While looking at
KUDU-2131, it occurred to me that we could solve the "batched fsyncs could lead to a tablet copy session timeout" problem entirely if we simply deferred the commit of the blocks until after we're done with the session. That would mean deferring it until after DownloadWALs. That's probably a good thing to do anyway:
- If DownloadWALs fails there's no reason to pay the fsync price and commit all those blocks.
- While DownloadWALs is running the kernel has more time to eagerly flush the blocks, so the fsyncs at the end could be cheaper.