The Go distribution includes a command, named
automates the downloading, building, installation, and testing of Go packages
and commands. This document talks about why we wrote a new command, what it
is, what it's not, and how to use it.
You might have seen early Go talks in which Rob Pike jokes that the idea for Go arose while waiting for a large Google server to compile. That really was the motivation for Go: to build a language that worked well for building the large software that Google writes and runs. It was clear from the start that such a language must provide a way to express dependencies between code libraries clearly, hence the package grouping and the explicit import blocks. It was also clear from the start that you might want arbitrary syntax for describing the code being imported; this is why import paths are string literals.
An explicit goal for Go from the beginning was to be able to build Go code using only the information found in the source itself, not needing to write a makefile or one of the many modern replacements for makefiles. If Go needed a configuration file to explain how to build your program, then Go would have failed.
At first, there was no Go compiler, and the initial development focused on building one and then building libraries for it. For expedience, we postponed the automation of building Go code by using make and writing makefiles. When compiling a single package involved multiple invocations of the Go compiler, we even used a program to write the makefiles for us. You can find it if you dig through the repository history.
The purpose of the new go command is our return to this ideal, that Go programs should compile without configuration or additional effort on the part of the developer beyond writing the necessary import statements.
The way to achieve the simplicity of a configuration-free system is to
establish conventions. The system works only to the extent that those conventions
are followed. When we first launched Go, many people published packages that
had to be installed in certain places, under certain names, using certain build
tools, in order to be used. That's understandable: that's the way it works in
most other languages. Over the last few years we consistently reminded people
(now replaced by
and its conventions: first, that the import path is derived in a known way from
the URL of the source code; second, that the place to store the sources in
the local file system is derived in a known way from the import path; third,
that each directory in a source tree corresponds to a single package; and
fourth, that the package is built using only information in the source code.
Today, the vast majority of packages follow these conventions.
The Go ecosystem is simpler and more powerful as a result.
We received many requests to allow a makefile in a package directory to provide just a little extra configuration beyond what's in the source code. But that would have introduced new rules. Because we did not accede to such requests, we were able to write the go command and eliminate our use of make or any other build system.
It is important to understand that the go command is not a general build tool. It cannot be configured and it does not attempt to build anything but Go packages. These are important simplifying assumptions: they simplify not only the implementation but also, more important, the use of the tool itself.
go command requires that code adheres to a few key,
First, the import path is derived in an known way from the URL of the
source code. For Bitbucket, GitHub, Google Code, and Launchpad, the
root directory of the repository is identified by the repository's
main URL, without the
http:// prefix. Subdirectories are named by
adding to that path.
For example, the Go example programs are obtained by running
git clone https://github.com/golang/example
and thus the import path for the root directory of that repository is
package is stored in a subdirectory, so its import path is
These paths are on the long side, but in exchange we get an automatically managed name space for import paths and the ability for a tool like the go command to look at an unfamiliar import path and deduce where to obtain the source code.
Second, the place to store sources in the local file system is derived
in a known way from the import path. Specifically, the first choice
unset, the go command will fall back to storing source code alongside the
standard Go packages, in
$GOPATH is set to a list of paths, the go command tries
<dir>/src/<import-path> for each of the directories in
Each of those trees contains, by convention, a top-level directory named
bin", for holding compiled executables, and a top-level directory
pkg", for holding compiled packages that can be imported,
and the "
src" directory, for holding package source files.
Imposing this structure lets us keep each of these directory trees
self-contained: the compiled form and the sources are always near each
These naming conventions also let us work in the reverse direction, from a directory name to its import path. This mapping is important for many of the go command's subcommands, as we'll see below.
Third, each directory in a source tree corresponds to a single package. By restricting a directory to a single package, we don't have to create hybrid import paths that specify first the directory and then the package within that directory. Also, most file management tools and UIs work on directories as fundamental units. Tying the fundamental Go unit—the package—to file system structure means that file system tools become Go package tools. Copying, moving, or deleting a package corresponds to copying, moving, or deleting a directory.
