Space Monkey dumps Python for Go | Hacker News Reader

Space Monkey dumps Python for Go

spacemonkey.com

338 pointsjtolds204 comments

This article doesn't address one of the biggest benefits I've noticed in switching from Python to Go.

I've been using Go for almost two years now. Like OP, I am/was primarily a Python developer. Like OP, my first Go project was a time-sensitive rewrite[0] of a project from Python (tornado) to Go.

Even though I was an experienced Python developer, the Go rewrite was marginally (~20%) faster[1]. But the real benefit came from the subsequent development - refactoring, rearchitecting, and maintaining the project on an ongoing basis. Go was designed to make it easy to scale the maintenance[2] of software, and on this one axis, it absolutely blows every other language and environment I've used out of the water.

For a fresh project, I'd say Go is about 10% slower to write than the equivalent Python[3] for the average case. But the time/cost savings are very quick to come thereafter.

[0] I would absolutely not recommend doing time-sensitive rewrites in general, but that decision was a separate matter.

[1] Some of this is due to the nature of rewrites in general, but the fact that it wasn't slower to use a language I'd never used before says something about the language.

[2] Scaling software development as teams grow is very different from scaling software as users grow.

[3] Assuming comparable familiarity with both languages, which is rarely the case, of course

This sounds really interesting, but can you elaborate on how Go scales better with maintenance?

(disclosure: I haven't fiddled with Go at all yet, so I know nothing about the language itself)

Several things make it easier to maintain:

- super fast compile times for fast developer iterations

- you can create an interface (set of methods) that your module owns, and apply it to objects created by other modules without recompiling. If you only need one method you can take that one. It encourages decoupling you from your dependencies.

- the code has one correct formatting convention and a tool that will auto-format your code

- many complex rules for numeric conversions that are implicit in C and cause no end of trouble are explicit and much simpler in Go.

- treating concurrency as a series of sequential process connected with channels makes it MUCH easier to reason about.

Don't the first 3 apply just as well to python? It compiles fast, and is duck-typed so interfaces are always there and implicit, and the indent-based blocking ends many arguments about formatting.

the difference is that if you're not careful when you cut-paste indented blocks in python, you can easily change logic.

If you cut and paste probably the misaligned indentation is the least problem. Cut and paste should be avoided if it's not done during a refactoring (so actual moving of code for better design).

Beginners usually think Python strong indentation is a weakness of the language. I actually find C++ freedom being more error prone:

    if (a < b);
        a = b

It's a not very frequent bug, but when it happens it takes you hours to spot. ;)

Yes and in Go (and hopefully all curly-brace languages of the future) this is actually a syntax error, you need the braces:

  if (a < b) {
     a = b;
  }

And you don't need the semicolons!

The thing I don't get...if braces are mandatory (Good), why keep the now completely, unambiguously, irrelevant () around the cond?

They're actually not required in Go.

You don't need those parens in Go, and go fmt will in fact remove them/

If you ran the code through gofmt, it would remove the () around the cond for you. I code go in SublimeText with GoSublime. On every save it runs the file through gofmt and reformats it for me. Keeps my code looking pretty with very little effort.

This would never happen to me because I frequently format my doc, and it wouldn't be indented.

Python executes, it does not compiles.

EDIT: Whoops, it's actually compiled into bytecode then executed by the VM.

This is not true.

As the_paul said, this is incorrect.

https://en.wikipedia.org/wiki/Python_(programming_language)#...

It actually "compiles" into bytecode, then consumed by the interpreter.

it's generous to call the CPython interpreter a VM - the binary encoding of Python isn't some crazy IL bytecode, its' really just python-as-binary. Language constructs converted into opcodes, strings with pre-calculated hashes, and local variables within a scope become a sort of vector... but otherwise, it's a prettymuch 1:1 mapping between Python language constructs and the bytecode form.

I thought you had meant that Go was "easier to maintain" than python - most of these don't apply to Python: Python has faster compilation times (0); interfaces without recompiling; formatting conventions; and numeric conversions is compared explicitly with C. Channels is the only one that applies to python.

Maybe he meant that having a typed language that is checked at compilation time is a huge plus for maintenance, and the usual downsides that come with compilation (mainly slowness) are pretty much inexistant with Go.

As some who generally prefers Python to Go quite strongly, the formatting situation is no where NEAR equivalent. With go it's trivial to setup your editor to perfectly reformat every time you save, or even on carriage return, and it does in a way that is 100% consistent and will never alter the logic of your code.

I'm not a python expert by any measure, but I've written a couple small projects with it. Python surprises for how well documented it is. In this specific case, PEP8 lays down most formatting conventions, so you'd expect some tooling to be available. Is this not the case?

A quick Google search brought up this, for instance:https://bitbucket.org/StephaneBunel/pythonpep8autoformat

The thing is, while this kind of thing is available for pretty much every language, it is standard for Go. As in, if you don't gofmt your code, everyone who looks at it will bug you to format it. That's a huge difference. There's basically zero code in the wild that you'd ever want to use that /isn't/ gofmted.

Exactly the point. Go is a VERY opinionated language. You see the term "idiomatic Go" used very regularly. The Go team and community push hard for everyone to follow a set way of doing things. This means all Go code must be formatted a certain way, handle errors in a certain way, etc. People even put pre-commit hooks into their VCS system that will reject .go commits unless they are properly formatted before commit.

I use Vim and there's a plugin just for that: making sure my code is PEP8 compliant.

Python is interpreted, not compiled... Of course it has 0 compile time.

I think he mean one single big ass executable is way easier to maintain than pip install & virtualenv everywhere.

Yes we can also make a single big ass executable out of any python code but it's slower and not supported out of the box.

And IMHO that's whats killing python - the fragmentation.

You are incorrect about compile times in many ways. First- python byte compiles to a weird VM. Then it executes that. Second- go compilation times are comparable to python byte compilation time but the result is a native executable.

Would you please expand on [2] please with concrete examples.

I also rewrote a tornado service in Go. The thing I found most surprising was that the length of the program was about the same as the Python version. I had expected it to be quite a bit longer.

I'm curious, how does it compare to Java with a good IDE (eg. Eclipse or IntelliJ)?

