We built a background daemon as a macOS menu bar app in Go, and the performance was surprisingly bad. The Go bindings for native UI frameworks ended up being massive RAM hogs. When we profiled it, we found that the GC essentially gave up under load, which explained why customers were reporting a simple menu bar app consuming 2.5GB+ of RAM on their Macs. We eventually abandoned the Go approach and switched to Electron. (Not-so) Surprisingly, both the DX and UX improved significantly for our use case. Personally, I’d still prefer Swift/C#/C++ for native desktop work (coming from a Qt C++ background), but given the business constraints at the time, Electron ended up being the most pragmatic choice.
> When we profiled it, we found that the GC essentially gave up under load
Hmm, the Go GC is really quite capable, so I wonder what kind of pathological load it was being presented with. Even then, when the GC "fails" it means elevated CPU load from collection.
The main thing I can think of would be the application "leaking" by having unintentional references (or worse, actually leaking through cgo bindings), or trashing the allocator to cause temporary spikes in between cleanups.
However, while I don't think Go was actually to blame here, I would never use native UI bindings to a language that isn't 1:1 compatible with the original design and memory management principles, as such bindings get disproportionaly large and complex. It just sets you up for a bad time.
I totally agree :) I don't blame Go either. We were already a pure Go shop with a lot of focus on backend and infra systems engineering and were trying to venture into the desktop app market for our device monitoring software. Once we validated our idea with a rather buggy MVP haha, we quickly switched over to Electron and deployed on all 3 desktop OSes properly.
That is a surprising use case about Go and Electron (!), I would have imagined no contest for the superior performance of a compiled language, even with garbage collection. But the mention of "bindings for native UI frameworks", it was probably running the compiled code in a very tight loop, stressing the runtime. In contrast, Chromium specializes in UI with years of optimization.
Recently for a specific purpose I was reviewing options including Tauri, various WebView bindings, and in the end had to admit that Electron is probably the best approach in terms of feature set, development speed, etc.
I remember we used https://github.com/getlantern/systray at some point along with some interaction with fyne.io. That said, I do not remember how we managed the lifetime but we did test it thoroughly before deploying to our initial beta users (or so we thought :P). The GC behavior started to happen whenever the app was in the background, which it was supposed to be for the majority of the time.
Been a while since I worked on it but I remember we used https://github.com/getlantern/systray at some point along with some interaction with fyne.io AFAIR.
We are using it for our apps, but I can see why people do not use it for new projects:
1. The state of C++ is not great. Few developers, C++ footguns, complicated build systems, and generally slow progress, see my https://arewemodulesyet.org/
2. How Qt presents and licenses itself. Either you go LGPL or you have to pay big money for a commercial license, which will then infect all other apps as well. For example, when you have two Qt apps that talk to each other you must license _both_ commercially.
3. The split of Widgets and QML makes the ecosystem fragmented, because Widgets will never die. Even the Qt devs themselves are split about this. You can see this when example code for a new feature uses Widgets. QtCreator is also a nice example, where they reverted some new QML code quite a while ago and have not substantially added any new QML code since then.
4. Tooling: We use QML for everything and the tooling is not great. The language server is still super flaky and breaks, and developer tooling like the Chrome Dev Tools is virtually nonexistent.
5. Packaging is still also not great but has gotten better in the last few versions where Qt creates a deployment cmake script for you, but you still need logic for your own (vcpkg) packages.
Those are not native (on desktop) in any sense of the word. They don't use native controls. For that, you want WX or SWT, but those come with their own sets of problems.
On Windows, it's not even obvious what native is any more, even Microsoft just uses Web views. Mac is a bit better, but there are still 4 UI libraries to choose from (AppKit, UIKit through Catalyst, native SwiftUI and Catalyst SwiftUI).
I'm personally a fan of AppKit and Win32, but those are "dated" apparently.
I am working on the UI library and bindings for Go. Still not finished, but currently, the same app can be compiled for Win32, Cocoa, GTK2, GTK3, GTK4, Qt5, Qt6, and Motif. There is a web browser control, a GL canvas, and a regular canvas. I still work on the native table control, though.
IUP has custom-drawn controls for tables and cells (additional controls), and it uses another CD (canvas draw) library for that, not internal IUP Draw functions. I also started rewriting that to use the core IUP drawing functions instead. I also added a few more drawing functions, for rounded rectangles, bezier curves, and gradients. But ALL drivers, including Motif, have native table controls, so I really want to add one.
Edit: Also, the GLcanvas control now has an EGL driver, with native Wayland support for GTK3, GTK4, and Qt6 (needs private headers). I modernized a bit of everything, added support for APPID, DARKMODE, etc. Linux uses xdg-open rather than hardcoding browsers. Win32 driver is not using the Internet Explorer web control but the WebView2 with custom loader, on GTK, you do not have to worry about the WebKitGTK, it will find the correct library with dlopen, etc, etc. But, there is still a lot to do.
gp was using a more restrictive definition of "native controls". I.e. "o/s builtin UI controls" vs "framework canvas painted elements".
For Windows, "native" would be the classic Win32 UI "common control" elements from "Comctl32.dll"[0] that is directly used by older GUI frameworks such as Windows Forms. Those map to classic Win32 API CreateWindow(L"BUTTON", ...). In contrast, the newer frameworks of WPF and Xamarin Forms and Qt Quick "paints controls on a canvas" which are not "native" and makes every app UI look different instead of standardized "look & feel" of common controls.
But others include custom-canvas painting UI objects as "native" -- as long as it's not Electron.
I wouldn't exactly call Flutter native. It uses its own rendering engine and doesn't necessarily behave like operating system native controls. It is not really different from using electron.
