Go (programming language)

History

The language was announced in November 2009. It is used in some of Google's production systems, as well as by other firms.

Two major implementations exist:

The "gc" toolchain has been self-hosting since version 1.5.

Go originated as an experiment by Google engineers Robert Griesemer, Rob Pike, and Ken Thompson to design a new programming language that would resolve common criticisms of other languages while maintaining their positive characteristics. The new language was to include the following features:

In later interviews, all three of the language designers cited their shared dislike of C++'s complexity as a primary motivation for designing a new language.

Go 1.0 was released in March 2012.

Go 1.7 added "one tiny language change" and one port to macOS 10.12 Sierra plus some experimental ports, e.g. for Linux on z Systems (linux/s390x). Some library changes apply; for example, Unicode 9.0 is now supported.

Language design

Go is recognizably in the tradition of C, but makes many changes to improve brevity, simplicity, and safety. Go consists of,

Syntax

Go's syntax includes changes from C aimed at keeping code concise and readable. A combined declaration/initialization operator was introduced that allows the programmer to write i := 3 or s := "Hello, world!", without specifying the types of variables. This contrasts with C's int i = 3; and const char *s = "Hello, world!";. Semicolons still terminate statements, but are implicit when the end of a line occurs. Functions may return multiple values, and returning a result, err pair is the conventional way a function indicates an error to its caller in Go. Go adds literal syntaxes for initializing struct parameters by name, and for initializing maps and slices. As an alternative to C's three-statement for loop, Go's range expressions allow concise iteration over arrays, slices, strings, maps, and channels.

Types

Go has a number of built-in types, including numeric ones (byte, int64, float32, etc.), booleans, and character strings (string). Strings are immutable; built-in operators and keywords (rather than functions) provide concatenation, comparison, and UTF-8 encoding and decoding. Record types can be defined with the struct keyword.

For each type T and each non-negative integer constant n, there is an array type denoted [n]T; arrays of differing lengths are thus of different types. Dynamic arrays are available as "slices", denoted []T for some type T. These have a length and a capacity specifying when new memory needs to be allocated to expand the array. Several slices may share their underlying memory.

Pointers are available for all types, and the pointer-to-T type is denoted *T. Address-taking and indirection use the & and * operators as in C, or happen implicitly through the method call or attribute access syntax. There is no pointer arithmetic, except via the special unsafe.Pointer type in the standard library.

For a pair of types K, V, the type map[K]V is the type of hash tables mapping type-K keys to type-V values. Hash tables are built into the language, with special syntax and built-in functions. chan T is a channel that allows sending values of type T between concurrent Go processes.

Aside from its support for interfaces, Go's type system is nominal: the type keyword can be used to define a new named type, which is distinct from other named types that have the same layout (in the case of a struct, the same members in the same order). Some conversions between types (e.g., between the various integer types) are pre-defined and adding a new type may define additional conversions, but conversions between named types must always be invoked explicitly. For example, the type keyword can be used to define a type for IPv4 addresses, which are 32-bit unsigned integers:

With this type definition, ipv4addr(x) interprets the uint32 value x as an IP address. Simply assigning x to a variable of type ipv4addr is a type error.

Constant expressions may be either typed or "untyped"; they are given a type when assigned to a typed variable, if the value they represent passes a compile-time check.

Function types are indicated by the func keyword; they take zero or more parameters and return zero or more values, all of which are typed. The parameter and return values determine a function type; thus, func(string, int32) (int, error) is the type of functions that take a string and a 32-bit signed integer, and return a signed integer (of default width) and a value of the built-in interface type error.

Any named type has a method set associated with it. The IP address example above can be extended with a method for converting an address to a human-readable representation, viz.:

Due to nominal typing, this method definition adds a method to ipv4addr, but not on uint32. While methods have special definition and call syntax, there is no distinct method type.

Interface system

Go provides two features that replace class inheritance.

The first is embedding, which can be viewed as an automated form of composition or delegation.

