$ bzr log -r 6230
------------------------------------------------------------
revno: 6230
committer: Venkatesh Duggirala<email address hidden>
branch nick: mysql-5.6
timestamp: Thu 2014-10-30 22:33:43 +0530
message:
Bug#19704825 TEMPORARY SLAVE TYPE CONVERSION TABLES RELEASED TO EARLY
Problem: The memory used in preparing slave type conversion temporary table
is getting released early and causing unexpected results
Analysis: As part of bug#18770469 fix, We introduced an event
m_event_mem_root (a special mem root), added to Log_event
class. While server is creating the temporary table, the memory needed
is allocated from this special mem root which will be freed in ~Log_event()
i.e., the scope of this memroot is one event. But it could happen
that in some cases, server might need to access this
conversion temporary table for next following event.
For eg: A nested row event (insert is causing insert in a trigger)
In this situation, the memory is getting delayed too early
and causing issues when the server is trying to access the temporary
table inside the trigger.
Fix: We cannot use a mem_root whose scope is limited to an event
execution in this situation. With some further analysis, found out
that clearing a thd->mem_root at the end of statement (upon applying
an event which has STMT_END_F flag) will solve out of memory problem
while running a long transactions (bug#18770469) and will also
make this reported problem (memory is getting released early) to go away.
Upstream fix in 5.6.22:
$ bzr log -r 6230 ------- ------- ------- ------- ------- ------- ------- ----
-------
revno: 6230
committer: Venkatesh Duggirala<email address hidden>
branch nick: mysql-5.6
timestamp: Thu 2014-10-30 22:33:43 +0530
message:
Bug#19704825 TEMPORARY SLAVE TYPE CONVERSION TABLES RELEASED TO EARLY
Problem: The memory used in preparing slave type conversion temporary table
is getting released early and causing unexpected results
Analysis: As part of bug#18770469 fix, We introduced an event
m_event_mem_root (a special mem root), added to Log_event
class. While server is creating the temporary table, the memory needed
is allocated from this special mem root which will be freed in ~Log_event()
i.e., the scope of this memroot is one event. But it could happen
that in some cases, server might need to access this
conversion temporary table for next following event.
For eg: A nested row event (insert is causing insert in a trigger)
In this situation, the memory is getting delayed too early
and causing issues when the server is trying to access the temporary
table inside the trigger.
Fix: We cannot use a mem_root whose scope is limited to an event
execution in this situation. With some further analysis, found out
that clearing a thd->mem_root at the end of statement (upon applying
an event which has STMT_END_F flag) will solve out of memory problem
while running a long transactions (bug#18770469) and will also
make this reported problem (memory is getting released early) to go away.