Author: Ian Lance Taylor
Last update: October 15, 2018
A proposal for how to make incompatible changes from Go 1 to Go 2 while breaking as little as possible.
Currently the Go language and standard libraries are covered by the Go 1 compatibility guarantee. The goal of that document was to promise that new releases of Go would not break existing working programs.
Among the goals for the Go 2 process is to consider changes to the language and standard libraries that will break the guarantee. Since Go is used in a distributed open source environment, we cannot rely on a flag day. We must permit the interoperation of different packages written using different versions of Go.
Every language goes through version transitions. As background, here are some notes on what other languages have done. Feel free to skip the rest of this section.
C language versions are driven by the ISO standardization process. C language development has paid close attention to backward compatibility. After the first ISO standard, C90, every subsequent standard has maintained strict backward compatibility. Where new keywords have been introduced, they are introduced in a namespace reserved by C90 (an underscore followed by an uppercase ASCII letter) and are made more accessible via a
#define macro in a header file that did not previously exist (examples are
_Complex, defined as
_Bool, defined as
<stdbool.h>). None of the basic language semantics defined in C90 have changed.
In addition, most C compilers provide options to define precisely which version of the C standard the code should be compiled for (for example,
-std=c90). Most standard library implementations support feature macros that may be #define’d before including the header files to specify exactly which version of the library should be provided (for example,
_ISOC99_SOURCE). While these features have had bugs, they are fairly reliable and are widely used.
A key feature of these options is that code compiled at different language/library versions can in general all be linked together and work as expected.
The first standard, C90, did introduce breaking changes to the previous C language implementations, known informally as K&R C. New keywords were introduced, such as
volatile (actually that might have been the only new keyword in C90). The precise implementation of integer promotion in integer expressions changed from unsigned-preserving to value-preserving. Fortunately it was easy to detect code using the new keywords due to compilation errors, and easy to adjust that code. The change in integer promotion actually made it less surprising to naive users, and experienced users mostly used explicit casts to ensure portability among systems with different integer sizes, so while there was no automatic detection of problems not much code broke in practice.
There were also some irritating changes. C90 introduced trigraphs, which changed the behavior of some string constants. Compilers adapted with options like -no-trigraphs and -Wtrigraphs.
More seriously, C90 introduced the notion of undefined behavior, and declared that programs that invoked undefined behavior might take any action. In K&R C, the cases that C90 described as undefined behavior were mostly treated as what C90 called implementation-defined behavior: the program would take some non-portable but predictable action. Compiler writers absorbed the notion of undefined behavior, and started writing optimizations that assumed that the behavior would not occur. This caused effects that surprised people not fluent in the C standard. I won’t go into the details here, but one example of this (from my blog) is signed overflow.
C of course continues to be the preferred language for kernel development and the glue language of the computing industry. Though it has been partially replaced by newer languages, this is not because of any choices made by new versions of C.
The lessons I see here are:
C++ language versions are also now driven by the ISO standardization process. Like C, C++ pays close attention to backward compatibility. C++ has been historically more free with adding new keywords (there are 10 new keywords in C++11). This works out OK because the newer keywords tend to be relatively long (
static_assert) and compilation errors make it easy to find code using the new keywords as identifiers.
C++ uses the same sorts of options for specifying the standard version for language and libraries as are found in C. It suffers from the same sorts of problems as C with regard to undefined behavior.
An example of a breaking change in C++ was the change in the scope of a variable declared in the initialization statement of a for loop. In the pre-standard versions of C++, the scope of the variable extended to the end of the enclosing block, as though it were declared immediately before the for loop. During the development of the first C++ standard, C++98, this was changed so that the scope was only within the for loop itself. Compilers adapted by introducing options like
-ffor-scope so that users could control the expected scope of the variable (for a period of time, when compiling with neither
-fno-for-scope, the GCC compiler used the old scope but warned about any code that relied on it).
Despite the relatively strong backward compatibility, code written in new versions of C++, like C++11, tends to have a very different feel than code written in older versions of C++. This is because styles have changed to use new language and library features. Raw pointers are less commonly used, range loops are used rather than standard iterator patterns, new concepts like rvalue references and move semantics are used widely, and so forth. People familiar with older versions of C++ can struggle to understand code written in new versions.
