When you see the effort some of those compagnies go to get some decent performance, chose the right language from the start would have been easier on that side, look at Facebook with Hack, Youtube ( Python ) that re-implemented a VM in Go, Shopify is also doing a lot of things to improve performance because of Ruby limitation, when you pick up a language with by default good performance you don't have to go that route.

Are there as many examples of the inverse, big successful companies that started off with a "fast" stack and didn't need to optimize the one they grew with? It seems like flexibility is more important than performance until you actually have a winning product.

but ruby is moving in this direction

- jit

- more functional tools

- better introspection

- sorbet

- ractor

- truffleruby, and jruby

Ruby is working hard on performance but Sorbet is hardly a common thing. The Rails community isn't gonna adopt types anytime soon imo. To be fair most likely never, at least not as a community. Some companies may go that route but I hardly anticipate a big movement.

Those are all grafted on afterwards, and they don't see too much use apart from the people who made them (ie Stripe with sorbet). Compare that to a language that has first-class support for all of these built in. As you learn such language you're forced to learn these concepts as well, so more people in the community use them, thereby increasing their effectiveness through sustained development because they're popular features.

Unfortunately, in the case of Shopify none of those tools will run rails any faster. I think they only use TruffleRuby for some internal tools.

Hopefully things will improve now ChrisSeaton is working there.

Sure, all web development should be C with some assembly sprinkles.

This is a bad faith argument, the parent is not saying to use C, but surely you can concede that there are better languages for performance and are useful for web development? Even Node is faster than Python and it's chugging along well in the web dev community.

Node is probably faster than Python but so what? It comes with its own set of problems and drawbacks. If you're optimising for performance I don't see why you would go for Node anyway, and if you're not optimising for performance than I still don't see why Node but to each his own :)

Practically every popular language is faster than Ruby and Python.

Ah yes, because C is the only fast language.

C++ is faster (due to inlining and templates).

C++ as it is written can be arbitrarily slower and may have hidden bottlenecks due to copy constructors and poorly thought out Stl usage. But C++ fanboys refuse to admit this.

*can be faster

It can also be slower :)

So can Java :)

web assembly ;)

> 120 Billion worth Shopify isn't enough, maybe Instagram then? Github ?

Do these sites also have back end services written in other languages for the heavy listing? Anecdotally, seems to be a lot of stories to that effect.

And as noted in the article, Javascript/NodeJS is in a different performance category from Python and Java, more like Java or Go. But the simplicity of Go might make the performance and memory usage more predictable.

Shopify is absolutely a Ruby / Rails company. If out of the thousands of developers that work there you have a team or two writing some performance heavy c++ / Go code that says absolutely nothing.

Wouldn't it actually say a lot? Engineering is about trade-offs, and the trade-off that says "use this special language for 1% of our most critical services instead of the company-wide standard language" is not "absolutely nothing".

Lines of Code is a bad proxy for mission critical.

All of Shopify is mission critical. The code that constitutes the Shopify monolith is mission critical for sure. The company won't function without it.

How do you define mission critical?

This is somewhat circular, but: when your team decides to rewrite a component in a more complicated or more obscure (at least at Shopify) language to increase throughput rather than spinning up more instances of the current solution, then you're talking about a component that's even more critical than the other critical components.

It's just common practice - in Ruby some popular gems that parse XML or JSON come with a C extension that does some performance heavy calculation. Most gems don't do this, but when performance is very important some do. Us Rubyists see no problem with that, and we don't take this as some kind of incentive to leave Ruby, maybe even the opposite; knowing that you can do 99% of your app in Ruby, and if ever needed, easily fill in the 1% with Rust/Go/C++ is very reassuring imo.

It entirely depends on what you're doing. There are two issues as I see it:

a) Lack of static analysis tools means having to do more testing (manual or automated) for simple mechanical errors.

b) For super high throughput low latency systems, they are not up to snuff. When I worked in real time bidding in ad-tech this was actually a serious concern. I'd reach for Rust (or C/C++) in this scenario now.

Speaking from experience, the JVM is fast enough for real time bidding in adtech. Most major ad exchanges require the latency to be below ~80ms, which is not that hard to achieve. In contrast to high-frequency trading, there is no competive advantage in having a lower latency.

Are compile times and program start-up times not a factor?

One of the things I really appreciate about golang (from a completely different field) is how quic teh builds are, and how fast binaries start up (it's like I wrote it in C).

Java can compile quickly, a few minutes at most when C++ would take hours, so I am tempted to say that's not a problem.

The startup time is negligible in my experience (few seconds for JVM or python imports). I have to take over slow starting applications from time to time and it's always because of loading data and doing stupid shit on startup, irrelevant of the language. It's not a problem for production because server application only reboot once in forever.

It's still a problem for microservices architectures, unfortunately, especially if you want to support dynamic scaling of some kind. A few seconds is nothing if you expect that your server will be up forever, but it becomes a lot if sometimes it goes down for a bit to move to a different machine, and that takes seconds for your customers.