One thing I really missed when I abandoned the Java world and went to C++/Python was the refactoring and formatting tools available in the IDE. I could set it up so it'd auto-format my code on save, so that I could rename methods with a single keystroke, so that I could pull out classes or add getters/setters, so that I could add parameters, etc. I know Go supports some of that through gofmt and go fix, but I'm curious if anyone is both an experienced Java dev and an experienced Go dev and can compare the continued maintenance costs of the two.

BTW, I would put the productivity premium of Go over Python at about 50%-2x, not 10%. I did a mid-size (~6 man-months, though much of that time went into interactions with other teams) green-field prototype in Go at Google, and found myself really missing constructs like list comprehensions, dynamic typing, strong support for literal data, ability to treat user-defined types just like the built-ins, etc. I do think you'll make some of that back in continued maintenance - I actually really enjoyed working with Go - but I didn't experience the claims of "Go is just as good as Python for prototyping", as someone with a fair amount of Python experience.

Doesn't Go come with an executable that reformats .go files to a canonical format?

That's the gofmt tool that we're all talking about.

The availible go ide's are still young but go includes a lot of really nice tooling in the core distribution. The gofmt tool enforces a canonical format for go code and I have sublime text set up (via gosublime) to reformat on save and use the gocode autocomplete/linting daemnon.

Gofmt is actually one of my favorite parts about go and I feel that it, along with the simple documentation system and godoc.org save me a ton of time making my code ready for others to read/use.

The code base behind gofmt and gofix can also be used for other tools, i haven't felt the need for a powerful refactoring tool but I'm sure one will show up eventually.

I do miss list comprehensions and operator overloading but I understand the dev's arguments against those.

missing constructs like list comprehensions, dynamic typing, strong support for literal data, ability to treat user-defined types just like the built-ins, etc.

Of those, I only see comprehensions as something that's severely lacking, though I am not a Python guru by any measure. Dynamic typing: Go has duck typing, which gives some of the same benefits. Different refactoring strategies that take advantage of duck typing over other things you'd do with dynamic languages (like just substituting an object of a different type willy-nilly) will probably make up for the rest. I don't see strong support for literal data as a big productivity gain. Go also does a rather good job of treating user defined types like the built-ins and vice-versa, with a few exceptions like maps.

I didn't experience the claims of "Go is just as good as Python for prototyping", as someone with a fair amount of Python experience.

A lot of your competence with Python for prototyping probably has to do with your competence with Python. A lot of the advantage of Go is what it doesn't have. It's a much simpler system, so it's easier to design around, especially as one needs to get closer to the metal.

The use-case for dynamic typing was that I was decoding a large JSON data structure that could easily have thousands of different fields nested 10+ levels deep. Go's standard json library usually recommends that you define a struct type to hold the result of the decode, tagging field names with the `json` annotation. This obviously doesn't work when you only care about a few dozen of the several thousand possible fields. There is fortunately an escape hatch where you can declare a field as interface{} to say "Any type of data goes here; I'll downcast it later, and just trust me on the type", but then you have to litter your code with downcasts.

Python just hands you back a dict of lists, and you pull out the information you want, no type declarations needed.

The "user-defined types like built-ins" lets you build libraries where you can use the basic language structures on the data structure itself, without having to convert to an intermediate form. For example, the Python bindings to Amazon S3 let you access files through dictionary notation. BeautifulSoup lets you iterate over child nodes with a for loop, or access attributes as a dict. NumPy lets you use standard arithmetic operators on matrices. Judicious use of this in libraries makes the resulting code much briefer.

Note on encoding/json: you only need to mention the fields you care about. If there are extra fields, the parser will just scan past them.

    >  I was decoding a large JSON data structure that could 
    > easily have thousands of different fields nested 10+ 
    > levels deep. Go's standard json library usually 
    > recommends that you define a struct type to hold the 
    > result of the decode, tagging field names with the 
    > `json` annotation. This obviously doesn't work when you 
    > only care about a few dozen of the several thousand 
    > possible fields.

Sure it does!

    type Response struct {
        LevelOne map[string]struct{
            LevelTwo map[int]struct{
                Value string `json:"eventual_value"`
            } `json:"level_two"`
        } `json:"level_one"`
    }

The response can have way more fields than just the ones you declare in your struct.

Compared to the Python version of:

  response['level_one']['level_two']['eventual_value']

I did ignore fields that I wasn't interested in; however, you do at least have to declare all the intermediate structs in the tree. With 10-ish levels of nesting, that's a lot of boilerplate; even in your example with 3 levels, you've still got 6 lines of code for what's a one-liner in Python.

It wouldn't be as much of a problem for a production system where you define your data structures once and then expect to amortize them over many, many changes. But I was prototyping, and the point of prototyping is to not spend time on boilerplate that isn't necessary to prove out your concept. A few hundred lines of type declarations just so I could decode JSON isn't exactly convenient.

Lines of code in type declarations isn't an interesting metric. Type safety, however, is.

Said otherwise, the python oneliner is likely not what you want, in real life you'd have to check for fields existence, purpose defaults or warn, etc. So you must likely end up with a few kings of code anyways.

Context of this thread is prototyping. For prototyping that one-liner is exactly what you want; if the fields don't exist just dump a stack trace and fix your code to use stuff that does exist.

I think the solution posed to you is at least much better than what your comment thread started with: defining the entire JSON structure with Go types. Actually, I think it's a lot better. Personally, I'd probably stop there---even if I were prototyping.

But, at your insistence, we can keep digging. It's easy to write a polymorphic function that arbitrarily picks out values from a Go `interface{}` value.

    func lookup(v interface{}, keys ...string) interface{} {
    	switch len(keys) {
    	case 0: return v
    	default:
    		dict, ok := v.(map[string]interface{})
    		if !ok {
    			panic(fmt.Sprintf("%T is not a dictionary", v))
    		}
    		return lookup(dict[keys[0]], keys[1:]...)
    	}
    }

And then we can use it like you would in Python:

    lookup(response, "level_one", "level_two", "eventual_value")

This doesn't completely remove the burden of type asserting (unfortunate), but it at least makes nested lookups much easier. Arguably, the nested lookups are probably where the type asserts hurt the most.

Full example: http://play.golang.org/p/4tWtqGRgU6

The use-case for dynamic typing was that I was decoding a large JSON data structure that could easily have thousands of different fields nested 10+ levels deep.

