Using syncreplica -export
to create common packets for multiple destinations
will reduce the number of packets sent
by every given server,
but not the number of packets received
(and thus only part of the number of packets shipped,
e.g. through firewalls and other restricted bandwidth channels).
In order to reduce this second component of the shipping load,
we have set up an import/export hub,
i.e. a vob server which imports all the packets it receives,
but doesn't forward them after succesfull reception.
Instead, it deletes them,
and relies upon it's own reexporting the changes later.
This means that by default, we want to route the packets through such a hub even when the destinations would be accessible, and thus a direct route would be possible. This is again a winning tradeoff in most cases.
One downside with this strategy
is however that it should allow for exceptions in interactive mode,
when the additional delay incurred by routing via the hub,
and waiting for the reexport,
is not acceptable
(one obvious case is chmaster commands
-- in my understanding, chepoch -actual
and similar commands do by-pass the shipping already).
Of course, we have already the workaround to use a dedicated client, with different ROUTE specifications in its own shipping.conf, to perform such commands.
But this is anyway errorprone and inconvenient.
-> The RFE would be for storage class dependent routes.
This could thus support various strategies, such as: