apitech
/
grafana
mirror of https://github.com/grafana/grafana
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

Docs: Update backend architecture contributor documentation (#51172)

Browse Source 
Co-authored-by: Sofia Papagiannaki <1632407+papagian@users.noreply.github.com>
pull/51650/head
Emil Tullstedt 3 years ago committed by GitHub
parent d63ffa314e
commit 0e7a495829
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 375 additions and 96 deletions
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Show Stats Download Patch File Download Diff File
  1. 169
      contribute/architecture/backend/communication.md
  2. 81
      contribute/architecture/backend/errors.md
  3. 221
      contribute/architecture/backend/package-hierarchy.md

169
contribute/architecture/backend/communication.md
Unescape View File

@ -2,134 +2,131 @@
Grafana uses a _bus_ to pass messages between different parts of the application. All communication over the bus happens synchronously. Grafana uses a _bus_ to pass messages between different parts of the application. All communication over the bus happens synchronously.
> **Deprecated:** The bus has officially been deprecated, however, we're still using the command/query objects paradigms. ## Commands and queries
There are three types of messages: _events_, _commands_, and _queries_. Grafana structures arguments to [services](services.md) using a command/query
separation where commands are instructions for a mutation and queries retrieve
records from a service.
## Events Services should define their methods as `func[T, U any](ctx context.Context, args T) (U, error)`.
An event is something that happened in the past. Since an event has already happened, you can't change it. Instead, you can react to events by triggering additional application logic to be run, whenever they occur. Each function should take two arguments. First, a `context.Context` that
carries information about the tracing span, cancellation, and similar
runtime information that might be relevant to the call. Secondly, `T` is
a `struct` defined in the service's root package (see the instructions
for [package hierarchy](package-hierarchy.md)) that contains zero or
more arguments that can be passed to the method.
> Because they happened in the past, event names are written in past tense, such as `UserCreated`, and `OrgUpdated`. The return values is more flexible, and may consist of none, one, or two
values. If there are two values returned, the second value should be
### Subscribe to an event either an `bool` or `error` indicating the success or failure of the
call. The first value `U` carries a value of any exported type that
In order to react to an event, you first need to _subscribe_ to it. makes sense for the service.
To subscribe to an event, register an _event listener_ in the service's `Init` method:
```go Following is an example of an interface providing method signatures for
func (s *MyService) Init() error { some calls adhering to these guidelines:
s.bus.AddEventListener(s.UserCreated)
return nil
}
func (s *MyService) UserCreated(event *events.UserCreated) error { ```
// ... type Alphabetical interface {
// GetLetter returns either an error or letter.
GetLetter(context.Context, GetLetterQuery) (Letter, error)
// ListCachedLetters cannot fail, and doesn't return an error.
ListCachedLetters(context.Context, ListCachedLettersQuery) Letters
// DeleteLetter doesn't have any return values other than errors, so it
// returns only an error.
DeleteLetter(context.Contxt, DeleteLetterCommand) error
} }
``` ```
**Tip:** Browse the available events in the `events` package. > Because we request an operation to be performed, command are written in imperative mood, such as `CreateFolderCommand`, `GetDashboardQuery` and `DeletePlaylistCommand`.
### Publish an event The use of complex types for arguments in Go means a few different
things for us, it provides us with the equivalent of named parameters
from other languages, and it reduces the headache of figuring out which
argument is which that often occurs with three or more arguments.
If you want to let other parts of the application react to changes in a service, you can publish your own events: On the flip-side, it means that all input parameters are optional and
that it is up to the programmer to make sure that the zero value is
useful or at least safe for all fields and that while it's easy to add
another field, if that field must be set for the correct function of the
service that is not detectable at compile time.
```go ### Queries with Result fields
event := &events.StickersSentEvent {
UserID: "taylor",
Count: 1,
}
if err := s.bus.Publish(event); err != nil {
return err
}
```
## Commands Some queries have a Result field that is mutated and populated by the
method being called. This is a remainder from when the _bus_ was used
