Go standards and style guidelines

This document describes various guidelines and best practices for GitLab projects using the Go language.

Overview

GitLab is built on top of Ruby on Rails, but we're also using Go for projects where it makes sense. Go is a very powerful language, with many advantages, and is best suited for projects with a lot of IO (disk/network access), HTTP requests, parallel processing, etc. Since we have both Ruby on Rails and Go at GitLab, we should evaluate carefully which of the two is best for the job.

This page aims to define and organize our Go guidelines, based on our various experiences. Several projects were started with different standards and they can still have specifics. They will be described in their respective README.md or PROCESS.md files.

Code Review

We follow the common principles of Go Code Review Comments.

Reviewers and maintainers should pay attention to:

defer functions: ensure the presence when needed, and after err check.
Inject dependencies as parameters.
Void structs when marshaling to JSON (generates null instead of []).

Security

Security is our top priority at GitLab. During code reviews, we must take care of possible security breaches in our code:

XSS when using text/template
CSRF Protection using Gorilla
Use a Go version without known vulnerabilities
Don't leak secret tokens
SQL injections

Remember to run SAST (ULTIMATE) on your project (or at least the gosec analyzer), and to follow our Security requirements.

Web servers can take advantages of middlewares like Secure.

Finding a reviewer

Many of our projects are too small to have full-time maintainers. That's why we have a shared pool of Go reviewers at GitLab. To find a reviewer, use the Engineering Projects page in the handbook. "GitLab Community Edition (CE)" and "GitLab Community Edition (EE)" both have a "Go" section with its list of reviewers.

To add yourself to this list, add the following to your profile in the team.yml file and ask your manager to review and merge.

projects:
  gitlab-ee: reviewer go
  gitlab-ce: reviewer go

Code style and format

Avoid global variables, even in packages. By doing so you will introduce side effects if the package is included multiple times.
Use go fmt before committing (Gofmt is a tool that automatically formats Go source code).

Automatic linting

All Go projects should include these GitLab CI/CD jobs:

go lint:
  image: golang:1.11
  script:
    - go get -u golang.org/x/lint/golint
    - golint -set_exit_status $(go list ./... | grep -v "vendor/")

Once recursive includes become available, you will be able to share job templates like this analyzer.

Dependencies

Dependencies should be kept to the minimum. The introduction of a new dependency should be argued in the merge request, as per our Approval Guidelines. Both License Management (ULTIMATE) and Dependency Scanning (ULTIMATE) should be activated on all projects to ensure new dependencies security status and license compatibility.

Modules

Since Go 1.11, a standard dependency system is available behind the name Go Modules. It provides a way to define and lock dependencies for reproducible builds. It should be used whenever possible.

When Go Modules are in use, there should not be a vendor/ directory. Instead, Go will automatically download dependencies when they are needed to build the project. This is in line with how dependencies are handled with Bundler in Ruby projects, and makes merge requests easier to review.

In some cases, such as building a Go project for it to act as a dependency of a CI run for another project, removing the vendor/ directory means the code must be downloaded repeatedly, which can lead to intermittent problems due to rate limiting or network failures. In these circumstances, you should cache the downloaded code between runs with a .gitlab-ci.yml snippet like this:

.go-cache:
  variables:
    GOPATH: $CI_PROJECT_DIR/.go
  before_script:
    - mkdir -p .go
  cache:
    paths:
      - .go/pkg/mod/

test:
  extends: .go-cache
  # ...

There was a bug on modules checksums in Go < v1.11.4, so make sure to use at least this version to avoid checksum mismatch errors.

ORM

We don't use object-relational mapping libraries (ORMs) at GitLab (except ActiveRecord in Ruby on Rails). Projects can be structured with services to avoid them. PQ should be enough to interact with PostgreSQL databases.

Migrations

In the rare event of managing a hosted database, it's necessary to use a migration system like ActiveRecord is providing. A simple library like Journey, designed to be used in postgres containers, can be deployed as long-running pods. New versions will deploy a new pod, migrating the data automatically.

Testing

Testing frameworks

We should not use any specific library or framework for testing, as the standard library provides already everything to get started. If there is a need for more sophisticated testing tools, the following external dependencies might be worth considering in case we decide to use a specific library or framework:

Subtests

Use subtests whenever possible to improve code readability and test output.

