| Supervisor | Niklas Gollenstede |
| IBR Group | VSS (Prof. Dietrich) |
| Type | Bachelor Thesis |
| Status | tentative |
General-purpose filesystems (FSes) are designed to handle a wide variety of use cases, and therefore have to support persistent in-place (or copy-on-write) modifications of files and directories. Their data structures and algorithms therefore have to be designed with safe mutability in mind, which increases complexity, may limit access concurrency and storage density, and must handle cache staleness. Read-only file systems can avoid these issues, and read-only application deployments are becoming increasingly popular. A problem that remains is that applications or systems deployed, for example, via Docker, Nix, OSTree, or as Flatpak are read-only in their execution, but still need to be written once for installation and garbage collected later. As such, they currently need to be installed either directly to general-purpose FSes, or as read-only FS images on top of a general-purpose FS. A FS that allowed each file and folder to be written only once and in bulk, and then treats them as read-only until eventually discarding them could support deployable read-only applications directly, while potentially retaining many benefits of read-only FSes. This thesis will investigate modifications to EROFS, the current state-of-the-art read-only FS, to make it support the write-once pattern. While the resulting FS should also directly work as efficient live storage for Docker/OCI, OSTree, Flatpak and others, some aspects should specifically be optimized for the use as Nix store -- managed by a single daemon but used by many running applications. | |