strange behavior of --inplace on ZFS

[Pavel Herrmann]

Hi

I am extending my ZFS+rsync backup to be able to handle large files (think 
virtual machine disk images) in an efficient manner. however, during testing I 
have found a very strange behavior of --inplace flag (which seems to be what I 
am looking for).

what I did: create a 100MB file, rsync, snapshot, change 1k in random location, 
rsync, snapshot, change 1K in other random location, repeat a couple times, 
`zfs list` to see how large my volume actually is.

the strange thing here is that the resulting size was wildly different 
depending on how I created the initial file. all modifications were done by the 
same command, namely
dd if=/dev/urandom of=testfile count=1 bs=1024 seek=some_num conv=notrunc

situation A:
file was created by running 
dd if=/dev/zero of=testfile bs=1024 count=102400
the resulting size of the volume is approximately 100MB times the number of 
snapshots

situation B:
file was created by running
dd if=/dev/urandom of=testfile count=102400 bs=1024
the resulting size of the volume is just a bit over 100MB

the rsync command used was
rsync -aHAv --delete --inplace root at remote:/test/ .

rsync on backup machine (the destination) is 3.1.0, remote has 3.0.9

there is no compression or dedup enabled on the zfs volume

anyone seen this behavior before? is it a bug? can I avoid it? can I make 
rsync give me disk IO statistics to confirm?

regards
Pavel Herrmann