saypi

README

saypi (say-pee-eye)

Cowsay API to demonstrate several Golang design [patterns].

API Docs

We haven't found a consistent pattern for documenting service APIs. Sharing client code helps but doesn't fulfill all the needs of good documentation. For now we rely on manually maintained markdown files such as the one in this repository.

Patterns

Repositories

In our monolithic Ruby codebase, we access data via an ActiveRecord-inspired library. Models are accessible to any part of the application and any part of the code can perform essentially arbitrary actions against the database (e.g. updating a particular field, running custom commands). The ORM provides a lot of power out of the box and abstracts away a ton of thinking about the interface with your persistence layer. I'd argue that this is a mistake -- persistence is one of the most challenging parts of developing a consistent and scalable application and shouldn't be treated as a generic problem.

In saypi, all access to the database is wrapped by repository types. The only way to communicate with the database is through the narrow interface defined on each repository. You'll never see a Save method allowing you to update arbitrary data on a model. This makes data access patterns more predictable, testable and refactorable. For example:

• You can't accidentally write code in an unrelated module that triggers a full table scan.
• You can easily stub out the database in unrelated unit tests.
• You can refactor data access patterns in one place, with a clearly testable interface. For example, I was able to add read replicas with minimal code churn.

Handler interface

We agree with Square's assessment that full-featured web frameworks are unncessary in Go. We found that the specific choice of mux and handler interface doesn't matter nearly as much as agreeing on a consistent pattern. For example, it's important that all of our applications provide the same request IDs in their log lines and can use the same package to record consistent metrics.

We use the context package to pass request-scoped values and our handlers use the signature func(ctx context.Context, rw http.ResponseWriter, req *http.Request). Goji makes it easy to mix-and-match our handlers and middleware with code that assumes an http.Handler.

Error handling

In our early Go applications we struggled to return useful errors to our users and log the right details for debugging. We tried various patterns that left us returning unhelpful errors, overly "helpful" errors that exposed us to SSRF vulnerabilities or, at best, poorly structured error data.

Errors may start deep in the application, far from the request handler. In our Ruby applications we would throw a UserError that bubbles all the way up the stack into error handling middleware. This deeply couples the entire application to a particular transport and assumptions about user permissions. In Go, we prefer to think of errors in our code as a separate concept from error responses returned to the user.

In this snippet, the repository translates a known database error into a package-specific type. For unknown errors it prepends useful context to the database error for later debugging.

var errCursorNotFound = errors.New("Invalid cursor")
...
func (r *repository) listUserMoods(...) ([]Mood, bool, error) {
...
	if err == sql.ErrNoRows {
		return nil, false, errCursorNotFound
	} else if err != nil {
		return nil, false, fmt.Errorf("listing mood: %v", err)
	}
...
}

Only when the error reaches the request handler do we determine how the error will translate into a serialized response to the user. If the controller recognizes the error, it will translate it into a concrete type that represents the error in a structured way. If the controller does not recognize the error it does not panic; it explicitly returns a failure to the client. In either case, a helper serializes the error into JSON and responds over the wire.

func (c *Controller) ListMoods(ctx context.Context, w http.ResponseWriter, r *http.Request) {
...
	if err == errCursorNotFound {
		respond.UserError(ctx, w, http.StatusBadRequest, usererrors.InvalidParams{{
			Params:  []string{cursorParam},
			Message: "must refer to an existing object",
		}})
		return
	} else if err != nil {
		respond.InternalError(ctx, w, err)
		return
	}
...
}

Every error response we return to the client is represented by a concrete type and each type corresponds to a unique string code. Each error also generates a human-readable message for easier debugging on the wire. The library can serialize and deserialize these types to JSON.

// InvalidParamsEntry represents a single error for InvalidParams
type InvalidParamsEntry struct {
	Params  []string `json:"params"`
	Message string   `json:"message"`
}

// InvalidParams represents a list of parameter validation
// errors. Each element in the list contains an explanation of the
// error and a list of the parameters that failed.
type InvalidParams []InvalidParamsEntry

// Code returns "invalid_params"
func (e InvalidParams) Code() string { return "invalid_params" }

// Message returns a joined representation of parameter messages.
// When possible, the underlying data should be used instead to
// separate errors by parameter.
func (e InvalidParams) Message() string {
	...
}

{
  "code": "invalid_params",
  "message": "Parameter `starting_after` must refer to an existing object.",
  "data": [
    {
      "params": ["starting_after"],
      "message": "must refer to an existing object",
    }
  ]
}

Applications can register custom error types in addition to common types provided by the library. Clients get access to structured, typed details on the error rather than having to parse arbitrary string messages and untyped metadata.