Yeah, prototyping with JSON is one place where dynamic languages are more nimble.

For example, the Python bindings to Amazon S3 let you access files through dictionary notation. BeautifulSoup lets you iterate over child nodes with a for loop, or access attributes as a dict. NumPy lets you use standard arithmetic operators on matrices. Judicious use of this in libraries makes the resulting code much briefer.

Yes, this is another specific area where dynamic languages do prototyping better and can produce shorter code. I did an exercise where someone implemented a routine in Clojure that I had written in Python. The Python was shorter! In part, this was due to list comprehensions. In part, it was due to an excellent library. What you describe is what many Smalltalkers did as well: create your own control structures and DSLs.

I've sometimes wondered what could be done with a dynamic superset language of Go, such that one could mostly add type definitions and compile to Go.

My biggest pain when programming servers in Go is when handling others' dynamic JSON. Multiple levels of switch-case to handle different types for the same JSON fields at different times. With the current eco systems, Objective C seems the best language, dynamic typing with optional static typing, or mostly static typing in the programs, with dynamic typing for cases like handling dynamic data. Other goods include ARC, Xcode, Apple's frameworks, etc. Concurrency seems to work well with dispatch queues etc. I wonder if Objective C could be as good as Go for servers.

Apple uses Java on its server-side.

My biggest pain when programming servers in Go is when handling others' dynamic JSON.

It should be simple enough to implement a dynamic JSON parsing DSL in Go. It would be slower and memory inefficient, but much more convenient for rapid development.

>A lot of the advantage of Go is what it doesn't have. It's a much simpler system, so it's easier to design around, especially as one needs to get closer to the metal.

Go doesn't have a REPL and that is objectively a serious dent on effective prototyping.

Experienced Java and Go dev here.

On formatting, gofmt is awesome. There's an IDE called LiteIDE which automatically calls gofmt for you, or you can just run it yourself whenever. Additionally, because there's deliberately no room to customize gofmt, you don't have to spend time in pointless meetings when "that guy" decides it's time to rewrite your company eclipse templates and argue over all the minutiae.

As far as refactoring, nobody's built an IDE integration that comes close to eclipse/IDEA for that as far as I know. It's totally doable in Go, in a way that it wouldn't be for Python (yay types), but I don't think anyone's done it. On the upside, doing it manually is less painful than it would be in typical Java code, due to type inference and having less code to begin with.

gofmt is the single biggest feature of Go that might make me try it out. Such a great idea, I wish every language would adopt it.

A lot have:

https://news.ycombinator.com/item?id=7600130

This isn't the same. Go defines a standard format for all code written in Go with nothing left up to programmer opinion. While a lot of great formatting tools exist for other languages, they require choices to be made which inevitably results in disagreement.

If you use C# in Visual Studio, you can press Ctrl+K, Ctrl+D to have the entire document reformatted. Surprisingly few people seem to know about this, judging by how often I've found myself using this on other people's source files and had everything shift around. (Or maybe I'm just unusual and everybody else has just played with the settings.)

The reformatting is actually quite aggressive (it will add/remove blanks and move braces and bits of code around), and so does a good job of keeping everybody's code looking consistent. The Visual Studio text editing functionality only occasionally reaches even the lofty heights of "pretty good", but I do think the document reformatting actually even exceeds that.

(Shame the settings are are per-user rather than per-project, but there you go. It wouldn't be Visual Studio if they didn't manage to fuck something up at the last moment.)

Every decent IDEs have a code formater, but not everyone uses the same IDE so its usefulness is limited. On the other hand, gofmt comes with the platform and make it easy to enforce a common standard between developers.

That said, I think that in C#'s specific case most developers should be using Visual Studio.

The default Visual Studio formatting sucks - wasted lines for opening braces when we have limited vertical real estate. Visual Studio should support per-project, or per-solution formatting configurations, which can be placed into the source code repository to ensure everyone on the team is using the same settings - MonoDevelop already does this and it's really useful for switching between projects which use different formats.

Four space tabs are like a splinter in my mind.

That suggests you don't understand what tabs are. Tabs are a single char, not x number of spaces. You can set your text editor to display tabs as any number of spaces you want, I set mine to two.

See: https://www.youtube.com/watch?v=sln-gJaURzk#t=1735

No need to be pedantic. The choice is more that just trivial display in and IDE.

For example,

https://github.com/zenazn/goji/blob/master/goji.go

I wasn't a huge fan of LiteIDE when I tried it. I prefer GoSublime[1] with SublimeCodeIntel[2], with goimports[3] for my gofmt command. The combo of the three is really awesome.

[1] https://github.com/DisposaBoy/GoSublime

[2] https://github.com/SublimeCodeIntel/SublimeCodeIntel

[3] https://github.com/bradfitz/goimports

What does SublimeCodeIntel provide that GoSublime doesn't already give you? GoSublime gives you Gocode support, so you have full auto-complete already.

For some reason I thought that was where I got the autocomplete from, but as you've pointed out it's totally unrelated.

Cool, just wanted to make sure. I just started with Golang and wanted to make sure I didn't miss something.

> There's an IDE called LiteIDE which automatically calls gofmt for you...

or on emacs, if you use the packaged go-mode.el, you can run gofmt on the "before-save-hook" to do that for you...

I'm surprised nobody has mentioned gofmt -r yet.

http://golang.org/cmd/gofmt/

There is actually a decent open source, third party Go plugin for Intellij that works pretty well. Our team has been using it successfully on a fairly large project.

It's slightly annoying to setup, but absolutely worth it.

Is there a book or tutorial or blog you would recommend for a Python person to learn Go? I'm certainly interested in what your thoughts are.

Here's a few links you might find helpful. They're not specific to Python developers, though.

http://www.golang-book.com/

http://www.golangbootcamp.com/book

http://talks.golang.org/2012

https://gobyexample.com/

Sure - I actually gave a talk last year on this exact topic. There's an audio recording somewhere, but here's the repository for the slides (which contains a link to the PDF): https://github.com/ChimeraCoder/go-for-pythonists

I wrote about my experience here:

http://blog.repustate.com/migrating-code-from-python-to-gola...

How does Go makes it easy to scale the maintenance of software?

