After number of connects/transactions/disconnects using JDBC, coordinator stacks up the queries and non-existing connections in the pg_stat_activity like
application_name | pgxc
client_addr | ::1
client_port | 52413
backend_start | 2013-02-14 13:16:28.183573+01
query_start | 2013-02-14 13:20:41.063212+01
state_change | 2013-02-14 13:20:41.06787+01
waiting | f
state | idle
query | SET SESSION AUTHORIZATION DEFAULT;RESET ALL;
the query is cleary executed from agent_reset_session(PoolAgent *agent) in poolmgr.c.
Transactions are succesfult completed but some connections are never returned to the pool it seems, so after # of cycles system refuses new connections.
The tests are too complicated to post here, hint could be that SET SCHEMA is called for each of the nerw connections - remember there was some bug related to SET.