Also, JIT languages have a very poor habit of doing a terrible first impression because of the warm-up time, especially in Java. If you are delivering applications to customers, that becomes a real burden - the very first time they use your shiny new application, everything is moving like molasses, until the JVM decides it's JIT time...

In a microservice architecture you'd probably have more than one instance running at any given time though, and do a rolling restart so there's always at least one instance available.

Yes there are niches where Ruby won't work, even big niches like Kernel development or real time systems maybe. But for web development, in general, these languages have proven themselves for so long it's getting quite ridiculous now to say they won't work. As for testing - I disagree. Frameworks like Rails are so easily testable it's a breeze. Java/Spring dependency injection is jumping through hoops just to provide a testable framework I find it hard to believe it's any easier.

Yes, I don't do web development. At least not mostly.

But I understand that that's what most people out there are doing. But they should also not delude theirselves that the toolsets appropriate for that environment are appropriate every where else. I see this bias a lot, on HN even, that everyone now is a "full stack developer" doing this kind of development.

Many of us work in entirely different stacks.

> But for web development, in general, these languages have proven themselves for so long it's getting quite ridiculous now to say they won't work.

And I don't think web development is where Go is getting traction. More in back end services and infrastructure.

> Java/Spring dependency injection is jumping through hoops just to provide a testable framework I find it hard to believe it's any easier.

As someone who develops with Spring everyday, I don't think it's the best solution for anything.

What are your main pain points with Spring?

Annotation oriented programming.

Having the code execution paths radically change by adding an annotation to a method or a class makes it very difficult to reason about what it will do when deployed.

If that annotation you found through Google does what you want and expect, that's great. But if it doesn't, or fails in unexpected ways, debugging it can be a nightmare.

This always baffled me -- when Guava came on the scene, and when Spring also adopted annotations...

The whole original point of dependency injection was to decouple dependency management from the code, to make it easier to test, and easier to reason about and analyze.

DI via annotations goes ahead and sticks them right back in there. And now we have, like you say, magic code that is difficult to reasona bout.

XML sucked, but annotations sucked more.

Its interesting that a java framework is so "magical", feels like it goes against the java mindset I know.

Yes, that's what makes it so terrible. All of the "action at a distance" complexity of Ruby meta-programming, with none of the concise and easy to read code!

just out of curiosity -- for ad-tech real time bidding, what's your network latency and bandwidth like? I can buy that C/C++ is needed if you are colocated to an exchange, but if you're bidding online, those few fractions of an ms you save in C/C++ vs node.js you could have also saved by locating your server closer to whatever ad-tech exchange you're bidding on.

The issue I had back in this line of work was with garbage collection and was an issue in the 99th percentile.

When you have an expectation from the exchange that you respond in under 80ms, and 25-50ms of that is eaten up by transport, you don't have a lot of time to mess around.

So you spend the first chunk of time optimizing I/O and how you're accessing it.

Then you start looking at computation -- improving caching, etc.

At a certain point you start noticing in your graphs that you're on average doing quite well. But there's those hiccups every Nth request...

And now you're in the game of fighting with your language's garbage collection algorithm...

One clean and dirty way to help with that.

Try to improve allocations where possible. There's often plenty of small allocations happening in Java that can be avoided (string usage is one major driver). Less allocations means less frequent garbage collections.

Then one hack is to disable the garbage collection. Let the software run up to consume all the 100 GB of memory of the server and crash and restart. There is no pause of garbage collection when there is no garbage collection.

If it's not enough, the last resort is to write native code or switch to C++.

At this level of scrutiny, replacing algorithms is appropriate and if the algorithms are built into the language then replacing the language may be one* way to clear the issue, but you have to do your homework to get there.

*sync.Pool, arena allocation strategies

when I worked in adtech, we located the servers close to the exchange. This indeed made much more difference to the latency than code optimizations.

Do u guys mean adtech or real time trading?

Realtime bidding on ad exchanges

> b) For super high throughput low latency systems, they are not up to snuff. When I worked in real time bidding in ad-tech this was actually a serious concern. I'd reach for Rust (or C/C++) in this scenario now.

What did you do? You wrote servers in C? you wrote a Redis for instance?

Optimized for GC and figured it out. But that point also coincided with a job switch; I went to another ad tech company, but on the exchange side rather than the bidding side. And all our ad server infrastructure for that was written in C/C++ (with embeded V8 JS for biz logic stuff). Then that company was bought by Google, and I worked on the exchange side at Google, too, where everything was also in C++.

My successor at the original startup rewrote everything in Python. And I watched from the exchange side as they struggled for two months to meet basic performance constraints. They eventually got it though. It certainly can be done.

It's worth pointing out this was 10 years ago. And in the meantime we've had the usual improvements in machine performance, and SSDs are a thing in data centres, etc.