If we don't defer and try again, we just get stuck.
I've had that patch locally for a long while, but it's hard to force
the connection loss to happen at the exact moment for this. The fix is more
based on logical reasoning based on similar issues here and in FatCRM.