query".
  • When using composition, the validators inside the composed rule set will be executed relatively to the element. ctx.Data will not be equal to the root data but to the parent element linked to the composed rule set. This affects comparison rules, whose comparison paths will now be relative. This effectively allows to re-use rule sets everywhere easily, even if they involve field comparison.
  • Validation is not always executed last as it used to. It is now executed following the same rules of ordering as regular middleware: you can now chose when validation occurs in the middleware stack.
  • The PostValidationHooks were removed.
  • Numeric validators now let you pick the exact Go type you want. The new validators will automatically check that the input value fits inside the corresponding type. For float validators, it is ensured that there are no loss of precision.
    • integer becomes: Int(), Int8(), Int16(), Int32(), Int64(), Uint(), Uint8(), Uint16(), Uint32(), Uint64().
    • numeric becomes: Float32(), Float64().
  • Array() doesn't have type parameters anymore. To validate array elements, add a path entry matching the array elements.
  • Size() validator and its derivatives such as Min(), Max(), etc now also work with objects and will validate its number of keys.
  • The new validation.Trim() validator trims strings.
  • The email validator now parses the email address with the net/mail standard package instead of using a regex.
  • Fixed a reflect error in the array validator if the first element of the array is invalid (e.g.: nil).
  • Removed the suffix "[]" from n-dimensional array elements field name in error messages.
  • The new validation.RequiredIf() validator allows to conditionally mark a field as required.
  • Type-dependent rules now support object type too.
  • The convention for the file in which the validation rules changed from request.go to validation.go.
  • Validators Exists()/Unique() now take a Gorm scope as parameter, letting you defined the table and condition yourself. You can preferably use a scope returned by a repository.
  • New validatorss ExistArray() and UniqueArray() efficiently validate an entire array against a database.
  • Arrays being supported as root element, the [] path is now valid.
  • It is now possible to validate all properties of an object without knowing their names using a wildcard * in the path (e.g: object.*).

Manual validation

  • validation.Validate and validation.ValidateWithExtra becomes validation.Validate(opts).
  • The validation.Options contains several options, as well as external dependencies such as the language, the database, the config, etc. Those will be passed to the validators so they can access them just like any regular component.
  • isJSON becomes ConvertSingleValueArrays, which does the same but the logic is different so the bool value will be the opposite.
  • This function now also returns a slice of errors. Those are not validation errors, they are actual execution errors.

Custom validators

  • Custom validators are implementations of the validation.Validator interface. They compose with validation.BaseValidator for the base values, but can override methods IsType() and IsTypeDependent().
  • ctx.Data is now any instead of map[string]any.
  • ctx.Extra["request"] becomes ctx.Extra[validation.ExtraRequest{}]
  • ctx.Extra is not scoped to the current validator only anymore. The same reference given in Options.Extra is shared between all validators.
  • Validator instances are not meant for re-use. Make sure to not persist any data inside a validator.
  • ctx.Valid() becomes ctx.Invalid (the boolean value is thus inverted).
  • ctx.Rule was removed. ctx.Rule.Params becomes validator struct fields. The values are passed to the validator constructor.
  • Don't panic inside validators. If you need to report an error, use ctx.AddError(). This error will be brought up in the returned values of validation.Validate().
  • Validation placeholders have been removed. Each validator can now dynamically return their placeholders by implementing MessagePlaceholders(*validation.Context) []string.
  • The :field placeholder remains unchanged.
  • The validation.Context now allow nested validation and batch array validation:
    • ctx.AddArrayElementValidationErrors() marks a child element to the field currently under validation as invalid.
    • ctx.AddValidationError() adds a validation error message at the given path.
    • ctx.AddValidationErrors() adds a *validation.Errors entire object to be merged into the errors bag of the parent validation.
    • ctx.Path() returns the exact path to the element under validation.
  • validation.GetFieldType() now returns constants. The bool field type is now supported too.

Structure conversion and mapping

The use of DTO (Data Transfer Object) and model mapping is now encouraged. This mechanism will help separate the data layer, reduce the risk of sensitive information leaks, and ease the communication between the different layers of your application. Working with strongly-typed structures

In short:

  • Services will receive and return DTOs only.
  • Repositories will receive and return models only.
  • There will be one DTO defined for each operation. (eg. CreateUser, UpdateUser, etc).
  • Models should not be partially updated to avoid temporal inconsistencies. Each update should retrieve the entire record, use model mapping to change the necessary fields, and save the model.

