I am testing keep alive on PostgreSQL 9.2 and 9.3 on Centos 6.X .
As I tested and saw until now, the keep alive functions as follow [if I understand correctly and it's not a bug] :
When a connection it's in idle state or in idle in transaction, if the connection with client it's broken for a number of keep alive, the backend will be terminated.
By the way if this could be ok in an OLTP enviroment, because the average time of a query is << the time of keep alive, in a DWH enviroment could be a problem.
Imagine your application server, where there is an ETL, will go down for 1 minute and your transactions are still running on the DWH database, that transactions could run for hours before the
keep alive will terminate them, because they are in transaction state and not idle or idle in transaction.