Better output in tests

When comparing expected and actual values in tests, use testify/require.Equal, testify/require.EqualError, testify/require.EqualValues, and others to improve readability when comparing structs, errors, large portions of text, or JSON documents:

type TestData struct {
    // ...
}

func FuncUnderTest() TestData {
    // ...
}

func Test(t *testing.T) {
    t.Run("FuncUnderTest", func(t *testing.T) {
        want := TestData{}
        got := FuncUnderTest()

        require.Equal(t, want, got) // note that expected value comes first, then comes the actual one ("diff" semantics)
    })
}

Table-Driven Tests

Using Table-Driven Tests is generally good practice when you have multiple entries of inputs/outputs for the same function. Below are some guidelines one can follow when writing table-driven test. These guidelines are mostly extracted from Go standard library source code. Keep in mind it's OK not to follow these guidelines when it makes sense.

Defining test cases

Each table entry is a complete test case with inputs and expected results, and sometimes with additional information such as a test name to make the test output easily readable.

Define a slice of anonymous struct inside of the test.
Define a slice of anonymous struct outside of the test.
Named structs for code reuse.
Using map[string]struct{}.

Contents of the test case

Ideally, each test case should have a field with a unique identifier to use for naming subtests. In the Go standard library, this is commonly the name string field.
Use want/expect/actual when you are specifcing something in the test case that will be used for assertion.

Variable names

Each table-driven test map/slice of struct can be named tests.
When looping through tests the anonymous struct can be referred to as tt or tc.
The description of the test can be referred to as name/testName/tn.

Benchmarks

Programs handling a lot of IO or complex operations should always include benchmarks, to ensure performance consistency over time.

CLIs

Every Go program is launched from the command line. cli is a convenient package to create command line apps. It should be used whether the project is a daemon or a simple cli tool. Flags can be mapped to environment variables directly, which documents and centralizes at the same time all the possible command line interactions with the program. Don't use os.GetEnv, it hides variables deep in the code.

Daemons

Logging

The usage of a logging library is strongly recommended for daemons. Even though there is a log package in the standard library, we generally use Logrus. Its plugin ("hooks") system makes it a powerful logging library, with the ability to add notifiers and formatters at the logger level directly.

Structured (JSON) logging

Every binary ideally must have structured (JSON) logging in place as it helps with searching and filtering the logs. At GitLab we use structured logging in JSON format, as all our infrastructure assumes that. When using Logrus you can turn on structured logging simply by using the build in JSON formatter. This follows the same logging type we use in our Ruby applications.

How to use Logrus

There are a few guidelines one should follow when using the Logrus package:

When printing an error use WithError. For example, logrus.WithError(err).Error("Failed to do something").
Since we use structured logging we can log fields in the context of that code path, such as the URI of the request using WithField or WithFields. For example, logrus.WithField("file", "/app/go).Info("Opening dir"). If you have to log multiple keys, always use WithFields instead of calling WithField more than once.

Tracing and Correlation

LabKit is a place to keep common libraries for Go services. Currently it's vendored into two projects: Workhorse and Gitaly, and it exports two main (but related) pieces of functionality:

gitlab.com/gitlab-org/labkit/correlation: for propagating and extracting correlation ids between services.
gitlab.com/gitlab-org/labkit/tracing: for instrumenting Go libraries for distributed tracing.

This gives us a thin abstraction over underlying implementations that is consistent across Workhorse, Gitaly, and, in future, other Go servers. For example, in the case of gitlab.com/gitlab-org/labkit/tracing we can switch from using Opentracing directly to using Zipkin or Gokit's own tracing wrapper without changes to the application code, while still keeping the same consistent configuration mechanism (i.e. the GITLAB_TRACING environment variable).

Context

Since daemons are long-running applications, they should have mechanisms to manage cancellations, and avoid unnecessary resources consumption (which could lead to DDOS vulnerabilities). Go Context should be used in functions that can block and passed as the first parameter.

Dockerfiles

Every project should have a Dockerfile at the root of their repository, to build and run the project. Since Go program are static binaries, they should not require any external dependency, and shells in the final image are useless. We encourage Multistage builds:

They let the user build the project with the right Go version and dependencies.
They generate a small, self-contained image, derived from Scratch.

Generated docker images should have the program at their Entrypoint to create portable commands. That way, anyone can run the image, and without parameters it will display its help message (if cli has been used).

Return to Development documentation.