C++ is of course an enormously popular language, and the ongoing language revision process has not harmed its popularity.
Besides the lessons from C, I would add:
I know less about Java than about the other languages I discuss, so there may be more errors here and there are certainly more biases.
Java is largely backward compatible at the byte-code level, meaning that Java version N+1 libraries can call code written in, and compiled by, Java version N (and N-1, N-2, and so forth). Java source code is also mostly backward compatible, although they do add new keywords from time to time.
The Java documentation is very detailed about potential compatibility issues when moving from one release to another.
The Java standard library is enormous, and new packages are added at each new release. Packages are also deprecated from time to time. Using a deprecated package will cause a warning at compile time (the warning may be turned off), and after a few releases the deprecated package will be removed (at least in theory).
Java does not seem to have many backward compatibility problems. The problems are centered on the JVM: an older JVM generally will not run newer releases, so you have to make sure that your JVM is at least as new as that required by the newest library you want to use.
Java arguably has something of a forward compatibility problem in that JVM bytecodes present a higher level interface than that of a CPU, and that makes it harder to introduce new features that cannot be directly represented using the existing bytecodes.
This forward compatibility problem is part of the reason that Java generics use type erasure. Changing the definition of existing bytecodes would have broken existing programs that had already been compiled into bytecode. Extending bytecodes to support generic types would have required a large number of additional bytecodes to be defined.
This forward compatibility problem, to the extent that it is a problem, does not exist for Go. Since Go compiles to machine code, and implements all required run time checks by generating additional machine code, there is no similar forward compatibility issue.
But, in general:
Python 3.0 (also known as Python 3000) started development in 2006 and was initially released in 2008. In 2018 the transition is still incomplete. Some people continue to use Python 2.7 (released in 2010). This is not a path we want to emulate for Go 2.
The main reason for this slow transition appears to be lack of backward compatibility. Python 3.0 was intentionally incompatible with earlier versions of Python. Notably,
Because Python is an interpreted language, and because there is no backward compatibility, it is impossible to mix Python 2 and Python 3 code in the same program. This means that for a typical program that uses a range of libraries, each of those libraries must be converted to Python 3 before the program can be converted. Since programs are in various states of conversion, libraries must support Python 2 and 3 simultaneously.
Python supports statements of the form
from __future__ import FEATURE. A statement like this changes the interpretation of the rest of the file in some way. For example,
from __future__ import print_function changes
So, we knew it already, but:
The Perl 6 development process began in 2000. The first stable version of the Perl 6 spec was announced in 2015. This is not a path we want to emulate for Go 2.
There are many reasons for this slow path. Perl 6 was intentionally not backward compatible: it was meant to fix warts in the language. Perl 6 was intended to be represented by a spec rather than, as with previous versions of Perl, an implementation. Perl 6 started with a set of change proposals, but then continued to evolve over time, and then evolve some more.
use feature which is similar to Python's
from __future__ import. It changes the interpretation of the rest of the file to use a specified new language feature.
Pedantically speaking, we must have a way to speak about specific language versions. Each change to the Go language first appears in a Go release. We will use Go release numbers to define language versions. That is the only reasonable choice, but it can be confusing because standard library changes are also associated with Go release numbers. When thinking about compatibility, it will be necessary to conceptually separate the Go language version from the standard library version.
As an example of a specific change, type aliases were first available in Go language version 1.9. Type aliases were an example of a backward compatible language change. All code written in Go language versions 1.0 through 1.8 continued to work the same way with Go language 1.9. Code using type aliases requires Go language 1.9 or later.
Type aliases are an example of an addition to the language. Code using the type alias syntax
type A = B did not compile with Go versions before 1.9.
Type aliases, and other backward compatible changes since Go 1.0, show us that for additions to the language it is not necessary for packages to explicitly declare the minimum language version that they require. Some packages changed to use type aliases. When such a package was compiled with Go 1.8 tools, the package failed to compile. The package author can simply say: upgrade to Go 1.9, or downgrade to an earlier version of the package. None of the Go tools need to know about this requirement; it's implied by the failure to compile with older versions of the tools.
