unresolved MDL deadlock with limited innodb thread concurrency

[Seppo Jaakola]

Hanging happens in 1.0beta evaluation with following use case:

* innodb_thread_concurrency=32
* ~50 client connections running short autocommit queries
  (UPDATE,DELETE,INSERT)
* one UPDATE with INNER JOIN, updating 'queue' table
* one DDL running: CREATE TABLE IF NOT EXISTS queue(...)

The DDL stays hanging in state: "Waiting for table metadata lock", killing this query resolves the deadlock.

[Seppo Jaakola (seppo-jaakola)]

development head has a number of fixes for DDL processing with respect to MDL usage. These fixes mostly dealt with issues found with FTWRL..
Change set 3567 adds one more fix for this: now BF thread can use wsrep_abort_thd() to kill local state MDL victim.

[Seppo Jaakola (seppo-jaakola)]

Hanging happens with 2.0 RC, with following regression test:

sqlgen --host gw --port 3307 --users 12 --create 1 --duration 3600 --rows 1000 --ac-frac 10 --rollbacks 0.1 --alters 0.001

and cluster topology: 2 nodes MM load balanced

[Seppo Jaakola (seppo-jaakola)]

MDL caused some issues, here is a potential fix: http://bazaar.launchpad.net/~codership/codership-mysql/wsrep-5.5/revision/3702