for sending commands and queries as well as for events.
A command is a request for an action to be taken. Unlike an event's fire-and-forget approach, a command can fail as it is handled. The handler will then return an error. All bus commands and queries had to implement the Go type
`func(ctx context.Context, msg interface{}) error`
and mutation of the `msg` variable or returning structured information in
`error` were the two most convenient ways to communicate with the caller.
> Because we request an operation to be performed, command are written in imperative mood, such as `CreateFolderCommand`, and `DeletePlaylistCommand`. All `Result` fields should be refactored so that they are returned from
the query method:
### Dispatch a command ```
type GetQuery struct {
To dispatch a command, pass the `context.Context` and object to the `DispatchCtx` method: Something int
```go Result ResultType
// context.Context from caller
ctx := req.Request.Context()
cmd := &models.SendStickersCommand {
UserID: "taylor",
Count: 1,
} }
if err := s.bus.DispatchCtx(ctx, cmd); err != nil {
if err == bus.ErrHandlerNotFound { func (s *Service) Get(ctx context.Context, cmd *GetQuery) error {
return nil // ...do something
} cmd.Result = result
return err return nil
} }
``` ```
> **Note:** `DispatchCtx` will return an error if no handler is registered for that command. should become
> **Note:** `Dispatch` currently exists and requires no `context.Context` to be provided, but it's strongly suggested to not use this since there's an ongoing refactoring to remove usage of non-context-aware functions/methods and use context.Context everywhere. ```
type GetQuery struct {
**Tip:** Browse the available commands in the `models` package. Something int
### Handle commands
Let other parts of the application dispatch commands to a service, by registering a _command handler_:
To handle a command, register a command handler in the `Init` function.
```go
func (s *MyService) Init() error {
s.bus.AddHandlerCtx(s.SendStickers)
return nil
} }
func (s *MyService) SendStickers(ctx context.Context, cmd *models.SendStickersCommand) error { func (s *Service) Get(ctx context.Context, cmd GetQuery) (ResultType, error) {
// ... // ...do something
return result, nil
} }
``` ```
> **Note:** The handler method may return an error if unable to complete the command. ## Events
> **Note:** `AddHandler` currently exists and requires no `context.Context` to be provided, but it's strongly suggested to not use this since there's an ongoing refactoring to remove usage of non-context-aware functions/methods and use context.Context everywhere. An event is something that happened in the past. Since an event has already happened, you can't change it. Instead, you can react to events by triggering additional application logic to be run, whenever they occur.
## Queries > Because they happened in the past, event names are written in past tense, such as `UserCreated`, and `OrgUpdated`.
A command handler can optionally populate the command sent to it. This pattern is commonly used to implement _queries_. ### Subscribe to an event
### Making a query In order to react to an event, you first need to _subscribe_ to it.
To make a query, dispatch the query instance just like you would a command. When the `DispatchCtx` method returns, the `Results` field contains the result of the query. To subscribe to an event, register an _event listener_ in the service's `Init` method:
```go ```go
// context.Context from caller func (s *MyService) Init() error {
ctx := req.Request.Context() s.bus.AddEventListener(s.UserCreated)
query := &models.FindDashboardQuery{ return nil
ID: "foo",
}
if err := bus.Dispatch(ctx, query); err != nil {
return err
} }
// The query now contains a result.
for _, item := range query.Results { func (s *MyService) UserCreated(event *events.UserCreated) error {
// ... // ...
} }
``` ```
> **Note:** `Dispatch` currently exists and requires no `context.Context` to be provided, but it's strongly suggested to not use this since there's an ongoing refactoring to remove usage of non-context-aware functions/methods and use context.Context everywhere. **Tip:** Browse the available events in the `events` package.
### Return query results ### Publish an event
To return results for a query, set any of the fields on the query argument before returning: If you want to let other parts of the application react to changes in a service, you can publish your own events:
```go ```go
func (s *MyService) FindDashboard(ctx context.Context, query *models.FindDashboardQuery) error { event := &events.StickersSentEvent {
// ... UserID: "taylor",
query.Result = dashboard Count: 1,
return nil }
if err := s.bus.Publish(event); err != nil {
return err
} }
``` ```

