ZFS swap space can trigger OOM much earlier.
killing what? I would insert "(random?) processes" here.
Contractions are discouraged in man pages, so use "does not" here.
It does not matter how much memory is required for write. What matter is that each write requires allocation of memory.
If the VM high watermark levels are tweaked, OOM from ZFS on swap can be entirely avoided. Don't know if this is relevant?