[rds-devel] failover

[Or Gerlitz]

Xiaohan Qin wrote:
> Is failover supported in RDS? If yes, does the support come from the protocol implementation or 
> by layers beneath it. Thanks. 
>   
RDS supports failover and the support comes from the protocol 
implementation, that is each message kept till (protocol level) ACK is 
received for it,  where following connection failure all inflight (un 
acked) messages are retransmitted.

This holds for the bcopy protocol. The zcopy (rdma) protocol 
transactions are unreliable and hence fail-over are not supported for 
them, in that case the app using rds has to do something.

Being an rdma-cm consumer, rds address resolution (resolving IP address 
to HW address) uses the IP stack neighbour discovery code, so for  
fail-over to work, there's a need for the arp to be sent through a 
secondary port in the case of local failure, and to be resolved to a 
peer secondary port in the case of remote failure. This is achieved 
through standard  Linux bonding in its HA mode.

Or.