It's true of course that programmers need to understand language additions, but the the tooling does not. Neither the Go 1.8 tools nor the Go 1.9 tools need to explicitly know that type aliases were added in Go 1.9, other than in the limited sense that the Go 1.9 compiler will compile type aliases and the Go 1.8 compiler will not. That said, the possibility of specifying a minimum language version to get better error messages for unsupported language features is discussed below.
We must also consider language changes that simply remove features from the language. For example, issue 3939 proposes that we remove the conversion
string(i) for an integer value
i. If we make this change in, say, Go version 1.20, then packages that use this syntax will stop compiling in Go 1.20. (If you prefer to restrict backward incompatible changes to new major versions, then replace 1.20 by 2.0 in this discussion; the problem remains the same.)
In this case, packages using the old syntax have no simple recourse. While we can provide tooling to convert pre-1.20 code into working 1.20 code, we can't force package authors to run those tools. Some packages may be unmaintained but still useful. Some organizations may want to upgrade to 1.20 without having to requalify the versions of packages that they rely on. Some package authors may want to use 1.20 even though their packages now break, but do not have time to modify their package.
These scenarios suggest that we need a mechanism to specify the maximum version of the Go language with which a package can be built.
Importantly, specifying the maximum version of the Go language should not be taken to imply the maximum version of the Go tools. The Go compiler released with Go version 1.20 must be able to build packages using Go language 1.19. This can be done by adding an option to cmd/compile (and, if necessary, cmd/asm and cmd/link) along the lines of the
-std option supported by C compilers. When cmd/compile sees the option, perhaps
-lang=go1.19, it will compile the code using the Go 1.19 syntax.
This requires cmd/compile to support all previous versions, one way or another. If supporting old syntaxes proves to be troublesome, the
-lang option could perhaps be implemented by passing the code through a convertor from the old version to the current. That would keep support of old versions out of cmd/compile proper, and the convertor could be useful for people who want to update their code. But it is unlikely that supporting old language versions will be a significant problem.
Naturally, even though the package is built with the language version 1.19 syntax, it must in other respects be a 1.20 package: it must link with 1.20 code, be able to call and be called by 1.20 code, and so forth.
The go tool will need to know the maximum language version so that it knows how to invoke cmd/compile. Assuming we continue with the modules experiment, the logical place for this information is the go.mod file. The go.mod file for a module M can specify the maximum language version for the packages that it defines. This would be honored when M is downloaded as a dependency by some other module.
The maximum language version is not a minimum language version. If a module require features in language 1.19, but can be built with 1.20, we can say that the maximum language version is 1.20. If we build with Go release 1.19, we will see that we are at less than the maximum, and simply build with language version 1.19. Maximum language versions greater than that supported by the current tools can simply be ignored. If we later build with Go release 1.21, we will build the module with
This means that the tools can set the maximum language version automatically. When we use Go release 1.30 to release a module, we can mark the module as having maximum language version 1.30. All users of the module will see this maximum version and do the right thing.
This implies that we will have to support old versions of the language indefinitely. If we remove a language feature after version 1.25, version 1.26 and all later versions will still have to support that feature if invoked with the
-lang=go1.25 option (or
-lang=go1.24 or any other earlier version in which the feature is supported). Of course, if no
-lang option is used, or if the option is
-lang=go1.26 or later, the feature will not be available. Since we do not expect wholesale removals of existing language features, this should be a manageable burden.
I believe that this approach suffices for language removals.
For better error messages it may be useful to permit the module file to specify a minimum language version. This is not required: if a module uses features introduced in language version 1.N, then building it with 1.N-1 will fail at compile time. This may be confusing, but in practice it will likely be obvious what the problem is.
That said, if modules can specify a minimum language version, the go tool could produce an immediate, clear error message when building with 1.N-1.
The minimum language version could potentially be set by the compiler or some other tool. When compiling each file, see which features it uses, and use that to determine the minimum version. It need not be precisely accurate.
