From Python to Go: migrating our entire API(blog.repustate.com)
blog.repustate.com
From Python to Go: migrating our entire API
http://blog.repustate.com/migrating-entire-api-go-python/
9 comments
It sounded like they were gunning to shave the maximum in response time from their HTTP APIs. From my understanding, Elixir/Erlang probably wouldn't be able to match an optimized Go process in raw efficiency. I get the sense that scaling horizontally is the big draw of the Erlang VM, not so much raw responsiveness.
If simplicity were the primary concern, they could have stuck with Python. You do get a much more distributed element with Elixir/Erlang (which certainly helps with volume), but I'm not sure they'd reach that 10ms response time average with it.
If simplicity were the primary concern, they could have stuck with Python. You do get a much more distributed element with Elixir/Erlang (which certainly helps with volume), but I'm not sure they'd reach that 10ms response time average with it.
If latency were the primary concern nothing would beat the BEAM platform. It was built for that purpose, low latency is absolutely critical when you make a phone call.
Of course that's a bit like a benchmark (meaning, it means absolutely nothing to you/your code). If there were any more latency in the Erlang solution it would be the same algorithmic issues that could pop up in any software.
Go is more efficient for CPU throughput, but BEAM has been fine tuned for decades to keep latency down and wouldn't be able to top for responsiveness.
My personal opinion is that they made a mistake. Reporting about swapping out a stack like this is weird because it looks like they don't know what they're doing. What's next month's blog? We're Moving to Rust? I'm glad they're satisfied, but you gain improvements anytime you do a rewrite- even in the same language. I wouldn't have migrated my existing stack anywhere. I'd use PyPy which would've done wonders for responsiveness on the existing product and adopt containers for deployment.
Of course that's a bit like a benchmark (meaning, it means absolutely nothing to you/your code). If there were any more latency in the Erlang solution it would be the same algorithmic issues that could pop up in any software.
Go is more efficient for CPU throughput, but BEAM has been fine tuned for decades to keep latency down and wouldn't be able to top for responsiveness.
My personal opinion is that they made a mistake. Reporting about swapping out a stack like this is weird because it looks like they don't know what they're doing. What's next month's blog? We're Moving to Rust? I'm glad they're satisfied, but you gain improvements anytime you do a rewrite- even in the same language. I wouldn't have migrated my existing stack anywhere. I'd use PyPy which would've done wonders for responsiveness on the existing product and adopt containers for deployment.
> I'd use PyPy which would've done wonders for responsiveness on the existing product and adopt containers for deployment.
That's an easy claim to make, but you don't know whether their stack would actually get much improvement out of PyPy. We've had trouble in the past using PyPy with C extensions. Trying to migrate to pure Python equivalents was a headache. Once we managed to get a PoC of a service running under PyPy, we saw very little improvement.
That's an easy claim to make, but you don't know whether their stack would actually get much improvement out of PyPy. We've had trouble in the past using PyPy with C extensions. Trying to migrate to pure Python equivalents was a headache. Once we managed to get a PoC of a service running under PyPy, we saw very little improvement.
It is an easy claim to make but just as easy for them to test as well, so it works out.
10ms HTTP response times in Phoenix (on elixir) is pretty normal.
I see a lot of these types of blogs but I'm always left with the same questions. Did they not try using PyPy? Using LXC and PyPy would be my first step, not swapping out my stack. It just seems more sensible to me considering that Go as a platform isn't perfect either. Nothing is.
But I'd be hesitant to leave a massive, thriving ecosystem like Python (especially if on 2.7 which has everything). Win some things, lose some others when you can just adopt containers and PyPy and gain the benefits that Go has without losing Python's advantages.
But I'd be hesitant to leave a massive, thriving ecosystem like Python (especially if on 2.7 which has everything). Win some things, lose some others when you can just adopt containers and PyPy and gain the benefits that Go has without losing Python's advantages.
They tried Cython first, so PyPy wouldn't have been faster than that. The big speedup for them was in making the code parallel, which is tricky in python and typically requires processes instead of threads and copying instead of sharing, either of which may eat most of the gains in this case.
their problem was using weird and heavy frameworks.
i wouldn't consider djangob for anything other than a heavily cached content site.
seems like it's a company that just plow through problems. i bet whatever the problem is in go they will find only next week.
python 2.7 and 3 have a marvelous, but not very well documented, multiprocessor module. it gives you more performance for that the cython and other shenanigans.
but i guess using the existing tool right doesn't give promotions.
... yeah I'm making a lot of assumptions about then but that's just because I'm salty today. they should still learn to use python multiprocessing module. :)
i wouldn't consider djangob for anything other than a heavily cached content site.
seems like it's a company that just plow through problems. i bet whatever the problem is in go they will find only next week.
python 2.7 and 3 have a marvelous, but not very well documented, multiprocessor module. it gives you more performance for that the cython and other shenanigans.
but i guess using the existing tool right doesn't give promotions.
... yeah I'm making a lot of assumptions about then but that's just because I'm salty today. they should still learn to use python multiprocessing module. :)
They cite their main reason for migrating as being self-hosting, which is hard in Python.
I've had a huge amount of success using pyinstaller for this problem, shipped many products into hugely diverse environments.
https://pyinstaller.org
I've had a huge amount of success using pyinstaller for this problem, shipped many products into hugely diverse environments.
https://pyinstaller.org
I think performance ended up being the big win:
"We migrated our entire API stack from Python (First Django then Falcon) to Go, reducing the mean response time of an API call from 100ms to 10ms We reduced the number of EC2 instances required by 85%"
"We migrated our entire API stack from Python (First Django then Falcon) to Go, reducing the mean response time of an API call from 100ms to 10ms We reduced the number of EC2 instances required by 85%"
I'm usually left wondering if anyone else know about the multiprocessing module. Specifically the Pool class.
