On 4 Sep 2023 13:57 -0700, from dpchr...@holgerdanske.com (David Christensen): > * I am using zfs-auto-snapsnot(8) for snapsnots. Are you using rsnapsnot(1) > for snapshots?
No. I'm using ZFS snapshots on the source, but not for backup purposes. (I have contemplated doing that, but it would increase complexity a fair bit.) The backup target is not snapshotted at the block storage or file system level; however, rsync --link-dest uses hardlinks to deduplicate whole files. > * du(1) of the backup file system matches ZFS properties 'referenced' and > 'usedbydataset'. This would be expected, depending on exact specifics (what data du traverses over and what your ZFS dataset layout is). To more closely match the the _apparent_ size of the files, you'd look at e.g. logicalreferenced or logicalused. > * I am unable to correlate du(1) of the snapshots to any ZFS properties -- > du(1) reports much more storage than ZFS 'usedbysnapshots', even when scaled > by 'compressratio'. This would also be expected, as ZFS snapshots are copy-on-write and thus in effect only bookkeep a delta, whereas du counts the apparent size of all files accessible under a path and ZFS snapshots allow access to all files within the file system as they appeared at the moment the snapshot was created. There are nuances and caveats involved but, as a first approximation, immediately after taking a ZFS snapshot the size of the snapshot is zero (plus a small amount of metadata overhead for the snapshot itself) regardless of the size of the underlying dataset, and the apparent size of the snapshot grows as changes are made to the underlying dataset which cause some data to be referenced only by the snapshot. In general, ZFS disk space usage accounting for snapshots is really rather non-intuitive, but it does make more sense when you consider that ZFS is a copy-on-write file system and that snapshots largely boil down to an atomic point-in-time marker for dataset state. (In ZFS, a dataset can be either a file system optionally exposed at a directory mountpoint or a volume exposed as a block device.) -- Michael Kjörling 🔗 https://michael.kjorling.se “Remember when, on the Internet, nobody cared that you were a dog?”