libfuse VS php-fuse

I suspect if you've run into problems with a lot of things built on FUSE, the problem is FUSE.

Yes, s3fs and sshfs can both leave the system in an unstable state. For example, there can be a dead mount which is impossible to unmount, and in severe cases, blocks a clean reboot.

A file system in user space (or in network space) should NEVER break the system, no matter what happens in user space (or in network space). Most network file systems try to respect this (albeit with mixed success). FUSE does not.

I'm not claiming FUSE cannot be made to work. Just that it's very bad since (1) plenty of smart people clearly failed to do so (2) the badness it leaves behind should be more than it's permitted to.

I can point to specific issues, but at the end of the day, that's neither here nor there. At the end of the day, something like:

https://github.com/libfuse/libfuse/blob/master/example/poll....

Is about a hundred times more complicated than it should be. It should not require memsets, pthread mutexes, or flags, and should probably have an implementation in a modern, high-level language. To a large extent, that's the point of moving things out of the kernel.

I'm even perhaps okay with being permitted to do low-level operations for a particularly performance-constrained subsystem, but that's not 95% of the uses of something like FUSE.

Footnote: I actually enjoyed writing low-level code like this a lot, when computers were in the single-digit to triple-digit MHz range, and we didn't need to worry about people breaking in over a ubiquitous worldwide internet, but I left that mindset behind decades ago. Right now, I want code to be stable, simple, auditable, and secure.

I wrote the original version for a previous employer mostly in Python.

I was about to recreate a new version in Rust. And started with fixing up libfuse https://github.com/libfuse/libfuse/pulls?q=author%3Amatthias... and the Rust equivalent https://github.com/cloud-hypervisor/fuse-backend-rs/pulls?q=...

Your project is also interesting. I don't plan on ever adding write support. The old Python version was already using git as a library via gitpython, instead of shelling out via the command line. The new version will use Rust's gix.

Performance, even for the old Python version, was pretty decent. That probably came from using git via a library and being careful about fuse caching. The old version also already supported opening arbitrary commits, tags and branches, they were represented as different folders.

s3fs allows Linux, macOS, and FreeBSD to mount an S3 bucket via FUSE(Filesystem in Userspace). s3fs makes you operate files and directories in S3 bucket like a local file system. s3fs preserves the native object format for files, allowing use of other tools like AWS CLI.

FUSEs your files

The situation is much worse than I had imagined; the parent project, libfuse, is also in need of a maintainer.

https://github.com/libfuse/libfuse

The introduction of FFI in PHP was done with the reference to LuaJIT. As a result, FFI's CData is a completely dynamic type, similar to stdClass, and does not get proper support from the IDE or static analysis. For this reason, php-fuse has introduced a muddy mechanism to map CData to PHP classes. https://github.com/sj-i/php-fuse/tree/master/src/FFI It is a shame that CData is such a dynamic type since PHP nowadays is a language that easily benefits from static types, unlike Lua or Ruby. I'm hoping that I or someone else can come up with a better solution, such as allowing C types to be mapped to user-defined CData subclasses, as the idea was seen in the early discussions when FFI was introduced. https://externals.io/message/102036#102050