Go 博客

Go, Open Source, Community

2015/07/08

Welcome

[This is the text of my opening keynote at Gophercon 2015. The video is available here.]

Thank you all for traveling to Denver to be here, and thank you to everyone watching on video. If this is your first Gophercon, welcome. If you were here last year, welcome back. Thank you to the organizers for all the work it takes to make a conference like this happen. I am thrilled to be here and to be able to talk to all of you.

I am the tech lead for the Go project and the Go team at Google. I share that role with Rob Pike. In that role, I spend a lot of time thinking about the overall Go open source project, in particular the way it runs, what it means to be open source, and the interaction between contributors inside and outside Google. Today I want to share with you how I see the Go project as a whole and then based on that explain how I see the Go open source project evolving.

Why Go?

To get started, we have to go back to the beginning. Why did we start working on Go?

Go is an attempt to make programmers more productive. We wanted to improve the software development process at Google, but the problems Google has are not unique to Google.

There were two overarching goals.

The first goal is to make a better language to meet the challenges of scalable concurrency. By scalable concurrency I mean software that deals with many concerns simultaneously, such as coordinating a thousand back end servers by sending network traffic back and forth.

Today, that kind of software has a shorter name: we call it cloud software. It's fair to say that Go was designed for the cloud before clouds ran software.

The larger goal is to make a better environment to meet the challenges of scalable software development, software worked on and used by many people, with limited coordination between them, and maintained for years. At Google we have thousands of engineers writing and sharing their code with each other, trying to get their work done, reusing the work of others as much as possible, and working in a code base with a history dating back over ten years. Engineers often work on or at least look at code originally written by someone else, or that they wrote years ago, which often amounts to the same thing.

That situation inside Google has a lot in common with large scale, modern open source development as practiced on sites like GitHub. Because of this, Go is a great fit for open source projects, helping them accept and manage contributions from a large community over a long period of time.

I believe much of Go's success is explained by the fact that Go is a great fit for cloud software, Go is a great fit for open source projects, and, serendipitously, both of those are growing in popularity and importance in the software industry.

Other people have made similar observations. Here are two. Last year, on RedMonk.com, Donnie Berkholz wrote about “Go as the emerging language of cloud infrastructure,” observing that “[Go's] marquee projects ... are cloud-centric or otherwise made for dealing with distributed systems or transient environments.”

This year, on Texlution.com, the author wrote an article titled “Why Golang is doomed to succeed,” pointing out that this focus on large-scale development was possibly even better suited to open source than to Google itself: “This open source fitness is why I think you are about to see more and more Go around ...”

The Go Balance

How does Go accomplish those things?

How does it make scalable concurrency and scalable software development easier?

Most people answer this question by talking about channels and goroutines, and interfaces, and fast builds, and the go command, and good tool support. Those are all important parts of the answer, but I think there is a broader idea behind them.

I think of that idea as Go's balance. There are competing concerns in any software design, and there is a very natural tendency to try to solve all the problems you foresee. In Go, we have explicitly tried not to solve everything. Instead, we've tried to do just enough that you can build your own custom solutions easily.

The way I would summarize Go's chosen balance is this: Do Less. Enable More.

Do less, but enable more.

Go can't do everything. We shouldn't try. But if we work at it, Go can probably do a few things well. If we select those things carefully, we can lay a foundation on which developers can easily build the solutions and tools they need, and ideally can interoperate with the solutions and tools built by others.

Examples

Let me illustrate this with some examples.

First, the size of the Go language itself. We worked hard to put in as few concepts as possible, to avoid the problem of mutually incomprehensible dialects forming in different parts of a large developer community. No idea went into Go until it had been simplified to its essence and then had clear benefits that justified the complexity being added.

In general, if we have 100 things we want Go to do well, we can't make 100 separate changes. Instead, we try to research and understand the design space and then identify a few changes that work well together and that enable maybe 90 of those things. We're willing to sacrifice the remaining 10 to avoid bloating the language, to avoid adding complexity only to address specific use cases that seem important today but might be gone tomorrow.

