This version of csnap-agent borrows the server template from csnap.c to
accept and serve multiple local connections from cluster devices. So
only a single named socket is needed for all csnap devices on a node
connecting to the same snapshot server.
The snapshot agent also needs to communicate with other cluster
components:
- Connection manager and/or resource manager, to request a snapshot
server connection
- Resource manager may request that the snapshot agent start a
snapshot server.
The same polling harness can handle this other traffic easily enough.
The question is, should those other components be taught to make their
connections to the named socket, in which case we get to reuse the
connection code as well? Should our csnap agent take it upon itself to
connect to those other components, or should it make its socket name
known and wait for connections? The latter makes our code smaller, but
is it right?
If we use the named socket for connection with other components, the
implication is that we will only communicate with local agents of those
components. On the other hand, if we use inet connections, we then
have to teach the csnap agent how to fail the connections over, which
really seems like somebody else's job.
I dunno. A reasonable next step is to manually feed the csnap agent
some csnap server addresses over a pipe instead of, or in addition to,
taking a server address on the command line the way it does now. This
will allow us to test server failover, at least the flavor of server
failover that doesn't need to consult the connection manager.
Regards,
Daniel
