This seems like a different problem than people are making it out to be.
SMR drives have slow random writes and paper over it with caching, until the cache gets full. Then in theory what should happen is that the actual write speed of the drive is exposed. That means resilvers would take a long time, but they should still finish.
What seems to be actually happening is that some of these drives have a firmware bug such that when caching is enabled and the cache gets full, the write speed drops to zero. The system then regards the drive as faulty and boots it out.
So it seems like this should be solvable with a firmware update that causes the drive to behave differently (slow rather than stopped) when the cache gets full.
This also implies that some other SMR drives with different firmware might not behave like this, and that it might not happen with ordinary RAID rebuilds as opposed to ZFS resilvering because then the writes should be almost entirely sequential (i.e. what SMR is good at) as opposed to ZFS which is more random.
It doesn't have to stop completely to trigger the problem case, does it? It seems like just being slow enough to trigger the containing system's timeout response would be enough.
SMR drives shouldn't be that slow. They're slower than PMR drives but it shouldn't be by so much that individual writes are taking tens of seconds.
What's probably happening is one of two things. Either the cache gets full and the drive is blocking while it flushes the cache to the disk, or the drive is advertising that it can do a large number of simultaneous write operations which it can't actually do all of in a reasonable amount of time when the cache is full, and then the ones queued last time out. In the first case they could have the firmware continue to process uncached writes when the cache is full, in the second case don't advertise the ability to do as many simultaneous writes.
Another alternative might be for the system to use a longer timeout value for these drives, but whether that's reasonable depends on how long it would actually have to be.