81
contribute/architecture/backend/errors.md
Unescape View File

@ -0,0 +1,81 @@
# Errors
Grafana introduced its own error type `github.com/grafana/grafana/pkg/util/errutil.Error`
in June 2022. It's built on top of the Go `error` interface extended to
contain all the information necessary by Grafana to handle errors in an
informative and safe way.
Previously, Grafana has passed around regular Go errors and have had to
rely on bespoke solutions in API handlers to communicate informative
messages to the end-user. With the new `errutil.Error`, the API handlers
can be slimmed as information about public messaging, structured data
related to the error, localization metadata, log level, HTTP status
code, and so forth are carried by the error.
## Basic use
### Declaring errors
For a service, declare the different categories of errors that may occur
from your service (this corresponds to what you might want to have
specific public error messages or their templates for) by globally
constructing variables using the `errutil.NewBase(status, messageID, opts...)`
function.
The status code loosely corresponds to HTTP status codes and provides a
default log level for errors to ensure that the request logging is
properly informing administrators about various errors occurring in
Grafana (e.g. `StatusBadRequest` is generally speaking not as relevant
as `StatusInternal`). All available status codes live in the `errutil`
package and have names starting with `Status`.
The messageID is constructed as `<servicename>.<error-identifier>` where
the `<servicename>` corresponds to the root service directory per
[the package hierarchy](package-hierarchy.md) and `<error-identifier>`
is a short identifier using dashes for word separation that identifies
the specific category of errors within the service.
To set a static message sent to the client when the error occurs, the
`errutil.WithPublicMessage(message string)` option may be appended to
the NewBase function call. For dynamic messages or more options, refer
to the `errutil` package's GoDocs.
Errors are then constructed using the `Base.Errorf` method, which
functions like the [fmt.Errorf](https://pkg.go.dev/fmt#Errorf) method
except that it creates an `errutil.Error`.
```go
package main
import (
"errors"
"github.com/grafana/grafana/pkg/util/errutil"
"example.org/thing"
)
var ErrBaseNotFound = errutil.NewBase(errutil.StatusNotFound, "main.not-found", errutil.WithPublicMessage("Thing not found"))
func Look(id int) (*Thing, error) {
t, err := thing.GetByID(id)
if errors.Is(err, thing.ErrNotFound) {
return nil, ErrBaseNotFound.Errorf("did not find thing with ID %d: %w", id, err)
}
return t, nil
}
```
Check out [errutil's GoDocs](https://pkg.go.dev/github.com/grafana/grafana@v0.0.0-20220621133844-0f4fc1290421/pkg/util/errutil)
for details on how to construct and use Grafana style errors.
### Handling errors in the API
API handlers use the `github.com/grafana/grafana/pkg/api/response.Err`
function to create responses based on `errutil.Error`s.
> **Note:** (@sakjur 2022-06) `response.Err` requires all errors to be
> `errutil.Error` or it'll be considered an internal server error.
> This is something that should be fixed in the near future to allow
> fallback behavior to make it possible to correctly handle Grafana
> style errors if they're present but allow fallback to a reasonable
> default otherwise.