Changes:

  • request.ToStruct() was removed.
  • The reflectutil package was removed.
  • typeutil.ToFloat64() and typeutil.ToString() were removed.
  • typeutil.Convert() new function converts anything into the desired type using JSON marshaling and unmarshaling. This allows raw validated request data to be safely converted to a DTO, or a model to be converted to a DTO.
  • typeutil.MustConvert() does the same but panics in case of error.
  • typeutil.Copy() new function deep-copies a DTO's non-zero fields into the given model. It is used for model mapping.

Examples:

// Raw request data to DTO
requestDTO := typeutil.MustConvert[*dto.RegisterUser](request.Data)

// Model to DTO
userDTO := typeutil.MustConvert[*dto.User](user)

// Model mapping
userModel = typeutil.Copy(userModel, userDTO)
  • The new type typeutil.Undefined is a utility type wrapping a generic value that can be used to differentiate between the absence of a field and its zero value, without using pointers. This handy type will help you easily handle optional fields in requests, because "undefined" is different from nil, which is also different from the zero-value of a type. All fields that are not required in a validated request should use this type in their corresponding DTO. A field that wasn't provided by the user in their request will not be copied by model mapping.
type UpdateUser struct {
	Email typeutil.Undefined[string] `json:"email"`
	Name  typeutil.Undefined[string] `json:"name"`
}

Authentication

Authentication was one of those features in Goyave that was using a bit too much magic and unnecessary reflection. After re-designing the architecture with layer separation in mind, it became clear that authentication wasn't meeting the new criteria. Indeed, authentication is part of the HTTP layer, but it was using the database, and made it impossible to detach this logic and move it to a repository. The package was therefore redesigned to respect the new requirements.

  • auth.Authenticator and the authentication middleware now take a generic parameter representing the authenticated user's DTO.
  • Authentication should now work with a user DTO instead of the user model. This way the model doesn't leak into other layers.
  • Authenticators now depend on a service implementing the UserService[T] interface. This service should implement a method that fetches the user by it's "username". This allows moving the database logic to repositories and keeping the single responsibility of authenticators correctly scoped to their own layer. Struct tags auth:"username" and auth:"password" are not used anymore and can be removed.
  • Authenticators' Authenticate() method now return a *T instead of taking a model as parameter and updating it in place.
  • auth.FindColumns was removed.
  • JWT generation was moved inside a new service *auth.JWTService, identified by the name auth.JWTServiceName.
    • This service is able to use file systems for signing keys.
    • This service is automatically registered if not already registered when using the JWTAuthenticator or JWTController.
    • Key storage is now exported so you can use it as well.
  • The JWTAuthenticator now stores a valid token's claims in request.Extra[auth.ExtraJWTClaims{}].
  • The authentication middleware now checks for the auth.MetaAuth route meta to know if the route matched requires authentication. This means the authentication middleware can be used as a global middleware and you have easy fine control over which route or subrouter requires or doesn't require authentication.
  • Support for password-protected RSA keys has been dropped.
  • auth.JWTController:
    • The controller now has a generic parameter representing the authenticated user's DTO.
    • This parameter is also used in auth.TokenFunc instead of a user of type any.
    • UsernameField and PasswordField were renamed to UsernameRequestField and PasswordRequestField respectively.
    • A new field named PasswordField defines the name of the T's struct field that holds the user's hashed password.
    • auth.JWTRoutes() was removed. The auth.JWTController implements RegisterRoutes() to make the login route register automatically.
    • The response body in case of invalid credentials now has a key "error" instead of "validationError".

Database

Database is an important component that can be used in many places. It was important to make sure this dependency was not global anymore. It has been decided to remove some of its features to incentivise developers to use better and more sustainable solutions.

  • database.GetConnection(), database.Conn(), database.Close() and database.SetConnection() were removed.
  • database.Migrate() was removed. Using automatic migrations is now discouraged.
  • database.RegisterModel(), database.GetRegisteredModels(), ClearRegisteredModels() were removed. Registering models was only useful for auto-migrations and the test suite, which was also removed in favor of better solutions.
  • database.View was removed because it isn't of any use anymore after the testing changes.
  • Database initializers (database.AddInitializer() and database.ClearInitializers()) were removed.
  • database.New() creates a new database instance using the given *config.Config.
  • database.NewFromDialector() creates a new database instance using a custom dialector. This can be used when writing tests involving database mocking.
  • The Gorm logger now prints to the structured logger using the same format and colors as the rest of the logs.
  • database.TimeoutPlugin is a new plugin added by default. It defines a timeout for Gorm operations. It can be configured with the database.defaultReadQueryTimeout and database.defaultWriteQueryTimeout configuration entries. Using a timeout context for the query will override the default timeout.
  • It is now encouraged to use the request's context.Context for database queries execution.
  • database.Paginator:
    • Paginators now take a generic parameter representing the model to paginate and can be created with database.NewPaginator().
    • database.PaginatorDTO[T] is a new structure that can be used to write a pagination response. database.Paginator can be converted to a database.PaginatorDTO easily with typeutil.Convert().
    • paginator.Find() now returns an error instead of a *gorm.DB.
    • paginator.UpdatePageInfo() now returns an error.
    • The page info and records query are now executed together inside a transaction to avoid possible temporal inconsistencies.
  • The database-related validation rules have been moved to the validation package.

