File systems have always attracted more research than other parts of the operating system and that is still the case. Entire conferences such as FAST, MSST, and NAS are devoted largely to file and storage systems.

A considerable amount of research addresses the reliability of storage and file systems. A powerful way to guarantee reliability is to formally prove the safety of your system even in the face of catastrophic events, such as crashes (Chen et al., 2017). Also, with the rising popularity of SSDs as the primary storage medium, it is interesting to look at how well they hold up in large enterprise storage systems (Maneas et al., 2020).

As we have seen in this chapter, file systems are complex beasts and developing new file systems is not easy. Many operating systems allow file systems to be developed in user space (e.g., the FUSE userspace filesystem framework on Linux), but the performance is generally much lower. With new storage devices such as low-level SSDs arriving on the market, the need for an agile storage stack is important and research is needed to develop high-performance file systems quickly (Miller et al., 2021). In fact, new advances in storage technology is driving much of the research on file systems. For instance, how do we build efficient file systems for new persistent memory (Chen et al., 2021; and Neal, 2021)? Or how can we speed up file system checking (Domingo, 2021)? Even fragmentation creates different issues on hard disks and SSDs and requires different approaches (Kesavan, 2019).

Storing increasing amounts of data on the same file system is challenging, especially in mobile devices, leading to the development of new methods to compress the data without slowing down the system too much, for instance by taking the access patterns of files into account (Ji et al., 2021). We have seen that as an alternative to per-file or per-block compression, some file systems today support deduplication across the entire system to prevent storing the same data twice. Unfortunately, deduplication tends to lead to poor data locality and trying to obtain good deduplication without performance loss due to lack of locality is difficult (Zou, 2021). Of course, with data deduplicated all over the place, it becomes much harder to estimate how much space if left or will be left when we delete a certain file (Harnik, 2019).