Using electron at least uses some UI primitives from chromium. Flutter has thrown away all the usability and robustness of existing components and just reimplemented everything. It absolutely is different from electron
"Native" seems to mean different things to different people. I'm mostly with you on this, but the tides are turning. In any case, the other 3 do use real native widgets.
Making good GUI software requires a lot of iteration and trial and error before you're satisfied with the UI and UX. With a web-based tech, you make a change, auto reload triggers, you see the change almost instantly, making tweaking very easy. If you're working with a large Qt codebase, every little change to a header file requires a long ass compile times. It's really frustrating when you spend an hour just tweaking a few controls when you know it could have taken 5 minutes. Also, the reactive model as seen in web frameworks like React or Vue is much superior to the typical flow of state management in retained mode GUI applications in desktop frameworks. Until we have a decent solution that solves these problems, people will continue using tech like Electron or OS web views.
It takes half a day to implement proper hot reload in QtQuick, which also has all the reactive features. Even less now that AI can just write it for you, and it’ll be more performant than Vite dev builds.
Coming to desktop app development from the web, I’ve got most of the same conveniences I’m used to like GammaRay as the inspector. The only real difference is I’m willing to wade through cmake and linking errors.
Even QWidgets is still super fast to develop with if you’re using PySide (although hot reload is a bit more difficult to implement and distribution becomes the nightmare).
But I already know how to write a web app, I don’t know how to write a desktop app. It’s faster to just write and wrap a web app, and as far as most people can tell, it works just fine.
To me, this argument always sounds like someone is being forced or threatened into creating a desktop app. It was never supposed to be easy; the goal is to create an app that users would want and will actually use.
I really enjoyed building small apps with wails.
Even though people would prefer that we all used native UI frameworks, the DX is simply incomparable to that of web technologies.
And for most apps using browser based rendering won't be an issue. People often underestimate how optimized mondern browsers really are. And because Chromium is not shipped the bundle size is managable.
Not wanting to use JS on the backend I tried both Tauri and Wails and found the simplicity of Go to just work perfectly for my use-cases
Hmm, the Go GC is really quite capable, so I wonder what kind of pathological load it was being presented with. Even then, when the GC "fails" it means elevated CPU load from collection.
The main thing I can think of would be the application "leaking" by having unintentional references (or worse, actually leaking through cgo bindings), or trashing the allocator to cause temporary spikes in between cleanups.
However, while I don't think Go was actually to blame here, I would never use native UI bindings to a language that isn't 1:1 compatible with the original design and memory management principles, as such bindings get disproportionaly large and complex. It just sets you up for a bad time.
Recently for a specific purpose I was reviewing options including Tauri, various WebView bindings, and in the end had to admit that Electron is probably the best approach in terms of feature set, development speed, etc.
This sounds to me like leaking resources from the binding. Which binding did you use and how did you manage it's lifetime?
Which bindings did you use ?
0 - https://fyne.io
Other discussions: - https://news.ycombinator.com/item?id=31785556 - https://news.ycombinator.com/item?id=19478079 - https://news.ycombinator.com/item?id=22291150
1. The state of C++ is not great. Few developers, C++ footguns, complicated build systems, and generally slow progress, see my https://arewemodulesyet.org/
2. How Qt presents and licenses itself. Either you go LGPL or you have to pay big money for a commercial license, which will then infect all other apps as well. For example, when you have two Qt apps that talk to each other you must license _both_ commercially.
3. The split of Widgets and QML makes the ecosystem fragmented, because Widgets will never die. Even the Qt devs themselves are split about this. You can see this when example code for a new feature uses Widgets. QtCreator is also a nice example, where they reverted some new QML code quite a while ago and have not substantially added any new QML code since then.
4. Tooling: We use QML for everything and the tooling is not great. The language server is still super flaky and breaks, and developer tooling like the Chrome Dev Tools is virtually nonexistent.
5. Packaging is still also not great but has gotten better in the last few versions where Qt creates a deployment cmake script for you, but you still need logic for your own (vcpkg) packages.
On Windows, it's not even obvious what native is any more, even Microsoft just uses Web views. Mac is a bit better, but there are still 4 UI libraries to choose from (AppKit, UIKit through Catalyst, native SwiftUI and Catalyst SwiftUI).
I'm personally a fan of AppKit and Win32, but those are "dated" apparently.
https://github.com/gen2brain/iup-go
For Windows, "native" would be the classic Win32 UI "common control" elements from "Comctl32.dll"[0] that is directly used by older GUI frameworks such as Windows Forms. Those map to classic Win32 API CreateWindow(L"BUTTON", ...). In contrast, the newer frameworks of WPF and Xamarin Forms and Qt Quick "paints controls on a canvas" which are not "native" and makes every app UI look different instead of standardized "look & feel" of common controls.
But others include custom-canvas painting UI objects as "native" -- as long as it's not Electron.
[0] https://learn.microsoft.com/en-us/windows/win32/controls/com...
GTK and KDE controls are native to GTK and KDE.
I can also think of QT and GTK for other languages too.
Coming to desktop app development from the web, I’ve got most of the same conveniences I’m used to like GammaRay as the inspector. The only real difference is I’m willing to wade through cmake and linking errors.
Even QWidgets is still super fast to develop with if you’re using PySide (although hot reload is a bit more difficult to implement and distribution becomes the nightmare).
https://github.com/mappu/miqt
Ya gotta be practical.
And for most apps using browser based rendering won't be an issue. People often underestimate how optimized mondern browsers really are. And because Chromium is not shipped the bundle size is managable.
Not wanting to use JS on the backend I tried both Tauri and Wails and found the simplicity of Go to just work perfectly for my use-cases
With Electron, for example, a stripped down Chromium is shipped. So what does the web view rendering with this package?
Why is that hidden behind a click and two walls of text?