We rewrote Twisted Deferred handling to be 30% faster. We optimized the Twisted Reactor event loop to not only do less constant work, but to prefer incoming I/O.
...We spent long nights poring over profiling readouts and traces, wrote our own monitoring system, and wrote our own benchmarking tools

... We enabled cgroups on our Python process to tightly control memory usage and swap contention without requiring a different Python memory allocator scheme.

This is very interesting for the rest of the Python/Twisted community - now that they are not using this (and so is'nt part of the secret sauce), I wonder if they are willing to push this out as a patch/blog post ?

We're planning on open sourcing a bunch of stuff. We'll have some follow up posts.

thanks !

really look forward to this..

It sounds like the cost of context switching between these very different workloads (crypto, disk I/O, network I/O, protocol buffer parsing and formatting) could be improved in Twisted.

Any idea where the overhead comes from? Twisted, or the Python interpreter itself? Is this a GIL performance issue? Or perhaps even lower -- something here is really hostile to CPU cache?

I realize this is a matter of taste, but my favorite async framework is still the kernel. Write small programs that do one thing well (and thus have pretty uniform workloads) and then let the kernel balance resource usage.

After continuing to hit walls with the standard Python Protocol Buffer library, we wrote our own that was 5x faster than the barely-documented C++ Python Protocol Buffer compiled module support.

Ugh yeah, the standard Python protobuf library is pure python and horribly slow. And it requires code generation -- in a dynamic language! The C++ one is faster, but also requires code generation and is just nasty to work with.

Not that this matters much to you at this point, but I have a small C/Python protobuf library that's 10x faster than the standard Python protobuf: https://github.com/acg/lwpb

PS. I see you're in SLC area -- me too. We should talk tech shop in person sometime!

Hi! Yes, you should totally swing by. Shoot us an email :)

I think the sort of performance issues we were hitting were honestly related to just Python runtime overhead. A Python function call is actually really expensive, and with Twisted Deferred handling, it's really hard to eliminate the massive amount of function calls each I/O event does.

We just had a really slow CPU so we did our best to eliminate as many Python function calls in hotspots as possible, but yeah that was challenging.

A Python function call is actually really expensive, and with Twisted Deferred handling, it's really hard to eliminate the massive amount of function calls each I/O event does.

Ah, makes total sense. Any particular reason you chose to use Twisted in a hardware appliance? Is there a web interface that's supposed to be super responsive?

The main codebase actually started out targeting desktop-class hardware, but the poor uptime of user's work machines (now laptops, soon tablets) actually made our business model not work, hence the dedicated hardware.

Did you ever get a chance to evaluate what it does under PyPy? In my experience, there's quite a few such cases where it eliminates most of the overhead, and it's not even uncommon to see it generate what appears to be optimal x86_64.

No, we really wanted to use PyPy, but if we couldn't get it to work on ARMv5 it was of no use to us :(

Why wouldn't PyPy work on ARMv5? I think PyPy folks would be happy to do ARMv5 port for a reasonable price.

You can find contact information at http://baroquesoftware.com/

Yeah, heh, Glyph pointed out the same thing. https://twitter.com/glyph/status/458758471505035265

Hindsight is 20/20 I guess.

I have my doubts that that would have solved all of our problems though, but that's certainly a different path we could have taken.

I like seeing people embrace static typing, but it makes me sad that it's a language where you can't even write a container (a map/dictionry/linked-list/etc) without giving up type safety.

I recognize that Golang is filling a gap, letting people get static typing without having to learn much new stuff, and there's value in that, but I think Go enthusiasts should read these great posts by Tikhon Jelvis with an open mind:

https://www.quora.com/Googles-programming-language-Go-seems-...

https://www.quora.com/Go-programming-language/Do-you-feel-th...

Quite frankly, those articles are absolutely terrible. Firstly, it's clear that the author assumes there exists some correct way to design a language. (I reject that.) Secondly, the author never acknowledges any of the trade offs that come from adding additional features to a language. (Hint: one of those trade offs is... there are more features!)

This isn't a revolutionary thought. More compile time safety generally means more complexity in the type system. We, as programmers, must decide when, where and how much we're willing to pay for that type safety.

You can go balls-to-the-wall with Idris or Agda or ATS and encode a whole mess of things directly into your types. Maybe you can completely remove out-of-bounds errors! Surely, this is better than the alternative, which admits less safety.

This isn't a fair comparison because all three of those languages are ridiculously hard to use for practical things. (Well, it's hard for me, anyway.)

Then there are other, more practical languages like Rust, Java, Haskell, OCaml, etc., that all provide a helluva lot more compile time type safety than Go does. (Or, to be precise, a lot more expressive power with respect to compile time safe polymorphism.) Some of the languages go even further; Rust eliminates data races and Haskell isolates "side effects" with monads.

All of these things are wonderful, because we get a lot more guarantees about ways in which our program cannot fail once it's compiled. But they aren't free. They come with a cost. Typically the cost is in greater language complexity. The programmer must write code that the compiler accepts. If you're in that language's wheelhouse, maybe the complexity isn't so bad. But when you want to write a program that isn't easily accepted by the compiler, you run into complexity. Maybe it's not that bad---maybe you just need to figure out how to write a monad transformer. Or maybe you just need to think a bit more carefully about the lifetime of a borrowed pointer. But other times, you need type families and multi-parameter type classes before you can make an unboxed vector with your own type: http://hackage.haskell.org/package/vector-0.10.9.1/docs/Data... --- The safety is great, but dammit, you're going to pay for it.

On some days and for some projects, I'm more than happy to pay the price for safety. It is immensely satisfying to write a program and be able to make all sorts of claims about its safety. (I just wrote a regexp implementation in Rust. Guaranteed no memory errors. No GC. Totally type safe. It's awesome.) But other times, I don't feel like paying the cost.

And that's OK. It's not sad. It's not "ugly." It's a legitimate and reasonable choice given a trade off.

Saying a language is "badly designed" just because it didn't include your preferred set of features is bad juju IMO.

I agree with much of what you said, but an expressive type system isn't just a safety mechanism. It's also a tool to write fast code productively. Without compile time metaprogramming some things will always be either extremely resource hungry or extremely repetitive.

It's absolutely true that there is always a trade-off. Metaprogramming makes code more difficult to understand and hence more prone to certain types of errors. That's true for runtime metaprogramming as well though.