Session

The new session mechanism is an abstraction over a transaction system allowing services to define and control business transactions without directly interacting with the database. This system allows services to depend on multiple repositories and/or services, and call them seemlessly in the same transaction without having to develop them in a specific way and without worrying about side-effects.

Sessions can be mocked too. This will help you test those more complex cases where something fails during a business transaction.

Localization

The simple localization system in Goyave already works for most cases so it didn't went through a big overhaul. It was mostly refactored for easier use and greater openness.

  • The language system is now structured with two important elements:
    • *lang.Languages contains multiple loaded languages and defines a default language.
    • *lang.Language represents a single language and contains the translated text.
    • You can get a translated line directly from a *lang.Language or as before through *lang.Languages by specifying the language name.
  • Language loading methods now take a file system as parameter. This allows the language files to be embedded.
  • request.Lang is now a *lang.Language instead of a string. This means you can get a language line directly from the request language:
request.Lang.Get("custom-line")
  • The validation message keys for array elements were changed. Replace .array with .element.
  • Type-dependent validators can also support the object type now.
  • fields.json is now a map[string]string. There is no object with "name" nor "rules" anymore.
  • Improved language files unmarshal error messages.

Logging

Way before Goyave v5 was being designed, the question of better logs was already there for the framework. Instead of several basic loggers, a unified logger with levels would bring many benefits. At the time, many popular logging libraries existed but none of them aligned on a single interface. Go 1.21 introduced log/slog, a standard for structured logging. The choice of using this new standard and integrate it into the framework was made.

Thanks to structured logging, logs will be easier to read both in development and in production. The framework comes with a custom slog handler for development, which formats the logs in a human-readable way. In production, the logs will be output in the JSON format, which is very easy to parse by cloud service providers, making your logs easily readable and filterable.

All logs in Goyave v5 are now unified and can take full advantage of log levels (DEBUG, INFO, WARN, ERROR). This means that Gorm also uses the same logger as the rest of the application. The format will now be consistent across your entire application.

  • All logs are written to os.Stderr by default.
  • Logging is not global anymore. Logger instances are now passed around, mostly through components or dependencies. You also have the option to pass logger instances through the context.Context.
  • goyave.Logger, goyave.ErrLogger and goyave.AccessLogger were removed.
  • goyave.dev/goyave/v5/slog defines *slog.Logger, a wrapper around the standard *log/slog.Logger. This wrapper helps gracefully handle errors and print them with more details.
  • When debug mode (configuration app.debug) is disabled, the minimum log level is bumped to Info and the Gorm logger is disabled.
  • The common and combined access logs middleware were changed:
    • log.Formatter now takes a *log.Context instead of many parameters.
    • The formatter doesn't have access to the response object itself anymore, only the response status code.
    • log.Formatter now returns a message and a []slog.Attr slice instead of just a message.
    • The access log writer now writes to the structured logger at the Info level. The message will remain the same, but slog attributes will be added to make the log richer.
    • These attributes won't be printed in debug mode to avoid clutter.

Error handling

In previous versions, error handling was relying on panic. This had the advantage of centralizing error handling in the recovery middleware and getting precise stack traces. However, this approach wasn't very sane and not idiomatic. Going forward, the solution will be more idiomatic, easier to test and won't compromise on code quality, while improving the DX even more.

Developers are now encouraged to return errors all the way up the call stack. At the highest level, the HTTP handlers will report errors with response.Error(). Error handling will still be centralized in a status handler to make error reporting easier.