This is just a suggestion, not a requirement. It would likely provide a better user experience as the language changes.
The Go language can also change in ways that are not additions or removals, but are instead changes to the way a specific language construct works. For example, in Go 1.1 the size of the type
int on 64-bit hosts changed from 32 bits to 64 bits. This change was relatively harmless, as the language does not specify the exact size of
int. Potentially, though, some Go 1.0 programs continued to compile with Go 1.1 but stopped working.
A redefinition is a case where we have code that compiles successfully with both versions 1.N and version 1.M, where M > N, and where the meaning of the code is different in the two versions. For example, issue 20733 proposes that variables in a range loop should be redefined in each iteration. Though in practice this change seems more likely to fix programs than to break them, in principle this change might break working programs.
Note that a new keyword normally cannot cause a redefinition, though we must be careful to ensure that that is true before introducing one. For example, if we introduce the keyword
check as suggested in the error handling draft design, and we permit code like
check(f()), that might seem to be a redefinition if
check is defined as a function in the same package. But after the keyword is introduced, any attempt to define such a function will fail. So it is not possible for code using
check, under whichever meaning, to compile with both version 1.N and 1.M. The new keyword can be handled as a removal (of the non-keyword use of
check) and an addition (of the keyword
In order for the Go ecosystem to survive a transition to Go 2, we must minimize these sorts of redefinitions. As discussed earlier, successful languages have generally had essentially no redefinitions beyond a certain point.
The complexity of a redefinition is, of course, that we can no longer rely on the compiler to detect the problem. When looking at a redefined language construct, the compiler cannot know which meaning is meant. In the presence of redefined language constructs, we cannot determine the maximum language version. We don't know if the construct is intended to be compiled with the old meaning or the new.
The only possibility would be to let programmers set the language version. In this case it would be either a minimum or maximum language version, as appropriate. It would have to be set in such a way that it would not be automatically updated by any tools. Of course, setting such a version would be error prone. Over time, a maximum language version would lead to surprising results, as people tried to use new language features, and failed.
I think the only feasible safe approach is to not permit language redefinitions.
We are stuck with our current semantics. This doesn‘t mean we can’t improve them. For example, for issue 20733, the range issue, we could change range loops so that taking the address of a range parameter, or referring to it from a function literal, is forbidden. This would not be a redefinition; it would be a removal. That approach might eliminate the bugs without the potential of breaking code unexpectedly.
Build tags are an existing mechanism that can be used by programs to choose which files to compile based on the release.
Build tags name release versions, which look just like language versions, but, speaking pedantically, are different. In the discussion above we've talked about using Go release 1.N to compile code with language version 1.N-1. That is not possible using build tags.
Build tags can be used to set the maximum or a minimum release, or both, that will be used to compile a specific file. They can be a convenient way to take advantage of language changes that are only available after a certain version; that is, they can be used to set a minimum language version when compiling a file.
As discussed above, though, what is most useful for language changes is the ability to set a maximum language version. Build tags don‘t provide that in a useful way. If you use a build tag to set your current release version as your maximum version, your package will not build with later releases. Setting a maximum language version is only possible when it is set to a version before the current release, and is coupled with an alternate implementation that is used for the later versions. That is, if you are building with 1.N, it’s not helpful to use a build tag of
!1.N+1. You could use a build tag of
M < N, but in almost all cases you will then need a separate file with a build tag of
Build tags can be used to handle language redefinitions: if there is a language redefinition at language version
1.N, programmers can write one file with a build tag of
!1.N using the old semantics and a different file with a build tag of
1.N using the new semantics. However, these duplicate implementations are a lot of work, it's hard to know in general when it is required, and it would be easy to make a mistake. The availability of build tags is not enough to overcome the earlier comments about not permitting any language redefinitions.
It would be possible to add a mechanism to Go similar to Python‘s
from __future__ import and Perl’s
use feature. For example, we could use a special import path, such as
import "go2/type-aliases". This would put the required language features in the file that uses them, rather than hidden away in the go.mod file.