Fourth, each package is built using only the information present in the source files. This makes it much more likely that the tool will be able to adapt to changing build environments and conditions. For example, if we allowed extra configuration such as compiler flags or command line recipes, then that configuration would need to be updated each time the build tools changed; it would also be inherently tied to the use of a specific tool chain.
Finally, a quick tour of how to use the go command, to supplement
the information in How to Write Go Code,
which you might want to read first. Assuming you want
to keep your source code separate from the Go distribution source
tree, the first step is to set
$GOPATH, the one piece of global
configuration that the go command needs. The
$GOPATH can be a
list of directories, but by far the most common usage should be to set it to a
single directory. In particular, you do not need a separate entry in
$GOPATH for each of your projects. One
support many projects.
Here’s an example. Let’s say we decide to keep our Go code in the directory
$HOME/mygo. We need to create that directory and set
$ mkdir $HOME/mygo $ export GOPATH=$HOME/mygo $
Into this directory, we now add some source code. Suppose we want to use
the indexing library from the codesearch project along with a left-leaning
red-black tree. We can install both with the "
$ go get code.google.com/p/codesearch/index $ go get github.com/petar/GoLLRB/llrb $
Both of these projects are now downloaded and installed into our
$GOPATH directory. The one tree now contains the two directories
src/github.com/petar/GoLLRB/llrb/, along with the compiled
pkg/) for those libraries and their dependencies.
Because we used version control systems (Mercurial and Git) to check
out the sources, the source tree also contains the other files in the
corresponding repositories, such as related packages. The "
subcommand lists the import paths corresponding to its arguments, and
the pattern "
./..." means start in the current directory
./") and find all packages below that directory
$ go list ./... code.google.com/p/codesearch/cmd/cgrep code.google.com/p/codesearch/cmd/cindex code.google.com/p/codesearch/cmd/csearch code.google.com/p/codesearch/index code.google.com/p/codesearch/regexp code.google.com/p/codesearch/sparse github.com/petar/GoLLRB/example github.com/petar/GoLLRB/llrb $
We can also test those packages:
$ go test ./... ? code.google.com/p/codesearch/cmd/cgrep [no test files] ? code.google.com/p/codesearch/cmd/cindex [no test files] ? code.google.com/p/codesearch/cmd/csearch [no test files] ok code.google.com/p/codesearch/index 0.239s ok code.google.com/p/codesearch/regexp 0.021s ? code.google.com/p/codesearch/sparse [no test files] ? github.com/petar/GoLLRB/example [no test files] ok github.com/petar/GoLLRB/llrb 0.231s $
If a go subcommand is invoked with no paths listed, it operates on the current directory:
$ cd $GOPATH/src/code.google.com/p/codesearch/regexp $ go list code.google.com/p/codesearch/regexp $ go test -v === RUN TestNstateEnc --- PASS: TestNstateEnc (0.00 seconds) === RUN TestMatch --- PASS: TestMatch (0.01 seconds) === RUN TestGrep --- PASS: TestGrep (0.00 seconds) PASS ok code.google.com/p/codesearch/regexp 0.021s $ go install $
go install" subcommand installs the latest copy of the
package into the pkg directory. Because the go command can analyze the
dependency graph, "
go install" also installs any packages that
this package imports but that are out of date, recursively.
Notice that "
go install" was able to determine the name of the
import path for the package in the current directory, because of the convention
for directory naming. It would be a little more convenient if we could pick
the name of the directory where we kept source code, and we probably wouldn't
pick such a long name, but that ability would require additional configuration
and complexity in the tool. Typing an extra directory name or two is a small
price to pay for the increased simplicity and power.
As the example shows, it’s fine to work with packages from many different
projects at once within a single
$GOPATH root directory.
As mentioned above, the go command is not a general-purpose build
tool. In particular, it does not have any facility for generating Go
source files during a build. Instead, if you want to use a tool like
yacc or the protocol buffer compiler, you will need to write a
makefile (or a configuration file for the build tool of your choice)
to generate the Go files and then check those generated source files
into your repository. This is more work for you, the package author,
but it is significantly less work for your users, who can use
go get" without needing to obtain and build
any additional tools.