221
contribute/architecture/backend/package-hierarchy.md
Unescape View File

@ -1,16 +1,217 @@
# Package hierarchy # Package hierarchy
The Go package hierarchy in Grafana should be organized logically (Ben Johnson's The Go packages in Grafana should be packaged by feature, keeping
[article](https://medium.com/@benbjohnson/standard-package-layout-7cdbc8391fc1) served as inspiration), according to the packages as small as reasonable while retaining a clear sole ownership
following principles: of a single domain.
- Domain types and interfaces should be in "root" packages (not necessarily at the very top, of the hierarchy, but [Ben Johnson's standard package layout](https://medium.com/@benbjohnson/standard-package-layout-7cdbc8391fc1) serves as
logical roots) inspiration for the way we organize packages.
- Sub-packages should depend on roots - sub-packages here typically contain implementations, for example of services
## Principles of how to structure a service in Grafana
[](services.md)
### Domain types and interfaces should be in local "root" packages
Let's say you're creating a _tea pot_ service, place everything another
service needs to interact with the tea pot service in
_pkg/services/teapot_, choosing a name according to
[Go's package naming conventions](https://go.dev/blog/package-names).
Typically, you'd have one or more interfaces that your service provides
in the root package along with any types, errors, and other constants
that makes sense for another service interacting with this service to
use.
Avoid depending on other services when structuring the root package to
reduce the risk of running into circular dependencies.
### Sub-packages should depend on roots, not the other way around
Small-to-medium sized packages should be able to have only a single
sub-package containing the implementation of the service. By moving the
implementation into a separate package we reduce the risk of triggering
circular dependencies (in Go, circular dependencies are evaluated per
package and this structure logically moves it to be per type or function
declaration).
Large packages may need utilize multiple sub-packages at the discretion
of the implementor. Keep interfaces and domain types to the root
package.
### Try to name sub-packages for project wide uniqueness
Prefix sub-packages with the service name or an abbreviation of the
service name (whichever is more appropriate) to provide an ideally
unique package name. This allows `teaimpl` to be distinguished from
`coffeeimpl` without the need for package aliases, and encourages the
use of the same name to reference your package throughout the codebase.
### A well-behaving service provides test doubles for itself
Other services may depend on your service, and it's good practice to
provide means for those services to set up a test instance of the
dependency as needed. Refer to
[Google Testing's Testing on the Toilet: Know Your Test Doubles](https://testing.googleblog.com/2013/07/testing-on-toilet-know-your-test-doubles.html) for a brief
explanation of how we semantically aim to differentiate fakes, mocks,
and stubs within our codebase.
Place test doubles in a sub-package to your root package named
`<servicename>test` or `<service-abbreviation>test`, such that the `teapot` service may have the
`teapottest` or `teatest`
A stub or mock may be sufficient if the service is not a dependency of a
lot of services or if it's called primarily for side effects so that a
no-op default behavior makes sense.
Services which serve many other services and where it's feasible should
provide an in-memory backed test fake that can be used like the
regular service without the need of complicated setup.
### Separate store and logic
When building a new service, data validation, manipulation, scheduled
events and so forth should be collected in a service implementation that
is built to be agnostic about its store.
The storage should be an interface that is not directly called from
outside the service and should be kept to a minimum complexity to
provide the functionality necessary for the service.
A litmus test to reduce the complexity of the storage interface is
whether an in-memory implementation is a feasible test double to build
to test the service.
### Outside the service root
Some parts of the service definition remains outside the
service directory and reflects the legacy package hierarchy.
As of June 2022, the parts that remain outside the service are:
#### Migrations
`pkg/services/sqlstore/migrations` contains all migrations for SQL
databases, for all services (not including Grafana Enterprise).
Migrations are written per the [database.md](database.md#migrations) document.
