So I think you should use 3.3.2, to avoid troubleI encountered two additional bugs and have rolled back to 3.3.9.
The first is a HEADER_DONE not set error on a single dataset of a ZFS pool. We're dumping in the neighborhood of 20 small datasets (one per user) on a server running zfsonlinux, and each run resulted in precisely one of these failing with the HEADER_DONE not set error. It was not the same dataset each time. I've never encountered this error before, and digging into the code it wasn't immediately apparent why this was happening. The error did not occur with last night's 3.3.9 run.
The other error, which was a (literal) showstopper, was a taper hang. The run two nights ago never finished. At the time of investigation there were 8 idle dumper threads and 3 taper threads, with the driver reporting that the taper was writing. Two of the tapers were in a futex wait channel and the third was in a poll_s... wchan. I ultimately sent sigterm to this last taper and amanda immediately wrapped things up with a taper protocol error. The logs didn't report anything unusual prior to the hang and there's no evidence of any i/o problems with the hardware or network. And the problem did not occur with 3.3.9.
So it seems 3.4 is not ready for primetime, at least not for us.