Keeping the language small enables more important goals. Being small makes Go easier to learn, easier to understand, easier to implement, easier to reimplement, easier to debug, easier to adjust, and easier to evolve. Doing less enables more.

I should point out that this means we say no to a lot of other people's ideas, but I assure you we've said no to even more of our own ideas.

Next, channels and goroutines. How should we structure and coordinate concurrent and parallel computations? Mutexes and condition variables are very general but so low-level that they're difficult to use correctly. Parallel execution frameworks like OpenMP are so high-level that they can only be used to solve a narrow range of problems. Channels and goroutines sit between these two extremes. By themselves, they aren't a solution to much. But they are powerful enough to be easily arranged to enable solutions to many common problems in concurrent software. Doing less—really doing just enough—enables more.

Next, types and interfaces. Having static types enables useful compile-time checking, something lacking in dynamically-typed languages like Python or Ruby. At the same time, Go's static typing avoids much of the repetition of traditional statically typed languages, making it feel more lightweight, more like the dynamically-typed languages. This was one of the first things people noticed, and many of Go's early adopters came from dynamically-typed languages.

Go's interfaces are a key part of that. In particular, omitting the ``implements'' declarations of Java or other languages with static hierarchy makes interfaces lighter weight and more flexible. Not having that rigid hierarchy enables idioms such as test interfaces that describe existing, unrelated production implementations. Doing less enables more.

Next, testing and benchmarking. Is there any shortage of testing and benchmarking frameworks in most languages? Is there any agreement between them?

Go's testing package is not meant to address every possible facet of these topics. Instead, it is meant to provide the basic concepts necessary for most higher-level tooling. Packages have test cases that pass, fail, or are skipped. Packages have benchmarks that run and can be measured by various metrics.

Doing less here is an attempt to reduce these concepts to their essence, to create a shared vocabulary so that richer tools can interoperate. That agreement enables higher-level testing software like Miki Tebeka's go2xunit converter, or the benchcmp and benchstat benchmark analysis tools.

Because there is agreement about the representation of the basic concepts, these higher-level tools work for all Go packages, not just ones that make the effort to opt in, and they interoperate with each other, in that using, say, go2xunit does not preclude also using benchstat, the way it would if these tools were, say, plugins for competing testing frameworks. Doing less enables more.

Next, refactoring and program analysis. Because Go is for large code bases, we knew it would need to support automatic maintenance and updating of source code. We also knew that this topic was too large to build in directly. But we knew one thing that we had to do. In our experience attempting automated program changes in other settings, the most significant barrier we hit was actually writing the modified program out in a format that developers can accept.

In other languages, it's common for different teams to use different formatting conventions. If an edit by a program uses the wrong convention, it either writes a section of the source file that looks nothing like the rest of the file, or it reformats the entire file, causing unnecessary and unwanted diffs.

Go does not have this problem. We designed the language to make gofmt possible, we worked hard to make gofmt's formatting acceptable for all Go programs, and we made sure gofmt was there from day one of the original public release. Gofmt imposes such uniformity that automated changes blend into the rest of the file. You can't tell whether a particular change was made by a person or a computer. We didn't build explicit refactoring support. Establishing an agreed-upon formatting algorithm was enough of a shared base for independent tools to develop and to interoperate. Gofmt enabled gofix, goimports, eg, and other tools. I believe the work here is only just getting started. Even more can be done.

Last, building and sharing software. In the run up to Go 1, we built goinstall, which became what we all know as "go get". That tool defined a standard zero-configuration way to resolve import paths on sites like github.com, and later a way to resolve paths on other sites by making HTTP requests. This agreed-upon resolution algorithm enabled other tools that work in terms of those paths, most notably Gary Burd's creation of godoc.org. In case you haven't used it, you go to godoc.org/the-import-path for any valid "go get" import path, and the web site will fetch the code and show you the documentation for it. A nice side effect of this has been that godoc.org serves as a rough master list of the Go packages publicly available. All we did was give import paths a clear meaning. Do less, enable more.