So if the lack of generics leads to an overuse of reflection (or worse to custom code generators) then the complexity you avoided in the language is re-introduced through the backdoor.

At the end of the day I think it's about the culture that gets established around a language. Go's creators are trying really hard to socially engineer a culture of simplicity, and that's laudable. But it reminds me of the early days of Java even though the Go community would rather have me think of C.

The big difference: C had compile time metaprogramming from the beginning, Java did not. Look which one turned out to be the simpler ecosystem in the end.

> you can't even write a container (a map/dictionry/linked-list/etc) without giving up type safety

Nit: they're typesafe (when done using interfaces), just not statically so.

I think people are rediscovering the benefits of static typing for performance and correctness reasons. The latest generation of statically typed languages[0] all have some level of type inferencing which helps.

If Python is fast enough for you, it’s a fantastic language. The problem is once performance or codebase demands scale, dynamic typing rears its ugly head and there are no simple solutions. At work we sidestep this issue by writing a plethora of tests, but now dynamic typing productivity gains are offset and we spin up a lot of AWS instances for performance.

[0] Go, Rust, Scala. Haskell and OCaml have had it for a while.

While web 2.0. and mobile is reaching maturity with winners are consolidating the market we see the trading of productivity with performance and reduction of operational costs... I m more curious of upcoming booms and which languages and frameworks they are going to bring.

C++ isn't exactly a new language but has also had type inference for a few years. I think people still think of C++ as that dusty old thing in the corner and forget that it is actually capable of a lot of the things users of these new languages get excited about.

C++'s auto is younger than most of these languages (let alone than Haskell or MLs), to say nothing of actual tooling support.

gcc has had auto under a different name (typeof) for years and years, though i don't think it got much use outside of programming competitions.

C++11 is almost a new language. Coming back to C++ now after almost a decade away is exciting.

Wouldn't java be the natural rewrite, for performance speed up? Java is betwen Python and C, in terms of ease-of-use and performance. They mention a C rewrite, but not a java one...

I guess the concept here is that Go is between Python and Java: (much of) the ease of use of Python + (much of) the performance of Java. The type system literally straddles the dynamic and static.

[ I also detect an enthusiasm for cool, new tech for its own sake (despite risks and cost, or whether it's actually better or not - hell let's find out!) ]

Since they are targeting an embedded ARM system, maybe they had concerns about running the JVM on such constrained hardware?

You mean like these?

https://www.aicas.com/cms/en/JamaicaVM

Actually, some ARM processors have instruction set extensions to execute Java bytecode in hardware. http://en.wikipedia.org/wiki/Jazelle

    > The type system literally straddles the dynamic and 
    > static

The Go type system is fully static...

Go has an escape hatch (interface{} and downcasts) that basically give it an embedding of a dynamic type system inside the static one. You end up having to use this a lot if you're doing things like decoding complex JSON data structures or working with containers, since Go doesn't have generics.

> Go has an escape hatch (interface{} and downcasts) that basically give it an embedding of a dynamic type system inside the static one.

No, that gives you a static type system with loopholes, same as Java, you have to use reflect to get a dynamic system in which you can invoke an arbitrary and unchecked method on an object (also same as Java).

Granted, because Go's type system is structural you could define an interface with just the method(s) you want, cast your object to that interface and then call the methods, and you can even define the interface anonymously and inline. Still, go requires your static types to match unless you use reflect.

> [ I also detect an enthusiasm for cool, new tech for its own sake (despite risks and cost, or whether it's actually better or not - hell let's find out!) ]

Why? Because they did not choose Java?

> Java is betwen Python and C, in terms of ease-of-use

Opinions vary highly on that claim. In fact, I've never seen anyone claim "ease-of-use" as one of Java's properties.

He is putting it higher then C. Its not like he is setting a particularly high bar.

"So in the bottom of the ninth, we decided to transliterate our 90k lines of Python directly to Go, line by line.

It took us about 4 weeks."

How many people worked on the project? What's the build time on a program of this size?

We had about 3 or 4 guys full time. On my silly little i3 laptop, the whole codebase builds in 18 seconds - I suspect most of that is our C extensions.

C extensions because something you wanted wasn't in go? Are you using cgo to build?

Presumably, because they were doing it line by line, and the Python version used C extensions (for performance reasons).

Yeah, we actually have a few cgo packages. We have OpenSSL bindings (we need to use our hardware accelerated crypto, go's crypto/tls lib doesn't support that (yet)), a faster CRC thing, we needed monotonic clocks and couldn't wait for Go 1.3, etc.

We'll be releasing a bunch of these. Particularly the OpenSSL bindings, which even without hardware acceleration are faster than Go's TLS library, and despite Heartbleed, OpenSSL is not vulnerable to timing attacks like crypto/tls is.

I wonder how many lines of code they ended up with afterward, although admittedly lines of code is a pretty silly metric. I think it is useful in orders of magnitude when describing the size of a project, 100 lines, 1,000 lines, 10,000 lines, 100,000 lines all speak volumes about effort. So I guess what I'm wondering if you got about a roughly similar sized code base or significant change difference?

EDIT: fixed my calculations again.

Without tests but with our supporting libraries, our codebase was 36784 lines of Python. Those same lines became 41717 lines of Go.

Counting lines of code is hard.

Actually in this case, comparing two languages doing exactly the same thing, LOC is a relevant measure (its a crap measure of programmer productivity). It could show the expressiveness of a language. Also I am a believer that less lines of code = less likely to have bugs.

Less code in general means fewer bugs. However, there's definitely an inflection point where you're trying to cram too much functionality into too few lines of code, and so the code you're writing is a lot more complex than it really needs to be. This is where Go shines, is in discouraging overly complex single lines of code that don't actually save the programmer any time.

Sure, python has list comprehensions, but I can tell you how often I've seen hugely over-complicated one line list comprehensions that were impossible to understand unless you were absolutely sure what it was supposed to be doing before you read the code. I've refactored some of those in my lifetime just to make them more readable. That's not to say that most simple list comprehensions aren't totally fine.

So, just saying fewer lines of code makes for fewer bugs is not 100% accurate. I'd say, less /functionality/ leads to fewer bugs. But except in extreme circumstances (like writing a whole framework to do one small task), actual lines of code does not correlate to the number of bugs.