The second are its interfaces, which provides runtime polymorphism. Interfaces provide a limited form of structural typing in the otherwise nominal type system of Go. Any type that implements all methods of an interface conforms to that interface. Go interfaces were designed after protocols from the Smalltalk programming language. Multiple sources use the term duck typing when describing Go interface. Although the term duck typing is not precisely defined and therefore not wrong, it usually implies that type conformance is not statically checked. Since conformance to a Go interface is checked statically by the Go compiler (except when performing a type assertion), the Go authors prefer to use the term structural typing.

An interface specifies a set of types by listing required methods and their types, and is satisfied by any type that has the required methods. Implementing types do not need to specify their implementing of interfaces, so if Shape, Square and Circle are defined as:

then both Square and Circle are implicitly a Shape and can be assigned to a Shape-typed variable. In formal language, Go's interface system provides structural rather than nominal typing. Interfaces can embed other interfaces with the effect of creating a combined interface that is satisfied by exactly the types that implement the embedded interface and any methods that the newly defined interface adds.

The Go standard library uses interfaces to provide genericity in several places, including the input/output system that is based on the concepts of Reader and Writer.

Besides calling methods via interfaces, Go allows converting interface values to other types with a run-time type check. The language constructs to do so are the type assertion, which checks against a single potential type, and the type switch, which checks against multiple types.

The empty interface interface{} is an important corner case because it can refer to an item of any concrete type. It is similar to the Object class in Java or C#, but with the difference that the empty interface is satisfied by any type, including built-in types like int (while in Java and C#, an Object variable can only hold instances of reference type). Code using the empty interface cannot simply call methods (or built-in operators) on the referred-to object, but it can store the interface{} value, try to convert it to a more useful type via a type assertion or type switch, or inspect it with Go's reflect package. Because interface{} can refer to any value, it is a limited way to escape the restrictions of static typing, like void* in C but with additional run-time type checks.

Interface values are implemented using pointer to data and a second pointer to run-time type information. Like some other types implemented using pointers in Go, interface values are nil if uninitialized.

Package system

In Go's package system, each package has a path (e.g., "compress/bzip2" or "golang.org/x/net/html") and a name (e.g., bzip2 or html). References to other packages' definitions must always be prefixed with the other package's name, and only the capitalized names from other packages are accessible: io.Reader is public but bzip2.reader is not. The go get command can retrieve packages stored in a remote repository such as GitHub., and developers are encouraged to develop packages inside a base path corresponding to a source repository (such as github.com/user_name/package_name) to reduce the likelihood of name collision with future additions to the standard library or other external libraries.

Proposals exist to introduce a proper package management solution for Go similar to Rust's cargo system or Node's npm system.

Concurrency: goroutines and channels

The Go language has built-in facilities, as well as library support, for writing concurrent programs. Concurrency refers not only to CPU parallelism, but also to asynchrony: letting slow operations like a database or network-read run while the program does other work, as is common in event-based servers.

The primary concurrency construct is the goroutine, a type of light-weight process. A function call prefixed with the go keyword starts a function in a new goroutine. The language specification does not specify how goroutines should be implemented, but current implementations multiplex a Go process's goroutines onto a smaller set of operating system threads, similar to the scheduling performed in Erlang.

While a standard library package featuring most of the classical concurrency control structures (mutex locks, etc.) is available, idiomatic concurrent programs instead prefer channels, which provide send messages between goroutines. Optional buffers store messages in FIFO order and allow sending goroutines to proceed before their messages are received.

Channels are typed, so that a channel of type chan T can only be used to transfer messages of type T. Special syntax is used to operate on them; <-ch is an expression that causes the executing goroutine to block until a value comes in over the channel ch, while ch <- x sends the value x (possibly blocking until another goroutine receives the value). The built-in switch-like select statement can be used to implement non-blocking communication on multiple channels; see below for an example. Go has a memory model describing how goroutines must use channels or other operations to safely share data.

The existence of channels sets Go apart from actor model-style concurrent languages like Erlang, where messages are addressed directly to actors (corresponding to goroutines); the actor style can be simulated in Go by maintaining a one-to-one correspondence between goroutines and channels, but the language allows multiple goroutines to share a channel, or a single goroutine to send and receive on multiple channels.