You'll notice that many of these tooling examples are about establishing a shared convention. Sometimes people refer to this as Go being “opinionated,” but there's something deeper going on. Agreeing to the limitations of a shared convention is a way to enable a broad class of tools that interoperate, because they all speak the same base language. This is a very effective way to do less but enable more. Specifically, in many cases we can do the minimum required to establish a shared understanding of a particular concept, like remote imports, or the proper formatting of a source file, and thereby enable the creation of packages and tools that work together because they all agree about those core details.

I'm going to return to that idea later.

Why is Go open source?

But first, as I said earlier, I want to explain how I see the balance of Do Less and Enable More guiding our work on the broader Go open source project. To do that, I need to start with why Go is open source at all.

Google pays me and others to work on Go, because, if Google's programmers are more productive, Google can build products faster, maintain them more easily, and so on. But why open source Go? Why should Google share this benefit with the world?

Of course, many of us worked on open source projects before Go, and we naturally wanted Go to be part of that open source world. But our preferences are not a business justification. The business justification is that Go is open source because that's the only way that Go can succeed. We, the team that built Go within Google, knew this from day one. We knew that Go had to be made available to as many people as possible for it to succeed.

Closed languages die.

A language needs large, broad communities.

A language needs lots of people writing lots of software, so that when you need a particular tool or library, there's a good chance it has already been written, by someone who knows the topic better than you, and who spent more time than you have to make it great.

A language needs lots of people reporting bugs, so that problems are identified and fixed quickly. Because of the much larger user base, the Go compilers are much more robust and spec-compliant than the Plan 9 C compilers they're loosely based on ever were.

A language needs lots of people using it for lots of different purposes, so that the language doesn't overfit to one use case and end up useless when the technology landscape changes.

A language needs lots of people who want to learn it, so that there is a market for people to write books or teach courses, or run conferences like this one.

None of this could have happened if Go had stayed within Google. Go would have suffocated inside Google, or inside any single company or closed environment.

Fundamentally, Go must be open, and Go needs you. Go can't succeed without all of you, without all the people using Go for all different kinds of projects all over the world.

In turn, the Go team at Google could never be large enough to support the entire Go community. To keep scaling, we need to enable all this ``more'' while doing less. Open source is a huge part of that.

Go's open source

What does open source mean? The minimum requirement is to open the source code, making it available under an open source license, and we've done that.

But we also opened our development process: since announcing Go, we've done all our development in public, on public mailing lists open to all. We accept and review source code contributions from anyone. The process is the same whether you work for Google or not. We maintain our bug tracker in public, we discuss and develop proposals for changes in public, and we work toward releases in public. The public source tree is the authoritative copy. Changes happen there first. They are only brought into Google's internal source tree later. For Go, being open source means that this is a collective effort that extends beyond Google, open to all.

Any open source project starts with a few people, often just one, but with Go it was three: Robert Griesemer, Rob Pike, and Ken Thompson. They had a vision of what they wanted Go to be, what they thought Go could do better than existing languages, and Robert will talk more about that tomorrow morning. I was the next person to join the team, and then Ian Taylor, and then, one by one, we've ended up where we are today, with hundreds of contributors.

Thank You to the many people who have contributed code or ideas or bug reports to the Go project so far. We tried to list everyone we could in our space in the program today. If your name is not there, I apologize, but thank you.

I believe the hundreds of contributors so far are working toward a shared vision of what Go can be. It's hard to put words to these things, but I did my best to explain one part of the vision earlier: Do Less, Enable More.

Google's role

A natural question is: What is the role of the Go team at Google, compared to other contributors? I believe that role has changed over time, and it continues to change. The general trend is that over time the Go team at Google should be doing less and enabling more.

In the very early days, before Go was known to the public, the Go team at Google was obviously working by itself. We wrote the first draft of everything: the specification, the compiler, the runtime, the standard library.

