I see.
> 2) As long as TCP close is precise, then this isn't really a new failure
> mode, since the receiver node could always crash after receiving the
> data
I think things is slightly different. in cited case, both receiver's
kernel and application-process crash at the same moment. in detour's
case, application-process is still alive after router crashed.
> 3) It is possible as Tom said, to replicate hard state between detour
> boxes if this is a real issue (I'm not so excited about this idea
> though)
replication is generally costly, and if original and replica are set
up in the same place to reduce the cost, natural disaster will broke
both router at the same moment.
as Stefan mentioned in 1), if we classify the packet and apply the
eager-ACK only to short lived flows (interactive flow is preferable),
replication might be needles.
thanks.
-- Kenichi Ishikawa