A new error wrapping mechanism was added. goyave.dev/goyave/v5/util/errors brings the *errors.Error type, which provide a convenient way to enrich errors with information useful for debugging and error reporting:

  • Caller frames are collected when the error happens. The stack traces will always precisely point to the exact origin of the error.
  • Any error reason can be used, be it a string, a struct, another error, map, []error, []any, etc.
  • Multiple errors can be wrapped into one single error. Nested errors are also supported.
  • *errors.Error is handled by the structured logger, which results in more detailed logs.

Error wrapping diagram

Request parsing

Request parsing was one of the rigid features that were convenient in simple cases, but hard to work around for advanced usage. In its redesign, the focus was put on allowing finer control over it.

  • Request parsing is not a core middleware automatically executed on all requests anymore.
  • This change allows more fine-grained control over request parsing. This middleware prevents upload streaming. Now that it can be removed, this use-case isn't blocked anymore.
  • The request's body reader is not reset anymore after the parse middleware is executed.
  • The max upload size can now be configured from the middleware (using MaxUploadSize) if you need to locally override the configuration entry server.maxUploadSize.
  • The parse middleware now accepts any type of root element if the request body is JSON. It previously only accepted objects.
router.GlobalMiddleware(&parse.Middleware{})
  • The request's parsed body and query are not grouped in the same place anymore. They are now respectively stored in request.Data and request.Query.
  • request.Data is now any. It previously always was an object (map[string]any).
  • request.Query is a map[string]any.

Compression

  • middleware.Gzip() was replaced by the new goyave.dev/goyave/v5/middleware/compress package.
  • The new compress.Middleware provides a generic basis for all types of compression and accepts multiple encoders. The encoder will be chosen depending on the request's Accept-Encoding header.
  • Out-of-the-box, only the gzip encoder is available. It is however very easy to implement a custom encoder for other compression algorithms.
compress := &compress.Middleware{
	Encoders: []compress.Encoder{
		&compress.Gzip{
			Level: gzip.BestCompression,
		},
	},
}
router.Middleware(compress)

Tests

As explained earlier in the release notes, tests were suffering from many flaws making them incredibly inconvenient, slow, hard to maintain. Proper unitary tests were not possible neither because of all the dependencies imposed by the previous versions of the framework.

  • Testing has been completely overhauled, and is now carefully considered in the design decisions.
  • Test utilities are designed with parallelism in mind for better efficiency.
  • Testing is now agnostic. You can use the testing library of your choice. You are not forced to use testify anymore.
  • The working directory is not set to the projects root directory anymore.
  • The GOYAVE_ENV environment variable is not set automatically anymore.
  • The goyave.TestSuite and other testing utilities were removed. The remaining ones were moved to the testutil package.
  • The new testutil package contains the new testing utilities.
    • You can easily create a test server with testutil.NewTestServer or testutil.NewTestServerWithOptions.
    • Tests servers are wrappers around *goyave.Server. They expose useful methods that allows you to test a request without starting the server and listening to a port or test a middleware. It can be passed around as the parent server for all your components.
    • When using a test server, the logs are automatically redirected to the test logs (t.Log()).
    • Test servers automatically close their database connection when their parent test ends.
  • A *goyave.Server can run in transaction mode with server.Transaction(). All the SQL queries will be executed from inside a transaction that can be rolled back. Tests involving the database should use this feature to isolate each test and avoid conflicts or race conditions at the database level.
  • The database can be mocked using server.ReplaceDB().
  • testutil.ReadJSONBody() new function helps reading and unmarshalling a test response body.
  • testutil.ToJSON() new function is a shortcut to marshal anything and create a reader from the result. It can be used for HTTP request tests.
  • Creating test requests and responses is easier and less restricted than before since more fields are exported and you don't require a test suite anymore.
  • Multipart forms for file uploads as well as testutil.CreateTestFiles now take a file system as parameter, allowing to use embedded files or mocked file systems.
  • Factories were updated.
    • Factories now take a generic parameter representing the type of the model they will generate;
    • The generator function is now expected to return a type matching the generic type of the factory instead of returning any.
    • As a result, type assertions are not required anymore with the results of Save() and Generate().
    • Overrides now use the same copying mechanism as model mapping instead of relying on mergo.
  • HTTP tests now simply use a http.Request generated with the standard httptest package instead of multiple custom methods (Get(), Post(), etc).
  • The new architecture coupled with the use of interfaces make it very easy to mock any part of your application for unit-testing.

Raw data exploration

Raw data exploration is mainly used in validation. The framework now allows any type of data to be sent as the root element of a request, so this mechanism had to be slightly changed to support that. On top of that, many methods were added to make exploration and comparison more convenient in validators, where they will be used more frequently with the revamp.

Raw data exploration using the goyave.dev/goyave/v5/util/walk package has received minor changes:

  • *Path.Name is now a *string instead of string to better handle elements that don't have a name (such as array elements).
  • Paths can now explore array root elements (paths such as [] or [].field).
  • walk.Path.Walk() callback now receives a *walk.Context instead of walk.Context
  • walk.Path.First() new method returns the *walk.Context for the first final element that matches the path.
  • walk.Path.Depth() new method returns the depth of the path.
  • walk.Path.Truncate() new method returns a clone of the n first steps of the path so the returned path's depth equals the given depth.
  • *walk.Path now implements a String() methods that returns a string representation of the path.
  • walk.Context's new method Break() indicates the walker to stop early.
  • walk.MustParse() new function returns a *walk.Path or panics if there is an error.
  • A wildcard * can now be used to explore all the properties of an object.

Websocket

The new architectural changes were an opportunity to make websockets more in line with how regular controllers would be implemented.

  • The connection close timeout is now configured by the server.websocketCloseTimeout entry.
  • websocket.Conn.CloseWithError() doesn't change the returned message to the error message if debug is enabled. The message will always be "Internal Server Error".
  • The connection close timeout is now shared between the close control message and the connection close handler. Previously, the timeout was reset after writing the close control message, effectively allowing double the value of the configured timeout.
  • Websockets now use controllers implementing websocket.Controller instead of lone handlers. Controllers must implement a method Serve(*websocket.Conn, *goyave.Request) error. This way, websocket handlers also benefit from the fact they are components.
  • Controllers can implement the websocket.Registrer interface to manually register the upgrade route if they want to.
  • Upgraders are now created with websocket.New(controller).
  • Options UpgradeErrorHandler, ErrorHandler, CheckOrigin, Headers are replaced with interfaces that can be implemented by the controller.

Filters

The filters library didn't allow decoupling of the HTTP layer and the data layer because of its dependency to the *goyave.Request. It was therefore impossible to move its uses to a repository, where they belongs. By creating a DTO for this specific use and changing the error handling, filters can now be properly integrated in the new architecture. They were also upgraded to take advantage of the generics.

  • filter.Settings, filter.Scope(), filter.ScopeUnpaginated() now take a generic parameter representing the model being filtered.
  • filter.Settings.CaseInsensitiveSort new option wraps the sort values in LOWER() if the column is a string, resulting in ORDER BY LOWER(column).
  • filter.Settings.DefaultSort new option allows to define a list of sorts to be used if the user request didn't contain any sort.
  • filter.Settings.Scope(), filter.Settings.ScopeUnpaginated(), filter.Scope(), filter.ScopeUnpaginated() now take a *filter.Request instead of a *goyave.Request. This request can be created from a HTTP query with the new function filter.NewRequest(query).
  • filter.Settings.Scope() and filter.Scope() now return an error instead of a *gorm.DB.
  • filter.Filter.Scope(), filter.Join.Scope(), filter.Sort.Scope() now take a filter.Blacklist instead of *filter.Settings.
  • filter.Validation is now a goyave.RuleSetFunc, and should be used with route.ValidateQuery().
  • The field fields in a request validated with filter.Validation will now always be a []string.
  • The page info and records SQL queries are now executed inside a transaction.
  • Validation error messages names had a "goyave-filter-" prefix added.

Miscellaneous

  • Minimum Go version will always be the second-to-latest version. At the time of release, this is 1.21.
  • The template project has been simplified as much as possible to removed all the unnecessary code. When you start a project, the first thing you do shouldn't be to remove things you don't need.
  • The util/sliceutil package was removed. Use samber/lo instead.
  • goyave.EnableMaintenance(), goyave.DisableMaintenance() and goyave.IsMaintenanceEnabled() were removed. Their use-case was rare. They were removed to reduce bloat. This kind of maintenance mode should be handled by the proxy, not the application.
  • utils.Map was removed.
  • The ratelimit package was removed. This implementation couldn't be used in a realistic production environment and didn't cover critical aspects of a real application.
  • Fixed panic status handler attempting to write to hijacked connections.
  • The recovery middleware now forces override of the HTTP status to 500. This prevents empty 200 responses if an error occurs before writing the body (such as trying to JSON marshal an unsupported type).
  • All tests were re-written not only to be more extensive, but also easier to read and maintain. The stability and reliability of the framework has been tested in production over a long period.
  • The framework now uses path.Join() instead of concatenation of paths.