NeutralinoJS: Lightweight Electron alternative using native browser controls

Even in the browser itself, there's a clear delineation between the set of components that change quickly, and the set of components that change slowly—a factoring of layers calling out to be done, that everyone is ignoring. There's no reason for a whole new web browser to be shipped if only the quickly-changing part has changed.

Analogy: why would you ship a new copy of X/Wayland just because your Desktop Environment has a new release? The new DE version uses the same old compositor, because the compositor doesn't change much. So why would it make sense for a compositor to be part of the DE and embedded in the release package for the DE, rather than just requesting the OS to make available to it the services of a compositor of a specific [ABI protocol] version? Even if the compositor and the DE have the same maintainers, it'd make far more sense for them to be two separate projects, with one just depending on the other.

Or, to put this another way: why should ChromeOS devices (= another build target of the Chromium codebase) replace the whole OS image every time some new namespaced CSS layout rule is being incubated? That stuff has nothing to do with running an OS. The parts of Chrome that are an OS—which includes a lot of the stuff that gets installed as "browser" on other OSes!—should really live as OS services, separate from "the browser" (really just the renderer + VM.)

Microsoft took a while but eventually learned their lesson about this, and these days Edge is just an app and updating it doesn't require a restart of your computer. Part of what enabled that was delineating the slower-moving parts of the browser (e.g. the network stack) from the faster-moving parts, and letting the slower-moving parts live in the OS while the faster-moving parts lived in the app. "Edge" is just a renderer+VM. So is Safari, for that matter—what you download when you download "Safari Technology Preview" on macOS is just a renderer+VM, which relies on the same OS-provided libraries for e.g. network caching that the system Safari installation does.

The context here is that "the framework" does things 'for' the application it is hosting that the application cannot control, caching various things in memory in a way that makes sense for a browser, but not very much sense for a custom application.

Try this: open a new Chrome instance; look at its memory usage; then load a bunch of tabs (try an Open All on a bookmarks folder), close them all again, and then look at Chrome's memory usage again. It will have increased.

Chrome isn't leaking memory; instead, it is doing the same thing that an Operating System does: weakly reserving memory to optimistically cache stuff, releasing it again if there's enough memory pressure.

This works okay if you only have Chrome itself running. (Even then, its cache fights a bit with the OS page cache. It's certainly no Postgres, intentionally getting the OS page-cache to do its caching for it.) But as soon as you have two separate Chromium instances running (e.g. Chrome itself, plus an Electron app), then each one is going to try to "optimistically" cache as much stuff as it can, until it runs into the memory pressure created by the other instance caching as much stuff as it can, and so each instance will unload just a bit to let the other one cache just a bit more—back and forth—forever. Together, they thrash memory, and it becomes much harder for them to actually accomplish the "weak" part of "weak reservation", instead ending up hoarding all the memory between them.

This is the pretty much the same problem you see if you try to e.g. run two memory-heavy JVM processes (e.g. ElasticSearch) on the same system. Chrome is just one of the few times you'll see this "recapitulation of memory management within the runtime" effect client-side, rather than server-side.

I haven't seen this one personally, but I get the same crash-loop from the Discord app.

Weirdly enough, the crashes don't seem to affect any of the app's functionality.

Mind you, there's (sometimes a little, sometimes a lot) more to Electron than a webview. Some apps are running quite a bit of un-sandboxed native code—sometimes even forking off other native processes†—to do what they do. Keybase is, IIRC, one of the apps that are quite large on the "native" side.

† I once designed a system that has Electron install an Erlang release to the client, register it as a service, and start it. The Erlang node runs a locally-bound web server; the Electron app then visits that web server. All the data-wrangling happens in Erlang land, with Electron just serving as an HTML renderer for the pages coming from the local server. And yet Electron's native side is essential, here, because it ultimately manages the Erlang release (installing it, updating it, etc.) Erlang doesn't have any good tooling for doing client-native stuff on its own; Erlang devs find the whole idea of deploying an Erlang release to a client machine funny.

You can't write your GUI in those languages.