Once Go was open sourced, though, our role began to change. The most important thing we needed to do was communicate our vision for Go. That's difficult, and we're still working at it.. The initial implementation was an important way to communicate that vision, as was the development work we led that resulted in Go 1, and the various blog posts, and articles, and talks we've published.

But as Rob said at Gophercon last year, "the language is done." Now we need to see how it works, to see how people use it, to see what people build. The focus now is on expanding the kind of work that Go can help with.

Google's primarily role is now to enable the community, to coordinate, to make sure changes work well together, and to keep Go true to the original vision.

Google's primary role is: Do Less. Enable More.

I mentioned earlier that we'd rather have a small number of features that enable, say, 90% of the target use cases, and avoid the orders of magnitude more features necessary to reach 99 or 100%. We've been successful in applying that strategy to the areas of software that we know well. But if Go is to become useful in many new domains, we need experts in those areas to bring their expertise to our discussions, so that together we can design small adjustments that enable many new applications for Go.

This shift applies not just to design but also to development. The role of the Go team at Google continues to shift more to one of guidance and less of pure development. I certainly spend much more time doing code reviews than writing code, more time processing bug reports than filing bug reports myself. We need to do less and enable more.

As design and development shift to the broader Go community, one of the most important things we the original authors of Go can offer is consistency of vision, to help keep Go Go. The balance that we must strike is certainly subjective. For example, a mechanism for extensible syntax would be a way to enable more ways to write Go code, but that would run counter to our goal of having a consistent language without different dialects.

We have to say no sometimes, perhaps more than in other language communities, but when we do, we aim to do so constructively and respectfully, to take that as an opportunity to clarify the vision for Go.

Of course, it's not all coordination and vision. Google still funds Go development work. Rick Hudson is going to talk later today about his work on reducing garbage collector latency, and Hana Kim is going to talk tomorrow about her work on bringing Go to mobile devices. But I want to make clear that, as much as possible, we aim to treat development funded by Google as equal to development funded by other companies or contributed by individuals using their spare time. We do this because we don't know where the next great idea will come from. Everyone contributing to Go should have the opportunity to be heard.

Examples

I want to share some evidence for this claim that, over time, the original Go team at Google is focusing more on coordination than direct development.

First, the sources of funding for Go development are expanding. Before the open source release, obviously Google paid for all Go development. After the open source release, many individuals started contributing their time, and we've slowly but steadily been growing the number of contributors supported by other companies to work on Go at least part-time, especially as it relates to making Go more useful for those companies. Today, that list includes Canonical, Dropbox, Intel, Oracle, and others. And of course Gophercon and the other regional Go conferences are organized entirely by people outside Google, and they have many corporate sponsors besides Google.

Second, the conceptual depth of Go development done outside the original team is expanding.

Immediately after the open source release, one of the first large contributions was the port to Microsoft Windows, started by Hector Chu and completed by Alex Brainman and others. More contributors ported Go to other operating systems. Even more contributors rewrote most of our numeric code to be faster or more precise or both. These were all important contributions, and very much appreciated, but for the most part they did not involve new designs.

More recently, a group of contributors led by Aram Hăvărneanu ported Go to the ARM 64 architecture, This was the first architecture port by contributors outside Google. This is significant, because in general support for a new architecture requires more design work than support for a new operating system. There is more variation between architectures than between operating systems.

Another example is the introduction over the past few releases of preliminary support for building Go programs using shared libraries. This feature is important for many Linux distributions but not as important for Google, because we deploy static binaries. We have been helping guide the overall strategy, but most of the design and nearly all of the implementation has been done by contributors outside Google, especially Michael Hudson-Doyle.

My last example is the go command's approach to vendoring. I define vendoring as copying source code for external dependencies into your tree to make sure that they doesn't disappear or change underfoot.

Vendoring is not a problem Google suffers, at least not the way the rest of the world does. We copy open source libraries we want to use into our shared source tree, record what version we copied, and only update the copy when there is a need to do so. We have a rule that there can only be one version of a particular library in the source tree, and it's the job of whoever wants to upgrade that library to make sure it keeps working as expected by the Google code that depends on it. None of this happens often. This is the lazy approach to vendoring.

