~mysqlatfacebook/mysqlatfacebook/5.1 : changes

~mysqlatfacebook/mysqlatfacebook/5.1 » Changes from revision 3875

From Revision 3753 to 3734

Rev	Summary	Authors	Date
3753	Add my.cnf var reset_seconds_behind_master and SHOW STATUS v... Add my.cnf var reset_seconds_behind_master and SHOW STATUS value Relay_log_io_connected When TRUE reset_seconds_behind_master preserves original behavior which is to reset Seconds_Behind_Master to zero when the SQL thread catches up to the IO thread. When FALSE (default value) the reset is not done to avoid SBM flip-flops between zero and the real lag. The SHOW STATUS value Relay_log_io_connected counts the number of times the IO thread connects to the master. This might detect network problems from frequent disconnects.	Mark Callaghan	13 years ago
3752	Fix compiler warning for buf_uncache_tablespace Fix compiler warning for buf_uncache_tablespace	Mark Callaghan	13 years ago
3751	Handle admission_control_enter failures that occur after que... Handle admission_control_enter failures that occur after query start When admision_control_enter call fails after query start the count of running queries is not incremented for the user so it should not be decremented by any calls to admission_control_exit that follow. If admission_control_enter fails at query start, then the exit function is not called This is only an issue for failures done when enter/exit are called for disk and network IO. The AC state is changed in the THD to prevent a decrement from being done in admission_control_exit. The debug macros are moved before a mutex is locked for processing of sleep(). The call to admission_control_exit was also moved before the lock to avoid potential deadlock in the future. Changed admission_control.test to make it deterministic. Updated other tests because SHOW PROCESSLIST state is given now that the stall in sleep() is done earlier.	Mark Callaghan	13 years ago
3750	Add tests for DROP DATABASE with innodb background drop tabl... Add tests for DROP DATABASE with innodb background drop table	Mark Callaghan	13 years ago
3749	Don't assert that admission_control_diskio_enter returns 0. Don't assert that admission_control_diskio_enter returns 0. It can return !0 when the connection times out or is killed while in the queue	Mark Callaghan	13 years ago
3748	Fix test diffs for IS.table_statistics Fix test diffs for IS.table_statistics	Mark Callaghan	13 years ago
3747	Improve reporting for slow InnoDB IO requests Improve reporting for slow InnoDB IO requests Change the limit for "slow" requests from 2 to 1 second. Any read, write or fsync that takes at least that long in service time is counted and reported as an "old" request. Reporting this for fsync is new and looks like this in SHOW INNODB STATUS File flushes: 1216 requests, 12.37 seconds, 10.17 msecs/r, 0 old	Mark Callaghan	13 years ago
3746	Add table level compression statistics for primary index pag... Add table level compression statistics for primary index pages.	Mark Callaghan	13 years ago
3745	Fix http://bugs.mysql.com/62534 by making the my.cnf innodb_... Fix http://bugs.mysql.com/62534 by making the my.cnf innodb_max_dirty_pages_pct variable a double This makes innodb_max_dirty_pages_pct a double with min,default,max values 0.001, 75, 99.999. This also adds support for MYSQL_SYSVAR_DOUBLE. Added more to the BUFFER POOL AND MEMORY section of SHOW INNODB STATUS Percent pages dirty: 8.939 --> this is all n_dirty_pages / used_pages Percent all pages dirty: 0.195 --> this is n_dirty_pages / all-pages Max dirty pages percent: 75.000 --> this is innodb_max_dirty_pages_pct Also made all of the buf use 3 digits of precision instead of 2 (%.2f -> %.3f)	Mark Callaghan	13 years ago
3744	Reduce stalls from FLUSH TABLES foo WITH MEMORY CACHE Reduce stalls from FLUSH TABLES foo WITH MEMORY CACHE This changes buf_LRU_free_tablespace to unlock & relock the buffer pool mutex after every innodb_uncache_table_batch pages to reduce stalls from other threads waiting for it. This adds the my.cnf variable innodb_uncache_table_batch. The innodb team reviewed the change. These changes are made so that running FLUSH TABLES foo WITH MEMORY CACHE prior to DROP TABLE foo might make DROP TABLE cause fewer stalls. And we can experiment with that in production.	Mark Callaghan	13 years ago
3743	Update rows_deleted for truncate table. Update table stats c... Update rows_deleted for truncate table. Update table stats counters for blackhole and heap tables.	Mark Callaghan	13 years ago
3742	Upgrade version to 5.1.53 Upgrade version to 5.1.53 This is done so that we can use version specific comments to exclude grants for the new columns in mysql.user (max_concurrent_transactions, max_concurrent_queries). This is OK as we will skip the real 5.1.53 when upgrading the branch.	Mark Callaghan	13 years ago