It eats a lot of RAM. If your app uses X RAM, with 2 processes it will use 2X RAM. It quickly gets expensive if you're using AWS or similar.
Go is going to be a lot faster than Python even without multiprocessor. The performance difference becomes even larger on multi-colored cpu's. Say, for example, if Go is 3 times faster than Python on one core then you throw in 4 or 8 cores, the win is even bigger.
That's true but how many tasks legit need multithreading? Not many in my experience. Over the course of my career I've come to the conclusion that multithreading is best left for systems programming and very few specific applications (such as a game engine and not much else).
For pretty much everything else that people use multithreaded applications for in the past (encoding, image editing)- multinode processing is vastly superior to multithreaded. Erlang calling out to C (if even necessary) is going to win there over most. Or you could use Python or whatever you wanted.
The Java/Go way of local multithreaded applications seems archaic to me.
For pretty much everything else that people use multithreaded applications for in the past (encoding, image editing)- multinode processing is vastly superior to multithreaded. Erlang calling out to C (if even necessary) is going to win there over most. Or you could use Python or whatever you wanted.
The Java/Go way of local multithreaded applications seems archaic to me.
How did debugging change, for the better or worse?
Did your tests speed up?
I wonder why they didn't go with Java. It would have been a good fit with the additional benefit of not having to pay the premium Go devs go for (Go dev here).
Did someone gather salary statistics? I could imagine it could also be the other way around: hordes of Go fans competing for a relatively small amount of jobs drives prices down.
Am I the only one who thinks it's crazy that people would compete for jobs because of a language? Especially a new language? Would anyone seriously give up a job just for the opportunity to program in another language?
Perhaps this is a bigger thing among corporate programmers? I've only worked at smaller companies where they are more open to experimenting if you talk them into it.
Perhaps this is a bigger thing among corporate programmers? I've only worked at smaller companies where they are more open to experimenting if you talk them into it.
Many programmers have fairly strong preferences about programming languages and definitely many people would consider a Go job more desirable than a Java job. Most of those people aren't currently Java programmers - it's said[1] that Go is largely attracting people from dynamic languages and not from C++/Java.
[1] See this essay from Rob Pike where he says Go is attracting programmers mostly from dynamic languages (though he raises C++ programmers as the chief contrast, but still mentions Java too) : http://commandcenter.blogspot.fi/2012/06/less-is-exponential...
[1] See this essay from Rob Pike where he says Go is attracting programmers mostly from dynamic languages (though he raises C++ programmers as the chief contrast, but still mentions Java too) : http://commandcenter.blogspot.fi/2012/06/less-is-exponential...
It's an interesting phenomenon indeed. Coming from C++ and Prolog, I disliked Java quite a bit, but took two positions where Java was the main language (because they were interesting NLP/ML positions). In the end it's a tool to get interesting work done.
I follow some PL lists and there seems to be a sizeable group of people who prefer a position in their favorite language. Of course, I don't know what portion of the overal market they represent.
I follow some PL lists and there seems to be a sizeable group of people who prefer a position in their favorite language. Of course, I don't know what portion of the overal market they represent.
I think it's the opposite. The corporate programmer might not care about Go vs Java. But in a smaller company, Go is going to be more desirable.
my favorite quote:
"Writing software in languages like Python or Ruby often seduces you into being ignorant of what’s going on under the hood because it’s just so easy to do pretty complex things, but languages like Go and C don’t hide that."
Damning with faint praise. Often, you don't need to know what's going on under the hood, and often it's not very interesting.
To tell the truth, Python is an easy target here, but I doubt 99.9% of the projects in C or C++ care to manage memory as closely as GHC does under the hood. (The same way as they don't manage concurrency.)
Some times, it's even better not to look.
Some times, it's even better not to look.
tl;dr - Python is to slow for us. (its about the implementation, not the language).
I expect the performance/parallelism complaints. It's interesting that the tl;dr also includes "packaging and distributing an app is a reason go > python".
Python needs a .war format and I'd like to see more people get behind https://github.com/pantsbuild/pex
Python needs a .war format and I'd like to see more people get behind https://github.com/pantsbuild/pex
The original Python .war equivalent was the .egg, but that's slowly losing ground to the .wheel.
Oh wait, I'm confusing .jar and .war, aren't I? I think .wheel has some .war-like characteristics, thus it's growing popularity, but an not sure.
Pretty neat trick porting over python unit tests to go. Any plan on sharing more details or open sourcing that part?
"Due to the similarity between Python and Go" - huh? Go and Python are almost perfectly orthogonal.
Really liked: the simplicity of Go's build, not having to spawn N processes on the server side, the compiler catching so many issues, the very strict formatting rules, all a major win over Python. However, didn't like how verbose the code turned out, especially for parsing or any data manipulation tasks. For server-side debugging/tools, it's also a bit of a build-your-own adventure.
Elixir I found to be much more pleasant to write and subsequently read, more well-suited for back-end tasks, great for generating and parsing data. Also, more things in the toolset for back-end deployment, in large part due to the OTP stack.
Perhaps Go's ideal use case is local system tools, compilers, or anything else that requires concurrency (let's say, creating tools like "ls," or Apache bench, or gcc, or cURL), and something like Elixir is more ideal for building actual servers.