In contrast, most projects outside Google take a more eager approach, importing and updating code using automated tools and making sure that they are always using the latest versions.

Because Google has relatively little experience with this vendoring problem, we left it to users outside Google to develop solutions. Over the past five years, people have built a series of tools. The main ones in use today are Keith Rarick's godep, Owen Ou's nut, and the gb-vendor plugin for Dave Cheney's gb,

There are two problems with the current situation. The first is that these tools are not compatible out of the box with the go command's "go get". The second is that the tools are not even compatible with each other. Both of these problems fragment the developer community by tool.

Last fall, we started a public design discussion to try to build consensus on some basics about how these tools all operate, so that they can work alongside "go get" and each other.

Our basic proposal was that all tools agree on the approach of rewriting import paths during vendoring, to fit with "go get"'s model, and also that all tools agree on a file format describing the source and version of the copied code, so that the different vendoring tools can be used together even by a single project. If you use one today, you should still be able to use another tomorrow.

Finding common ground in this way was very much in the spirit of Do Less, Enable More. If we could build consensus about these basic semantic aspects, that would enable "go get" and all these tools to interoperate, and it would enable switching between tools, the same way that agreement about how Go programs are stored in text files enables the Go compiler and all text editors to interoperate. So we sent out our proposal for common ground.

Two things happened.

First, Daniel Theophanes started a vendor-spec project on GitHub with a new proposal and took over coordination and design of the spec for vendoring metadata.

Second, the community spoke with essentially one voice to say that rewriting import paths during vendoring was not tenable. Vendoring works much more smoothly if code can be copied without changes.

Keith Rarick posted an alternate proposal for a minimal change to the go command to support vendoring without rewriting import paths. Keith's proposal was configuration-free and fit in well with the rest of the go command's approach. That proposal will ship as an experimental feature in Go 1.5 and likely enabled by default in Go 1.6. And I believe that the various vendoring tool authors have agreed to adopt Daniel's spec once it is finalized.

The result is that at the next Gophercon we should have broad interoperability between vendoring tools and the go command, and the design to make that happen was done entirely by contributors outside the original Go team.

Not only that, the Go team's proposal for how to do this was essentially completely wrong. The Go community told us that very clearly. We took that advice, and now there's a plan for vendoring support that I believe everyone involved is happy with.

This is also a good example of our general approach to design. We try not to make any changes to Go until we feel there is broad consensus on a well-understood solution. For vendoring, feedback and design from the Go community was critical to reaching that point.

This general trend toward both code and design coming from the broader Go community is important for Go. You, the broader Go community, know what is working and what is not in the environments where you use Go. We at Google don't. More and more, we will rely on your expertise, and we will try to help you develop designs and code that extend Go to be useful in more settings and fit well with Go's original vision. At the same time, we will continue to wait for broad consensus on well-understood solutions.

This brings me to my last point.

Code of Conduct

I've argued that Go must be open, and that Go needs your help.

But in fact Go needs everyone's help. And everyone isn't here.

Go needs ideas from as many people as possible.

To make that a reality, the Go community needs to be as inclusive, welcoming, helpful, and respectful as possible.

The Go community is large enough now that, instead of assuming that everyone involved knows what is expected, I and others believe that it makes sense to write down those expectations explicitly. Much like the Go spec sets expectations for all Go compilers, we can write a spec setting expectations for our behavior in online discussions and in offline meetings like this one.

Like any good spec, it must be general enough to allow many implementations but specific enough that it can identify important problems. When our behavior doesn't meet the spec, people can point that out to us, and we can fix the problem. At the same time, it's important to understand that this kind of spec cannot be as precise as a language spec. We must start with the assumption that we will all be reasonable in applying it.

This kind of spec is often referred to as a Code of Conduct. Gophercon has one, which we've all agreed to follow by being here, but the Go community does not. I and others believe the Go community needs a Code of Conduct.

