It seems you cannot go a week without hearing about the next, allegedly faster, so called “web framework” for Node.js. Yes we all know Express is slow, but can yet another “web framework” really improve I/O performance? Beyond evading the overhead of Express, no, it cannot. To reach further you need to dig deep and redesign, not just slap a new layer on top.
Express is one of the oldest, so called “web framework” for Node.js. It builds on the out-of-box features Node.js provide, adding a nice App-centered programming interface to manage URL routes, parameters, methods and the like.
It is productive and elegant, sure, but lacks in performance. Emerging are projects like Fastify, and hundreds alike. They all aim to provide what Express does, at a lower performance penalty. But that’s exactly what they are; a penalty. Not an improvement. They’re still strictly limited to what Node.js can provide, and that’s not much as compared to the competition:
No so called “web framework” for Node.js, whether Fastify or not, can surpass the red line. This is a pretty low upper limit as compared to trending alternatives like Golang.
Luckily Node.js supports C++ addons, Google V8 bindings that link JavaScript to C++ and allows your JavaScript to invoke any behavior, even behavior that’s not provided by Node.js itself.
This makes it possible to extend and redefine what’s possible to do with JavaScript. It opens up for JavaScript that performs to the full extent made possible by Google V8, not limited to what the Node.js “core programmers” have decided is good enough.
I’m releasing brand new code, µWebSockets.js, available on GitHub today:
https://github.com/uNetworking/uWebSockets.js
It’s an alternative web server for JavaScript backends, written in ~6 thousand lines of C and C++, surpassing in performance the best of Golang by a large margin. Bitfinex.com already ported both of their trading APIs (REST & WebSocket) and are currently gradually putting it in production.
Paolo Ardoino from Bitfinex wanted to interject that “it’s a damn pretty cool project”.
This work is made possible solely thanks to sponsors; BitMEX, Bitfinex and Coinbase have made this work possible. Thanks to them, we now have a new release!
This is a new project, new code licensed Apache 2.0, successor to what’s known as “uws”. It’s an entire stack, from OS kernel to Google V8, a complete bypass that brings stable, secure, standards-compliant, fast and lightweight I/O for Node.js:
In this layered software design, where every layer depend only on the previous one, it becomes very easy to track and fix issues and/or extend with new support.
µSockets itself even has three sub layers, going from eventing to networking to crypto, each sub layer only aware of the previous one. This makes it possible to swap out parts, fix bugs, add alternative implementations all without changing any high level code.
Feeling tired of OpenSSL? Fine, swap it out by replacing ssl.c and its 600 lines of code. No other layer even knows what SSL is, so bugs are easy to locate.
Internal sub layers of µSockets
This differs greatly from how Node.js is implemented, with its “all-and-everything-in-one” design. In one source file of Node.js you can find libuv calls, syscalls, OpenSSL calls, V8 calls. It’s all mixed up in a big mess with no sense of isolated purpose. This makes it hard to make any real change.
Following is an oversimplification, many concepts left out for brevity, but should give you an idea of what µWebSockets.js is all about:
It is possible to outperform, in some regards, Golang’s Gorilla WebSockets on an SSL vs. non-SSL basis. That is, your JS code on SSL can message faster than Golang can without SSL (in some regards). I think that’s pretty cool.
Socket.IO is in many ways the “real-time” equivalent of Express. Both are old, elegant and popular, but also very slow:
Most of what Socket.IO helps you with boils down to pub/sub, the feature to emit messages to a room of multiple recipients, and to receive likewise.
Fallbacks are completely pointless today as every browser supports, and has supported for ages, WebSockets. SSL traffic cannot be interpreted by corporate proxies, and will pass through just like any Http traffic would, so WebSockets over SSL is definitely not going to be blocked. You can still have fallbacks, but they are pointless and incur unnecessary complexity.
One goal with µWebSockets.js is to provide features similar to those found in Socket.IO to make it somewhat simple to swap away completely, without any wrapper on top. This while not enforcing any particular non-standard protocol.
Many companies, most of them, struggle with some kind of pub/sub problem when it comes to WebSockets. Sadly, efficient pub/sub did not make the deadline for this release, but is coming shortly. Very high priority. It’s going to be really fast (benchmarks already put it faster than Redis). Keep an eye out!
Polishing, adding features and correcting mistakes. There’s going to be a period of introduction where things will maybe not fit in completely from the start. Keep in mind, it’s a large project consisting of many thousands of lines of C++ and C in three different repositories:
This project is used by companies with huge stress on I/O. Stability and security is (naturally & obviously) of highest priority to the project. Make sure to report stability issues in early point-releases now that this code is a major and big release with tons of changes.
#node-js #go