+1. In my limited testing, when I had different flags (default 128k recordsize, no compression)… I was filling the metadata device 3x-4x as fast - so it would have filled before the HDD. When I used recordsize=512k and compress=lz4 then 5GB/TB worked. There wasn’t a lot of wiggle-room (6GB/TB maybe would have been safer) but the HDD %-full was climbing faster than metadata %-full… which in the end is all you need.
But that was metadata-only: no small files. And for a 4TB node copied from ext4 to zfs.