But what should it say?

I believe the most important overall statement we can make is that if you want to use or discuss Go, then you are welcome here, in our community. That is the standard I believe we aspire to.

If for no other reason (and, to be clear, there are excellent other reasons), Go needs as large a community as possible. To the extent that behavior limits the size of the community, it holds Go back. And behavior can easily limit the size of the community.

The tech community in general and the Go community in particular is skewed toward people who communicate bluntly. I don't believe this is fundamental. I don't believe this is necessary. But it's especially easy to do in online discussions like email and IRC, where plain text is not supplemented by the other cues and signals we have in face-to-face interactions.

For example, I have learned that when I am pressed for time I tend to write fewer words, with the end result that my emails seem not just hurried but blunt, impatient, even dismissive. That's not how I feel, but it's how I can come across, and that impression can be enough to make people think twice about using or contributing to Go. I realized I was doing this when some Go contributors sent me private email to let me know. Now, when I am pressed for time, I pay extra attention to what I'm writing, and I often write more than I naturally would, to make sure I'm sending the message I intend.

I believe that correcting the parts of our everyday interactions, intended or not, that drive away potential users and contributors is one of the most important things we can all do to make sure the Go community continues to grow. A good Code of Conduct can help us do that.

We have no experience writing a Code of Conduct, so we have been reading existing ones, and we will probably adopt an existing one, perhaps with minor adjustments. The one I like the most is the Django Code of Conduct, which originated with another project called SpeakUp! It is structured as an elaboration of a list of reminders for everyday interaction.

"Be friendly and patient. Be welcoming. Be considerate. Be respectful. Be careful in the words that you choose. When we disagree, try to understand why."

I believe this captures the tone we want to set, the message we want to send, the environment we want to create for new contributors. I certainly want to be friendly, patient, welcoming, considerate, and respectful. I won't get it exactly right all the time, and I would welcome a helpful note if I'm not living up to that. I believe most of us feel the same way.

I haven't mentioned active exclusion based on or disproportionately affecting race, gender, disability, or other personal characteristics, and I haven't mentioned harassment. For me, it follows from what I just said that exclusionary behavior or explicit harassment is absolutely unacceptable, online and offline. Every Code of Conduct says this explicitly, and I expect that ours will too. But I believe the SpeakUp! reminders about everyday interactions are an equally important statement. I believe that setting a high standard for those everyday interactions makes extreme behavior that much clearer and easier to deal with.

I have no doubts that the Go community can be one of the most friendly, welcoming, considerate, and respectful communities in the tech industry. We can make that happen, and it will be a benefit and credit to us all.

Andrew Gerrand has been leading the effort to adopt an appropriate Code of Conduct for the Go community. If you have suggestions, or concerns, or experience with Codes of Conduct, or want to be involved, please find Andrew or me during the conference. If you'll still be here on Friday, Andrew and I are going to block off some time for Code of Conduct discussions during Hack Day.

Again, we don't know where the next great idea will come from. We need all the help we can get. We need a large, diverse Go community.

Thank You

I consider the many people releasing software for download using “go get,” sharing their insights via blog posts, or helping others on the mailing lists or IRC to be part of this broad open source effort, part of the Go community. Everyone here today is also part of that community.

Thank you in advance to the presenters who over the next few days will take time to share their experiences using and extending Go.

Thank you in advance to all of you in the audience for taking the time to be here, to ask questions, and to let us know how Go is working for you. When you go back home, please continue to share what you've learned. Even if you don't use Go for daily work, we'd love to see what's working for Go adopted in other contexts, just as we're always looking for good ideas to bring back into Go.

Thank you all again for making the effort to be here and for being part of the Go community.

For the next few days, please: tell us what we're doing right, tell us what we're doing wrong, and help us all work together to make Go even better.

Remember to be friendly, patient, welcoming, considerate, and respectful.

Above all, enjoy the conference.

Russ Cox 编写