// This makes an HTTP request to the application
err := cli.SetMood(&test.Mood)

switch usererr := client.UserError(err).(type) {
case usererrors.InvalidParams:
	for _, param := range usererr {
		log.Printf("%s: %s", param.Params, param.Message)
	}
case usererrors.NotFound:
	log.Printf("You must be dreaming. There is no such %s.", usererr.Resource)
case nil:
	log.Printf("I have no idea what to do with a %s.", err)
}

For a somewhat more complex example, consider returning additional information from the repository layer and translating it into a user-visible message. In this case, if deletion fails due to a uniqueness violation, we return an error listing the conflicting IDs.

type conflictErr struct{ IDs []string }
func (e conflictErr) Error() string { ... }

func (r *repository) DeleteMood(userID, name string) error {
	if isBuiltin(name) {
		return errBuiltinMood
	}

	queryArgs := struct{ UserID, Name string }{userID, name}
	if err := doDelete(r.deleteMood, queryArgs); err != nil {
		if dbErr, ok := err.(*pq.Error); !ok || dbErr.Code != dbErrFKViolation {
			return err
		}

		var lineIDs []string
		if err := r.findMoodLines.Select(&lineIDs, queryArgs); err != nil {
			return fmt.Errorf("listing lines for mood %q and user %q: %v", name, userID, err)
		}

		return conflictErr{lineIDs}
	}

	return nil
}

Suppose, for example, we don't want to expose those IDs to the same clients who can delete moods. The request handler can translate the error from the repository into a sanitized message to the client that only contains a count of the conflicting IDs.

func (c *Controller) DeleteMood(ctx context.Context, w http.ResponseWriter, r *http.Request) {
	userID := mustUserID(ctx)
	name := pat.Param(ctx, "mood")

	err := c.repo.DeleteMood(userID, name)
	if conflict, ok := err.(conflictErr); ok {
		respond.UserError(ctx, w, http.StatusBadRequest, usererrors.ActionNotAllowed{
			Action: fmt.Sprintf("delete a mood associated with %d conversation lines", len(conflict.IDs)),
		})
	}
	...
}

For a more privileged client, we might return a completely different error type that includes the actual IDs. In the client, we can type assert to retrieve the original error. Though our error has limited structure this time, it has enough that the client could use it to customize the message to the user.

// ActionNotAllowed describes an action that is not permitted.
type ActionNotAllowed struct {
	Action string `json:"action"`
}

// Code returns "action_not_allowed"
func (e ActionNotAllowed) Code() string { return "action_not_allowed" }

// Message returns a string describing the disallowed action
func (e ActionNotAllowed) Message() string {
	return fmt.Sprintf("You may not %s.", e.Action)
}

{
  "code": "action_not_allowed",
  "message": "You may not delete a mood associated with 3 conversation lines.",
  "data": {
    "action": "delete a mood associated with 3 conversation lines",
  }
}

err := cli.DeleteMood("cross")
if action, ok := client.UserError(err).(usererrors.ActionNotAllowed); ok {
    log.Printf("Seriously? You think you can just %s?", action.Action)
}

This pattern means that:

• Low-level modules are not tightly coupled to the transport and user permissions.
• Error responses contain information for human debuggers as well as structured information for programmatic clients.
• Golang clients can easily reify the original error type and manipulate structured data.
• We don't panic.

Initialization and the main method

Get out of your main method as quickly as possible! Command-line arguments and environment variables offer an untyped interface that makes testing difficult and error-prone.

The main method should do little more than read configuration from the environment. Saypi pushes the limit of what a main method should do by binding to a port and setting up graceful shutdown. Initialization such as connecting to the database should be handled by a separate App type that takes typed configuration from main and implements http.Handler. The App is easy to instantiate in tests without mucking with the environment and string-based configuration. It makes it easy to write functional tests that exercise initialization paths without deeply coupling packages in your app to each other.

Missing pieces

We've learned a lot working Go into new services over the past year but there are still a few places where we don't quite feel like we've landed on the right patterns.

• Configuration: We've used a mix of command-line flags, environment variables and JSON files to configure our applications. Across Stripe, we like environment variables as a default but some services have complex configuration that's easier to represent in a file. Passing secrets to the application introduces further wrinkles.
• Healthchecking: Deep healthchecks risk falsely marking every instance of a service as down rather than entering a degraded state while very shallow healthchecks can mask problems on a single instance. This isn't a Go specific problem, but we'll likely have to solve it in Go as our primary language for new services.
• Testing: We tried suite-based interfaces like Gocheck and Ginkgo but felt like we were fighting the language. We much prefer the stdlib testing interfaces but haven't figured out how much to cede to an assertion library like testify.assert. For now, our rule is that anything goes as long as the tests are of the form func Test*(t *testing.T).

Want to help us solve these problems and much more?

TODOs and Qs

Boring TODO

• Package descriptions
• Dependency management (vendor experiment?)
• Make this runnable with a Heroku button

Maybe TODO

• frontend interface, Go as a static fileserver, JS tests running against stub
• Use 201 created with Location header to force generating internal URLs? Return next/prev URLs for pagination in the Location header like Greenhouse?
• Generate public URLs for a conversation (maybe have auth package support returning a public version of any url?)
• Object creation limits
• Write an example of refactoring list queries to use a read-replica
• Server hosts its own Swagger API docs
• Support rendering conversations as text instead of JSON?
• Generate spans in client or at least take a context for cancellation
• Client with a much more stubbable interface (e.g. Service thing from verificator)
• Client that's more generic so we could use a single client across services

Maybe include these patterns?

• API clients across multiple languages
• DI?
• Dependency management? It's just being angry about Godep...