I wrote some Python to run on an OpenWRT (MIPS-based, 400MHz) router that interfaced with an Xbee radio and ran a web interface. I quickly discovered that what I expected to be IO-bound was in fact CPU-bound - reformatting the data and all the related text operations were fast enough after optimization, but only by a hair.

Our solution was to switch to a dedicated ARM board (beaglebone) attached to the router. But I'm definitely going to take a look at using a compiled language now, as the codebase is still very small.

I've got a completely data-free feeling that Ruby people hitting performance bottlenecks seem to reach for node whilst Python people have a tendency to look at Go.

Does anyone else think that this rings true? And if so - is it a cultural thing?

There are a _lot_ of Rubyists interested in Go, I'm not sure we can draw any real conclusions about the relative popularity of the move at this time.

I have the feeling that both camps are going for Go instead of Node.

My only data-point is the shop I work for, and folks in the community that I know come from a Ruby background, such as Dominik Honnef, Ben Johnson, Mitchell Hashimoto.

My impression is that people writing Web interfaces reach for Node while people writing CPU-heavy programs reach for Go. It so happens that Ruby people are more often writing simple Web apps while Python is somewhat more likely to be used for more computationally intensive tasks, but I think that's about as far as the correlation goes.

I think that depends on whether they're more of a "Rubyist" or primarily Rails developers who know "just enough Ruby". I think Node (and especially a framework like Meteor or Express) is more approachable for those who spend most of their time in a framework like Rails.

JavaScript people go for Node.js - they have nothing to loose.

After all of this optimization, we got up to 1.2 MB/s.

Ouch. That was a lot of work for 0.2 MB/s. The next 2.8 MB/s was also a lot of work, but it seems conceptually more straightforward.

Switching to Go though they were able to get 4MB/s. Even though they claimed to do "line-by-line" translation, I wonder how much of that speed-up was language/libraries vs rearchitecting.

Unless there is some reason to disbelieve the line-by-line claim, can't we assume it's all due to the more performant implementation?

Python is a joy when it's fast enough, but it's not surprising that it isn't good at throughput on small devices.

I didn't mean to imply they were lying in the post, more that it's probably hard to do a line-by-line translation from Python to Go. The nature of having to convert to things like channels in Go might help to re-architect the program even at a micro-level.

That's actually kind of shocking. I thought it would take way more work to port from Python to Go, and a 400% speedup is way less than I'd have expected (though this is string handling, and the Python was already brutally optimised with C modules where it counted).

We're I/O bound now (as we should have been to begin with), so speedup isn't the linear function you might expect here.

I'm curious what the memory usage was like before and after.

Doh, I totally should have mentioned that, cause yeah that was big too. We went from about 80-95mb resident memory (and our cgroups config doesn't allow for much swap) to about 50-60mb on average, which is a metric we weren't even trying to get lower. So, yeah, that's another win we should have mentioned.

One thing to mention, since the OP talks about developing for ARM:

I was surprised recently: first that Go itself bundles a cross compiler (you can easily download Go on x86 and compile ARM binaries with it), second that this made developing for Raspberry Pi and similar platforms very productive, and third that many Linux distributions have QEMU and set up binfmt translation so ARM-compiled binaries run on x86 with no extra effort.

The qemu binfmt extensions inside a chroot are so great.

For the uninitiated: https://wiki.debian.org/EmDebian/CrossDebootstrap#QEMU.2Fdeb...

Can someone tell me how much control GOOG has over the development of Go? I'm interested, but worry that is this another GOOG "open source" project that they have a huge say in.

Even if Go is completely controlled by Google today, there are enough companies using it that they can fork if Google turns evil.

Looks pretty open to me:

here's the current bleeding edge: https://code.google.com/p/go/source/browse

here's the issue tracker: https://code.google.com/p/go/issues/list

looks like it's being used by the dev team.

If you would like to fork the golang project, run this command:

  hg clone https://code.google.com/p/go

This is pretty clearly not Android.

Note also that Googlers have stated that they use the public version like everyone else.

Is anyone else getting SSL errors on the site?

We're getting slammed :( We're spinning up a new webserver.

This would have never happened with Python ;)

Our website is Apache + Python :)

Yeah, was tongue in cheek. I also happen to be a pretty big golang enthusiast, anyway.

lol, I just thought it was an amusing situation.

I'm curious if they tried Cython. I've read that you can achieve up to 35% speedup just by compiling python code, plus you can type pyx file to get near C performance.

I never tried it, this would have been a nice use case.

We did. We used Cython where we could, but Cython doesn't work with Twisted (or at least inline deferreds, which holy we had everywhere), due to different generator semantics. Edit: in fact, our internal protobuf lib was all generated Cython. We had a code-gen inception.

"So in the bottom of the ninth, we decided to transliterate our 90k lines of Python directly to Go, line by line."

A question here: Is it always a practical approach to transliterate line by line when transitioning to another language or because "Go [is] semantically very similar to Python"?

Usually it's too much work to do a refactor with the new language and a conversion.

Much easier to convert first (with whatever warts the existing code base has) then do a refactor.

So I doubt their approach had too much to do with Go per se and more with wanting to get something out sooner rather than later.

Both language semantics and language best practices are pretty similar in Python and Go, and I am sure this helps a lot. Transliterating to something like Java would be hard because their interfaces work differently enough from duck typing to force a lot of refactoring. And switching to something even more different, like Haskell or Clojure, would just make transliteration impossible.

I once consulted for a company whose previous consultant had tried to help them do a line-for-line translation (using off-site random-people-in-a-sweatshop online programmers for hire) from ASP to Rails. They pitched it as "transcoding". I will argue that this was not practical (though I can appreciate the idea that at first glance it almost sounds amazingly scalable ;P).

If you did a line for line transliteration, did you use any of Go's concurrency features such goroutines or channels? I ask because I find that when I translate a program from Python to Go that it's beneficial to structure the program differently so that I can use goroutines.

With Twisted (and with our threadpool worker pools), many patterns translated directly into Goroutine usage in a much cleaner way. Where with Python we were using our own helper libraries, Go's stdlib and the language itself were often more than enough.

We didn't end up using channels too much. Deferreds got translated into Futures that we wrote a small library for. Many of our Go-specific utility classes do use channels heavily though.