3741	Backport fix for bugs.mysql.com/59338 from 5.1.56 Backport fix for bugs.mysql.com/59338 from 5.1.56 This makes mysql write INSERT IGNORE and INSERT .. ON DUPLICATE KEY UPDATE to the binlog even when no rows are changed to match what is done elsewhere and what is documented. I changed some tests from the original patch to avoid relying on so many new include files and to make the INSERT IGNORE test check that binlog events are written when no rows are inserted.	Mark Callaghan	13 years ago
3740	The read to cache the value of queries running should be as ... The read to cache the value of queries running should be as close as possible to the use. Over time code was added to admission_control_enter and the gap between them grew too much.	Mark Callaghan	13 years ago
3739	Reduce buffer pool mutex stalls during drop table. Reduce buffer pool mutex stalls during drop table. This is a workaround for http://bugs.mysql.com/51325. It adds the my.cnf variable innodb_drop_table_phase1 that determines whether buf_LRU_drop_page_hash_for_tablespace is called during drop table processing. It was always called in the past to reduce the overhead of removing pages from the adaptive hash index. But it increases the time for which the buffer pool mutex is locked. The call isn't required as buf_LRU_invalidate_tablespace repeats the work. This also adds a printf for monitoring DROP TABLE perf which eventually might be removed. It starts with "InnoDB: buf_LRU_invalidate_tablespace". This also adds two SHOW STATUS counters that measure the time for which the buffer pool mutex is locked during DROP TABLE: Innodb_drop_table_phase1_seconds - time in buf_LRU_drop_page_hash_for_tablespace Innodb_drop_table_phase2_seconds - time in buf_LRU_invalidate_tablespace after the call to buf_LRU_drop_page_hash_for_tablespace. On a server with a 60G buffer pool and 3.2M pages used the buffer pool mutex is locked for ~0.5 seconds (0.25 in buf_LRU_drop_page_hash_for_tablespace, and 0.25 in buf_LRU_invalidate_tablespace). So setting innodb_drop_table_phase1=OFF should cut the mutex lock time from 0.5 to 0.25 seconds because the first call is not made. The "optimization" of calling buf_LRU_drop_page_hash_for_tablespace attempts to reduce the overhead from removing pages from the hash index. It also removes pages from the hash index and it can be expensive for buf_LRU_invalidate_tablespace to remove a page from the hash index because to do that it must release the buffer pool mutex and then restart from the end of the LRU (a "rescan"). But, fseg_free_extent -> btr_search_drop_page_hash_when_freed earlier in the DROP TABLE processing removed all/most of the pages for this tablespace so the rescan should be unlikely and the new monitoring lets us confirm that.	Mark Callaghan	13 years ago
3738	This applies the patch for http://bugs.mysql.com/54538 This applies the patch for http://bugs.mysql.com/54538 Fix Bug #54538 - use of exclusive innodb dictionary lock limits performance. This patch doesn't get rid of the need to acquire the dict_sys->mutex but reduces the need to keep the mutex locked for the duration of the query to fsp_get_available_space_in_free_extents() from ha_innobase::info(). To get this diff, do this on the 5.1 bzr trunk for mysql: bzr log -p 3351.61.2	Mark Callaghan	13 years ago
3737	Make the default TRUE so that connections wait in the admiss... Make the default TRUE so that connections wait in the admission control queue after a disk read unless they have an active transaction. This reduces the chance that admission control limits will be exceeded	Mark Callaghan	13 years ago
3736	When TC is enabled do not enforce AC waits for connections t... When TC is enabled do not enforce AC waits for connections that have undo slots TC/AC == transaction/admission control. I don't want convoys to form when running transactions stall in the AC queue. This was requested in the review with db-ops. This moves the check from query start to all calls to admission_control_enter. It does not hurt sysbench QPS and it might help at high concurrency. Results below ac128 - max_concurrent_queries = max_concurrent_transactions = 128 tc0 - innodb_thread_concurrency=0 tc128 - innodb_thread_concurrency=128 Results with binary from 1 week ago: 8 16 32 64 128 256 512 1024 concurrent clients 15527 24117 36648 35742 34355 32654 28821 28508 ac128 15491 24066 35982 35655 34341 32219 26856 15424 tc0 15508 23977 35766 35270 34314 33724 32169 32080 tc128 Results with this change: 8 16 32 64 128 256 512 1024 concurrent clients 15530 23976 36539 35313 34094 32769 32324 31602 ac128 15422 24050 36407 35358 33833 32811 27292 15613 tc0 15446 23951 35687 35432 33920 33696 32251 31899 tc128	Mark Callaghan	13 years ago
3735	Remove the page_splits stats for innodb compression Remove the page_splits stats for innodb compression	Mark Callaghan	13 years ago
3734	Try to fix intermittent test failures Try to fix intermittent test failures	Mark Callaghan	13 years ago