From these tools one can build concurrent constructs like worker pools, pipelines (in which, say, a file is decompressed and parsed as it downloads), background calls with timeout, "fan-out" parallel calls to a set of services, and others. Channels have also found uses further from the usual notion of interprocess communication, like serving as a concurrency-safe list of recycled buffers, implementing coroutines (which helped inspire the name goroutine), and implementing iterators.

Concurrency-related structural conventions of Go (channels and alternative channel inputs) are derived from Tony Hoare's communicating sequential processes model. Unlike previous concurrent programming languages such as Occam or Limbo (a language on which Go co-designer Rob Pike worked), Go does not provide any built-in notion of safe or verifiable concurrency. While the communicating-processes model is favored in Go, it is not the only one: all goroutines in a program share a single address space. This means that mutable objects and pointers can be shared between goroutines; see § Lack of race condition safety, below.

Suitability for parallel programming

Although Go's concurrency features are not aimed primarily at parallel processing, they can be used to program shared memory multi-processor machines. Various studies have been done into the effectiveness of this approach. One of these studies compared the size (in lines of code) and speed of programs written by a seasoned programmer not familiar with the language and corrections to these programs by a Go expert (from Google's development team), doing the same for Chapel, Cilk and Intel TBB. The study found that the non-expert tended to write divide-and-conquer algorithms with one go statement per recursion, while the expert wrote distribute-work-synchronize programs using one goroutine per processor. The expert's programs were usually faster, but also longer.

Lack of race condition safety

There are no restrictions on how goroutines access shared data, making race conditions possible. Specifically, unless a program explicitly synchronizes via channels or other means, writes from one goroutine might be partly, entirely, or not at all visible to another, often with no guarantees about ordering of writes. Furthermore, Go's internal data structures like interface values, slice headers, hash tables, and string headers are not immune to race conditions, so type and memory safety can be violated in multithreaded programs that modify shared instances of those types without synchronization.

Instead of language support, safe concurrent programming thus relies on conventions; for example, Chisnall recommends an idiom called "aliases xor mutable", meaning that passing a mutable value (or pointer) over a channel signals a transfer of ownership over the value to its receiver.

Omissions

Go deliberately omits certain features common in other languages, including (implementation) inheritance, generic programming, assertions, pointer arithmetic, and implicit type conversions.

Of these language features, the Go authors express an openness to generic programming, explicitly argue against assertions and pointer arithmetic, while defending the choice to omit type inheritance as giving a more useful language, encouraging instead the use of interfaces to achieve dynamic dispatch and composition to reuse code. Composition and delegation are in fact largely automated by struct embedding; according to researchers Schmager et al., this feature "has many of the drawbacks of inheritance: it affects the public interface of objects, it is not fine-grained (i.e, no method-level control over embedding), methods of embedded objects cannot be hidden, and it is static", making it "not obvious" whether programmers will not overuse it to the extent that programmers in other languages are reputed to overuse inheritance.

Regarding generic programming, some built-in functions are in fact type-generic, but these are treated as special cases; Rob Pike calls this a weakness of the language that may at some point be changed. The Google team that designs the language built at least one compiler for an experimental Go dialect with generics, but did not release it.

After initially omitting exceptions, the exception-like panic/recover mechanism was eventually added to the language, which the Go authors advise using for unrecoverable errors such as those that should halt an entire program or server request, or as a shortcut to propagate errors up the stack within a package (but not across package boundaries; there, error returns are the standard API).

Criticism

Go critics assert that:

The language designers argue that these trade-offs are important to Go's success, and explain some particular decisions at length, though they do express openness to adding some form of generic programming in the future, and to pragmatic improvements in areas like standardizing ways to apply code generation. Regarding GC, Go defenders point to pause-time reduction in later versions (e.g. Go 1.6), while acknowledging their GC algorithm is not hard real-time.