I'm curious as to how they did a line-by-line port of the twisted program -- twisted's async (plus coroutines-light w/ via generators) seems like a significantly different model than goroutines + channels. I wonder if they actually implemented Deferred in Go?

We implemented a future class that had very similar functionality to a deferred. We used goroutines heavily but channels not so much.

I suspect we can improve go runtime scheduler performance by starting to go through and replace pieces with channels where appropriate.

I really hope the go compilers will soon support shared object output so we can write python extensions in go instead of having to fully dump one for the other. I know it was a maybe for go 1.3 but I haven't heard much chatter and fear it will get pushed backed.

It's not going to be in 1.3 but the chatter is they might attempt something for 1.4

I hope so, I have some analytical code I really want to wrap for R and Python without the overhead other methods require.

Did you release your faster version of the Protocol Buffer library ?

While we're planning on open sourcing some things, I'm not sure we'll ever open source that one. Even though it was much faster, it didn't see much time in production due to the less than desirable impact it made (go Amdahl's law), and it still has some bugs and was highly customized to our specific protocol buffer definitions.

Internally, Google has something faster for Python, but they haven't released it yet. You might try putting pressure on them here: https://code.google.com/p/protobuf/issues/detail?id=434

http://yz.mit.edu/wp/fast-native-c-protocol-buffers-from-pyt... talks about the issues some more with some links to other rewrite attempts.

would you have gone with thrift if you had to do it all over again ? I ask because I saw the bug description and the glacial response ;)

No, I think we're honestly even more sold on protocol buffers' wire format than before.

That said, if I did want to go looking elsewhere, I think the main serialization format I am really interested in is Cap'n Proto: http://kentonv.github.io/capnproto/

What are the disadvantages of using Go instead of Python?

Testing can be more difficult in some cases, if you haven't carefully structured your code. In python you can always monkey patch whatever outside dependencies you have, whereas in Go, you can't do that without structuring your code to allow it. Careful use of interfaces and dependency injection can let you do it, but if you don't know to use that style, you can get big blocks of code that become hard to test because they expect the whole operating environment to exist, which may be hard to do during testing.

If you want to write code in a Domain Specific Language (even if that domain is "math"), it's nearly impossible in Go. You can't do operator overloading or all the really crazy hacking on the functionality of basic parts of the language to make it work in fundamentally different ways (the way you can with numpy etc). Go will always look like types, functions, and methods.

Right now, Go has no user-defined generics.. this is actually not a problem for many projects, but if you happen to need a lot of specialized containers for different types (like red black trees, graphs, etc), then it can be a barrier. There are ways around it, probably the best way is just code generation to duplicate the container code per type that you need. The other option is just using interface{}, which is basically void* or "Object" for Go, and casting in and out of it... but that can be slow.

There's probably some stuff I'm missing, but those are the major points that I can think of.

My guess is that if you've written code that's hard to test in Go, it's going to be hard to write good, non-brittle tests in Python. It's not so much that you need to write code amenable to dependency-injection to write good tests, you just need to write modular code.

As a fan of DSLs, I actually do appreciate how easy it is to understand any Go that anyone has written, since there aren't any cutesy DSL tricks anywhere. So, yeah, it's hard/impossible to do DSLs, but that might not be a Bad Thing.

Generics is a (frequently discussed) downside. It's definitely a tradeoff. For every container type we had, we had to make a copy of it or a typesafe wrapper around it for every possible instantiation of the container with different types. If you're used to coding by making flexible containers and so on, that's definitely much more brittle and challenging in Go. Go peeps will tell you that's part of a tradeoff they're willing to take.

Why not just bump your processing power?

No offense, but that was a hideous amount of effort for just 20% in one area. An extra dollar or two spent on a hardware upgrade wins for everything.

> provided that your problems are not algorithmic complexity

What does that even mean?

It means "provided the reason your code is too slow isn't that its performance scales very badly as a function of a parameter that, in practice, is not small".

If your code takes time proportional to 2^2^n then for n=4 that's 2^16 x constant, which is probably fine for any reasonable constant, and for n=6 that's 2^64 x constant, which is probably too much for any reasonable constant, so if you're running into problems with n>=6 then moving to a different language or faster hardware or more-micro-optimized code is not going to help you.

On the other hand, if your code takes time of order n^2 and typical values of n are in the thousands or millions, then the difference between C and Python, or Core i7 and Z80, or sloppily written code and highly-bummed code with an inner loop in hand-tuned assembler, may make a big difference.

Thanks for the explanation. I am aware of that; however, the sentence I quoted does not make sense grammatically; "algorithmic complexity" is not an adjective.

I did consider adding a final paragraph saying something like "I am assuming here that you were sincerely asking a question rather than merely being rude and snarky, but of course that may not be true", but decided to play it straight.

So, having established that you were merely complaining about a grammatical slip, and that despite asking "what does ... even mean?" you knew perfectly well what it meant:

Yeah, I agree, it would have been more correct had they written something like "provided that your problems are not a matter of algorithmic complexity" or "provided that your problem is not algorithmic complexity". Congratulations, you found a mistake. (Though not, I think, anything to do with thinking "algorithmic complexity" an adjective.) But (1) so what? and (2) if I'm right in detecting a subtext of "man, if they can make that mistake, why should I take anything else they say seriously?", then I strongly disagree; anyone can perpetrate the occasional grammatical error. (Just out of curiosity, I looked at some of your recent comments, and found one having looked through considerably less text than the length of the article in which you found an error. I dare say one could do the same for mine. To err is human.)

Points taken. In my favor; however, English is not my native language :) Though I do definitely make lots of mistakes in my native one.

Can someone explain why SpaceMonkey is better than Dropbox?

Space Monkey is cheaper for more space - $49/yr for 1TB. Further, because there's a device in your home, getting your data synced off your laptop is a much faster process - you aren't bottlenecked by whatever Dropbox is throttling your upstream at. Not to mention all of the DLNA, Samba, home media system integration we're working on.

And lack of oligarchs, who signed-off on torture, on the company board.

Now that can't be a bad thing! SpaceMonkey needs to use it as a selling point.

It would be good if you explained this more on your site. I couldn't figure out what the device was for, nor whether it was required, etc. I'd suggest a home page that clearly explains what you do and how it works. It's also not clear what happens price-wise after the first year. Is there some fee to pay for the device or just the $49/yr for the storage? Etc.

