On Sat, Feb 1, 2025 at 8:40 PM Dale <rdalek1...@gmail.com> wrote: > > Rich Freeman wrote: > > > > Now, if you were running btrfs or cephfs or some other exotic > > filesystems, then it would be a whole different matter, > > I could see > some RAID systems having issues but not some of the more advanced file > systems that are designed to handle large amounts of data.
Those are "RAID-like" systems, which is part of why they struggle when full. Unlike traditional RAID they also don't require identical drives for replication, which can also make it tricky when they start to get full and finding blocks that meet the replication requirements is difficult. With a COW approach like btrfs you also have the issue that altering the metadata requires free space. To delete a file you first write new metadata that deallocates the space for the file, then you update the pointers to make it part of the disk metadata. Since the metadata is stored in a tree, updating a leaf node requires modifying all of its parents up to the root, which requires making new copies of them. It isn't until the entire branch of the tree is copied that you can delete the old version of it. The advantage of this approach is that it is very safe, and accomplishes the equivalent of full data journaling without actually having to make more than one write of things. If that operation is aborted the tree just points at the old metadata and the in-progress copies are inside of free space, ignored by the filesystem, and thus they just get overwritten the next time the operation is attempted. For something like ceph it isn't really much of a downside since this it is intended to be professionally managed. For something like btrfs it seems like more of an issue as it was intended to be a general-purpose filesystem for desktops/etc, and so it would be desirable to make it less likely to break when it runs low on space. However, that's just one of many ways to break btrfs, so... :) In any case, running out of space is one of those things that becomes more of an issue the more complicated the metadata gets. For something simple like ext4 that just overwrites stuff in place by default it isn't a big deal at all. -- Rich