Conventions and code style

The Go authors put substantial effort into molding the style and design of Go programs:

Language tools

Go includes the same sort of debugging, testing, and code-vetting tools as many language distributions. The Go distribution includes, among other tools,

It also includes profiling and debugging support, runtime instrumentation (to, for example, track garbage collection pauses), and a race condition tester.

There is an ecosystem of third-party tools that add to the standard distribution, such as gocode, which enables code autocompletion in many text editors, goimports (by a Go team member), which automatically adds/removes package imports as needed, errcheck, which detects code that might unintentionally ignore errors, and more. Plugins exist to add language support in widely used text editors, and at least one IDE, LiteIDE, is branded as "a simple, open source, cross-platform Go IDE."

Hello world

Here is a Hello world program in Go:

Concurrency example

The following simple program demonstrates Go's concurrency features to implement an asynchronous program. It launches two "goroutines" (lightweight threads): one waits for the user to type some text, while the other implements a timeout. The select statement waits for either of these goroutines to send a message to the main routine, and acts on the first message to arrive (example adapted from Chisnall).

Projects using Go

Some notable open-source applications in Go include:

Some notable open-source frameworks using Go:

Other notable companies and sites using Go (generally together with other languages, not exclusively) include:

Reception

The interface system, and the deliberate omission of inheritance, were praised by Michele Simionato, who likened these language characteristics to those of Standard ML, calling it "a shame that no popular language has followed [this] particular route in the design space".

Dave Astels at Engine Yard wrote:

Go is extremely easy to dive into. There are a minimal number of fundamental language concepts and the syntax is clean and designed to be clear and unambiguous. Go is still experimental and still a little rough around the edges.

Ars Technica interviewed Rob Pike, one of the authors of Go, and asked why a new language was needed. He replied that:

It wasn't enough to just add features to existing programming languages, because sometimes you can get more in the long run by taking things away. They wanted to start from scratch and rethink everything. ... [But they did not want] to deviate too much from what developers already knew because they wanted to avoid alienating Go's target audience.

Go was named Programming Language of the Year by the TIOBE Programming Community Index in its first year, 2009, for having a larger 12-month increase in popularity (in only 2 months, after its introduction in November) than any other language that year, and reached 13th place by January 2010, surpassing established languages like Pascal. By June 2015, its ranking had dropped to below 50th in the index, placing it lower than COBOL and Fortran. But as of January 2017, its ranking had surged to 13th again, indicating significant growth in popularity and adoption. Go was awarded TIOBE programming language of the year 2016.

Regarding Go, Bruce Eckel has stated:

The complexity of C++ (even more complexity has been added in the new C++), and the resulting impact on productivity, is no longer justified. All the hoops that the C++ programmer had to jump through in order to use a C-compatible language make no sense anymore -- they're just a waste of time and effort. Go makes much more sense for the class of problems that C++ was originally intended to solve.

A 2011 evaluation of the language and its gc implementation in comparison to C++ (GCC), Java and Scala by a Google engineer found that:

Go offers interesting language features, which also allow for a concise and standardized notation. The compilers for this language are still immature, which reflects in both performance and binary sizes.

The evaluation got a rebuttal from the Go development team. Ian Lance Taylor, who had improved the Go code for Hundt's paper, had not been aware of the intention to publish his code, and says that his version was "never intended to be an example of idiomatic or efficient Go"; Russ Cox then did optimize the Go code, as well as the C++ code, and got the Go code to run slightly faster than C++ and more than an order of magnitude faster than the "optimized" code in the paper.

Naming dispute

On 10 November 2009, the day of the general release of the language, Francis McCabe, developer of the Go! programming language (note the exclamation point), requested a name change of Google's language to prevent confusion with his language, which he had spent 10 years developing. McCabe raised concerns that "the 'big guy' will end up steam-rollering over" him, and this concern resonated with the more than 120 developers who commented on Google's official issues thread saying they should change the name, with some even saying the issue contradicts Google's motto of: Don't be evil.

On 12 October 2010, the issue was closed by Google developer Russ Cox (@rsc) with the custom status "Unfortunate" accompanied by the following comment:

"There are many computing products and services named Go. In the 11 months since our release, there has been minimal confusion of the two languages."