#### API endpoints
`pkg/api/api.go` contains the endpoint definitions for the most of
Grafana HTTP API (not including Grafana Enterprise).
## Practical example ## Practical example
The `pkg/plugins` package contains plugin domain types, for example `DataPlugin`, and also interfaces The following is a simplified example of the package structure for a
such as `RequestHandler`. Then you have the `pkg/plugins/managers` subpackage, which contains concrete implementations service that doesn't do anything in particular.
such as the service `PluginManager`. The subpackage `pkg/plugins/backendplugin/coreplugin` contains `plugins.DataPlugin`
implementations. None of the methods or functions are populated and in practice most
packages will consist of multiple files. There isn't a Grafana-wide
convention for which files should exist and contain what.
`pkg/services/alphabetical`
```
package alphabetical
type Alphabetical interface {
// GetLetter returns either an error or letter.
GetLetter(context.Context, GetLetterQuery) (Letter, error)
// ListCachedLetters cannot fail, and doesn't return an error.
ListCachedLetters(context.Context, ListCachedLettersQuery) Letters
// DeleteLetter doesn't have any return values other than errors, so it
// returns only an error.
DeleteLetter(context.Contxt, DeltaCommand) error
}
type Letter byte
type Letters []Letter
type GetLetterQuery struct {
ID int
}
// Create queries/commands for methods even if they are empty.
type ListCachedLettersQuery struct {}
type DeleteLetterCommand struct {
ID int
}
```
`pkg/services/alphabetical/alphabeticalimpl`
```
package alphabeticalimpl
// this name can be whatever, it's not supposed to be used from outside
// the service except for in Wire.
type Svc struct { … }
func ProviceSvc(numbers numerical.Numerical, db db.DB) Svc { … }
func (s *Svc) GetLetter(ctx context.Context, q root.GetLetterQuery) (root.Letter, error) { … }
func (s *Svc) ListCachedLetters(ctx context.Context, q root.ListCachedLettersQuery) root.Letters { … }
func (s *Svc) DeleteLetter(ctx context.Context, q root.DeleteLetterCommand) error { … }
type letterStore interface {
Get(ctx.Context, id int) (root.Letter, error)
Delete(ctx.Context, root.DeleteLetterCommand) error
}
type sqlLetterStore struct {
db.DB
}
func (s *sqlStore) Get(ctx.Context, id int) (root.Letter, error) { … }
func (s *sqlStore) Delete(ctx.Context, root.DeleteLetterCommand) error { … }
```
## Legacy package hierarchy
> **Note:** A lot of services still adhere to the legacy model as outlined below. While it is ok to
> extend existing services based on the legacy model, you are _strongly_ encouraged to structure any
> new services or major refactorings using the new package layout.
Grafana has long used a package-by-layer layout where domain types
are placed in **pkg/models**, all SQL logic in **pkg/services/sqlstore**,
and so forth.
This is an example of how the _tea pot_ service could be structured
throughout the codebase in the legacy model.
- _pkg/_
- _api/_
- _api.go_ contains the endpoints for the
- _tea_pot.go_ contains methods on the _pkg/api.HTTPServer_ type
that interacts with the service based on queries coming in via the HTTP
API.
- _dtos/tea_pot.go_ extends the _pkg/models_ file with types
that are meant for translation to and from the API. It's not as commonly
present as _pkg/models_.
- _models/tea_pot.go_ contains the models for the service, this
includes the _command_ and _query_ structs that are used when calling
the service or SQL store methods related to the service and also any
models representing an abstraction provided by the service.
- _services/_
- _sqlstore_
- _tea_pot.go_ contains SQL queries for
interacting with stored objects related to the tea pot service.
- _migrations/tea_pot.go_ contains the migrations necessary to
build the
- _teapot/\*_ contains functions or a service for doing
logical operations beyond those done in _pkg/api_ or _pkg/services/sqlstore_
for the service.
The implementation of legacy services varies widely from service to
service, some or more of these files may be missing and there may be
more files related to a service than those listed here.
Some legacy services providing infrastructure will also take care of the
integration with several domains. The cleanup service both
provides the infrastructure to occasionally run cleanup scripts and
defines the cleanup scripts. Ideally, this would be migrated
to only handle the scheduling and synchronization of clean up jobs.
The logic for the individual jobs would be placed with a service that is
related to whatever is being cleaned up.

Write Preview
Loading…
Cancel
Save