Was the CPU pegged at 100% when using Python then?

yep

fyi..login link fails on the site.

Well it is kind of obvious that a compiled language is going to be faster than an interpreted one, especially the way how these interpreters work (https://wiki.python.org/moin/GlobalInterpreterLock). You can fool yourself with Twisted (or in Ruby with EventMachine, Goliath etc.) but it gets so just a bit ahead. The surprising fact for me is that Go is not as much faster. I was expecting a bigger gap in the performance between Go and Python. Understanding where your bottlenecks are is crucial and it is not super hard in Go. https://www.datadoghq.com/2014/04/go-performance-tales/ I guess the SpaceMonkey guys might further improve the performance just by doing a thorough analysis on their code.

They wrote a naive, line-by-line translation of their python code in Go. There's probably still a lot of low-hanging optimization to be had.

Exactly. I am kind of wondering why this 1:1 translation came up why not just follow best practices and implement a functionality instead...

The path they chose was much more predictable than your suggestion. I.e., after their team spent three days at it, their estimate for the total translation was probably within 50% of the month it actually took them. I don't expect they could have been as accurate in estimating an entirely new implementation.

Right on. We immediately had not totally horrible estimates for how long it would take.

Further, we have always really been inspired by Joel Spolsky's article on rewrites: http://www.joelonsoftware.com/articles/fog0000000069.html In fact, I can't help but wonder if we would have attempted a Go rewrite sooner if not for that article.

I've got a project I'm porting out of Erlang to Go. It's not a transliteration, since it can't be, but by design it's a drop in replacement, modulo a smidge of configuration for each (dozen lines of config, tops). As much as I'd like to GO WILD AND FIX ALL THE THINGS!, it's advantageous to be able to switch back and forth freely during QA and early deployment, because you need a smooth transition.

Out of curiosity what is you motivation for porting something from Erlang to Go?

I was expecting a bigger gap in the performance between Go and Python.

Perhaps we've come a pretty long ways in the interpreted languages department? Or, that the cost of interpreting is vastly less than speed losses due to IO or memory access?

"Interpreted is always slower than compiled lol right guise?" was a tired refrain ten years ago, much less today.

CPython's bytecode interpreter has not really improved at all in 10 years relative to JITs and compilers.

That's a problem with CPython, not interpreted languages as a whole.

Maybe they'll make it faster in Python 3000, right?

They make it pretty damn fast. But non-JIT runtimes have strict limitations on how fast they can go, unless you design your language specifically to optimize for interpreter speed (like lua).

PyPy is probably your best bet if you want a faster python. The official CPython will never have speed as a primary focus due to a number of self imposed limitations like:

Since it's a reference implementation it should be easy read and learn from.

They're not really willing to accept patches that speed up some things if they at the same time slow down other things (this has been the main problem with all the GIL removal patches that have shown up over the years).

They're not willing to accept patches that break any existing code or libraries.

PyPy on the other hand have non of these limitations and happily break all three, making it great for a subset of python code out there.

Well and more so, our workload really should be mostly I/O bound. We freed up the CPU with Go, but CPU is no longer the current bottleneck.

Python has benefited from years of throughout optimization of the language for real-world use. Go is still under heavy development and is not as optimized as other languages that have devs focused on optimization (that explains why node.js is a bit faster than Go for handling HTTP requests at the moment).

  > (that explains why node.js is a bit faster than Go for handling HTTP requests at the moment)

Pretty sure node's http parsing is handled by http-parser[1], which is written in C (originally pulled from nginx as I recall).

Go's http parser is written in Go.

[1]: https://github.com/joyent/http-parser

If I remember correctly the initial versions of node had the ragel-based http parser from mongrel but then ryan rewrote it. http-parser is hand-written.

maybe not entirely hand written? source[1]

[1]: https://github.com/joyent/http-parser/blob/master/LICENSE-MI...

    > (that explains why node.js is a bit faster than Go for handling HTTP requests at the moment)

I'm not sure that is true any more:

http://www.techempower.com/benchmarks/#section=data-r8&hw=i7...

At any rate, "handling HTTP requests" can't possibly be your bottleneck, given network communication, serialization, etc.

> it is kind of obvious that a compiled language is going to be faster than an interpreted one

While it may be obvious, the interesting question is why exactly. There's been a good presentation from Alex Gaynor [0] on this topic. TL;DR: Dynamic languages ar slow not because they can't be optimized, but because developers make a worse use of memory and have sucky algorithms, inducing way more mem allocation and copy than needed. If you can be smart about those, then your dynamic language can match static languages.

[0] https://speakerdeck.com/alex/why-python-ruby-and-javascript-...

There are certain things that you can't really avoid using an interpreted language. How would you go around Ruby's GIL for example when you specifically need multiple threads (not talking about green threads). I agree that some of the stuff is on the developers developing libraries and performance is not kept in mind but there are hard limits you can't really do anything about.

JRuby, RubyMotion

> How would you go around Ruby's GIL for example when you specifically need multiple threads (not talking about green threads).

Using any of the Ruby implementations that don't have a GIL. Notably, JRuby, MacRuby/RubyMotion, and (I think) Rubinius.

Rubinius does not have a GIL, you are correct.

Except for the fact there are no 'interpreted' or 'compiled' languages. What we have are implementations.

This is not merely nitpicking. Any turing-complete language can be interpreted or compiled. Language specs tend to make one side easier, but that doesn't mean that Python cannot be compiled (and it is actually compiled to bytecode). You can have something similar to a GIL in a compiled language too.

One of the things that makes it easier to increase performance is having type declarations. This makes it much easier for the compiler to reason about the code, which can lead to increased efficiency. But optional type annotations can accomplish just as much.

And the website takes more than 30s to load...

At least it does load. A lot of sites fall over on a good HN'ing.

And a lot of sites survive just fine. I don't think "at least it loads" is something to be proud of, especially when the topic is performance.

Except their website isn't in Go. So it doesn't even pertain to the conversation.

> Our website is Apache + Python

We just doubled our instance count, we should be good now.

LGTM, and your response time to "support" requests is certainly great! I'll leave my grumpy comment, so that the context does not get lost.

Reply on news.ycombinator.com