Using --proxy-pool-no-change-user crashes MySQL database

[Brian Chang]

Hi, I used a lua script to keep back-end connections to MySQL database server persistent (i.e. to form a back-end connection pool), while allowing the clients to connect to MySQL-proxy by allocating a back-end server connection from the pool.

When using "--proxy-pool-no-change-user" option, hoping each new client connection does not need to incur the extra overhead to authenticate with back-end db server, it actually complains about packet_id unexpected (because con->client's last_packet_id and the actual incoming packet_id is different). So I force the con->client to reset in MySQL-proxy code to fix that. But then another problem arises: the backend MySQL DB server got crashed after I stress tested using clients issuing different SQL statements. My suspicion is that when "network_mysqld_con_reset_command_response_state()" and/or "network_mysqld_con_command_states_init()" are used to clean up connection's state in order to let a server connection take a new query from client side, the server connection can be in a state in which server is still expecting some data to complete a previous query. Now, by resetting the state of server connection and by being fed with some new query from a new client connection, the server becomes confused. I actually see the backend MySQL server crashes at a point when it tries to roll-back, and somehow gets segfault. I know "--proxy-pool-no-change-user" leaves temporary tables on the server connection per document, but do not realize it can actually crash and corrupt back-end MySQL database (i.e., restarting MySQL database server fails, and needs to go thru recovery process).


In short, looks like "--proxy-pool-no-change-user" is still not a robust feature, given that it can literally corrupt back-end MySQL server.