Added counter clean-up code to readwritesplit closeSession. If operation counters have leaked, that is, due to some error in query routing some counter have been left positive, it is zeroed at session ending. This ensures that there is no garbage in global counter even if individual session wouldn't clean up its counters.
Fixed bug in session command resul handling. In case where backend sent error message the session command cursor wasn't updated properly.
Added check to bref_clear_state, if bref's waiter counter would go negative, decrement to global operation counter is skipped.
The flag is part of the active operation counting, which is utilized in load balancing. The active operation count per backend is used by default as criteria when router chooses to which backend a query should be routed.
All prepare commands are executed in every backend server currently connected.
All executes are routed to master. If stmt type was recorded in prepare phase in rwsplit router, read-only stmts could be routed to slaves.
COM_STMT_PREPARE gets arbitrary number of response packets from backend database. Since statements are prepared in every backend server and only one multi-packet response can be replied to client, redundant multi-packet responses are discarded. This is done in router. Mechanisms from session command handling are utilized with little changes: router must identify when response consists of multiple packets so that it knows to calculate the number of packets in response and that it is able to discard correct number of packets.
Information to the reply-handling router is provided by backend protocol, which includes a ordered list of commands of commands sent to protocol-owning backend server. A command is stored to protocol struct in mysql_backend.c:gw_MySQLWrite_backend if the statement buffer's type has GWBUF_TYPE_SINGLE_STMT set in mysql_client.c:route_by_statement. GWBUF_TYPE_SINGLE_STATEMENT indicates that there is single statement in the buffer, as opposite to Read Connection router, which accepts streaming input from client.
New rules for selecting Master/Slaves are based on MySQL depth in the
replication tree.
The only Master is the root server with SERVER_MASTER bit set
All other servers even with both SLAVE and MASTER bit will be selected
as slaves