This would provide a way to describe the set of language additions required by the file. It's more complicated, because instead of relying on a language version, the language is broken up into separate features. There is no obvious way to ever remove any of these special imports, so they will tend to accumulate over time. Python and Perl avoid the accumulation problem by intentionally making a backward incompatible change. After moving to Python 3 or Perl 6, the accumulated feature requests can be discarded. Since Go is trying to avoid a large backward incompatible change, there would be no clear way to ever remove these imports.
This mechanism does not address language removals. We could introduce a removal import, such as
import "go2/no-int-to-string", but it's not obvious why anyone would ever use it. In practice, there would be no way to ever remove language features, even ones that are confusing and error-prone.
This kind of approach doesn't seem suitable for Go.
One of the benefits of a Go 2 transition is the chance to release some of the standard library packages from the Go 1 compatibility guarantee. Another benefit is the chance to move many, perhaps most, of the packages out of the six month release cycle. If the modules experiment works out it may even be possible to start doing this sooner rather than later, with some packages on a faster cycle.
I propose that the six month release cycle continue, but that it be treated as a compiler/runtime release cycle. We want Go releases to be useful out of the box, so releases will continue to include the current versions of roughly the same set of packages that they contain today. However, many of those packages will actually be run on their own release cycles. People using a given Go release will be able to explicitly choose to use newer versions of the standard library packages. In fact, in some cases they may be able to use older versions of the standard library packages where that seems useful.
Different release cycles would require more resources on the part of the package maintainers. We can only do this if we have enough people to manage it and enough testing resources to test it.
We could also continue using the six month release cycle for everything, but make the separable packages available separately for use with different, compatible, releases.
Still, some parts of the standard library must be treated as core libraries. These libraries are closely tied to the compiler and other tools, and must strictly follow the release cycle. Neither older nor newer versions of these libraries may be used.
Ideally, these libraries will remain on the current version 1. If it seems necessary to change any of them to version 2, that will have to be discussed on a case by case basis. At this time I see no reason for it.
The tentative list of core libraries is:
I am, perhaps optimistically, omitting the net, os, and syscall packages from this list. We'll see what we can manage.
The penumbra standard library consists of those packages that are included with a release but are maintained independently. This will be most of the current standard library. These packages will follow the same discipline as today, with the option to move to a v2 where appropriate. It will be possible to use
go get to upgrade or, possibly, downgrade these standard library packages. In particular, fixes can be made as minor releases separately from the six month core library release cycle.
The go tool will have to be able to distinguish between the core library and the penumbra library. I don't know precisely how this will work, but it seems feasible.
When moving a standard library package to v2, it will be essential to plan for programs that use both v1 and v2 of the package. Those programs will have to work as expected, or if that is impossible will have to fail cleanly and quickly. In some cases this will involve modifying the v1 version to use an internal package that is also shared by the v2 package.
Standard library packages will have to compile with older versions of the language, at least the two previous release cycles that we currently support.
The ability to support
go get of standard library packages will permit us to remove packages from the releases. Those packages will continue to exist and be maintained, and people will be able to retrieve them if they need them. However, they will not be shipped by default with a Go release.
This will include packages like
and perhaps other packages that do not seem to be widely useful.
We should in due course plan a deprecation policy for old packages, to move these packages to a point where they are no longer maintained. The deprecation policy will also apply to the v1 versions of packages that move to v2.
Or this may prove to be too problematic, and we should never deprecate any existing package, and never remove them from the standard releases.
If the above process works as planned, then in an important sense there never will be a Go 2. Or, to put it a different way, we will slowly transition to new language and library features. We could at any point during the transition decide that now we are Go 2, which might be good marketing. Or we could just skip it (there has never been a C 2.0, why have a Go 2.0?).
Popular languages like C, C++, and Java never have a version 2. In effect, they are always at version 1.N, although they use different names for that state. I believe that we should emulate them. In truth, a Go 2 in the full sense of the word, in the sense of an incompatible new version of the language or core libraries, would not be a good option for our users. A real Go 2 would, perhaps unsurprisingly, be harmful.