A lot of the hard to test and validate problems with filesystems is finding all the edge cases and testing them, and they will be able to skirt a lot of that to start by reusing XFS as the main filesystem layer and focusing on the block and redundancy layers to start with (layers 0-6 in the PDF) and doing smarts with stratisd. The big thing they will need to add to XFS or emulate is the ability to online shrink an existing filesystem.
That should free them up to try and do some innovation on the block / redundancy side and potentially develop a new more flexible replacement for LVM2 that meets the needs of the current filesystems and hybrid volume oriented systems. The storage landscape is changing so fast now with SSD, NVMe and Optane etc that have speeds that demolish the assumptions previous filesystems were designed around, and 10TB+ drives being cheap so that average servers having 100's of TB directly attached is going to be normal and having updated tools that can make use of that much space in an intelligent way is going to be a key development area.
It would also be interesting to see some development around systems that can start on a single physical node and then expand into a clustered system for increased redundancy and performance, and some greater hybridization of object based and file based storage. Taking the needs of the various virtualization environments (Docker/Xen/kvm/esxi) and exposing the features they have had to build on top of existing POSIX layers in the filesystem directly could help advance the filesystem landscape as a whole.
How we are using and thinking about storage has changed so much in the past few years that I am glad open source friendly firms are still putting resources behind work like this.
That should free them up to try and do some innovation on the block / redundancy side and potentially develop a new more flexible replacement for LVM2 that meets the needs of the current filesystems and hybrid volume oriented systems. The storage landscape is changing so fast now with SSD, NVMe and Optane etc that have speeds that demolish the assumptions previous filesystems were designed around, and 10TB+ drives being cheap so that average servers having 100's of TB directly attached is going to be normal and having updated tools that can make use of that much space in an intelligent way is going to be a key development area.
It would also be interesting to see some development around systems that can start on a single physical node and then expand into a clustered system for increased redundancy and performance, and some greater hybridization of object based and file based storage. Taking the needs of the various virtualization environments (Docker/Xen/kvm/esxi) and exposing the features they have had to build on top of existing POSIX layers in the filesystem directly could help advance the filesystem landscape as a whole.
How we are using and thinking about storage has changed so much in the past few years that I am glad open source friendly firms are still putting resources behind work like this.