关于删除MySQL Logs的一点记录
今天尝试着重现了此问题,这里简要分析下原因。
重现步骤:
1. 构造slow log (将long_query_time设成了0);
2. 观察删rm slow log瞬间,tps/qps变化;
3. 观察执行flush slow logs瞬间,tps/qps变化;
4. 记录flush slow logs执行时, pstack打出的调用栈情况;
第一步,没啥好说的。
第二步,tps/qps没啥变化。
第三步,会发现tps/qps瞬间跌0,如下所示:
[ 639s] threads: 32, tps: 1121.00, reads/s: 15843.98, writes/s: 4490.99[ 640s] threads: 32, tps: 792.99, reads/s: 10803.89, writes/s: 3150.97[ 641s] threads: 32, tps: 0.00, reads/s: 0.00, writes/s: 0.00[ 642s] threads: 32, tps: 0.00, reads/s: 0.00, writes/s: 0.00[ 643s] threads: 32, tps: 471.01, reads/s: 6860.08, writes/s: 1908.02
mysql命令行会发现,flush slow logs执行时间刚好为3s左右。
第四步,我们看下pstack的输出结果,只记录相关的:
610 Thread 5 (Thread 0x2afdc4101700 (LWP 30762)):611 #00x0000003c6e40a7d6 in pthread_rwlock_rdlock () from /lib64/libpthread.so.0612 #10x0000000000825135 in inline_mysql_rwlock_rdlock ()613 #20x0000000000838004 in LOGGER::lock_shared() ()614 #30x00000000008283bf in LOGGER::slow_log_print(THD*, char const*, unsigned int, unsigned long long) ()615 #40x0000000000832b30 in slow_log_print(THD*, char const*, unsigned int, unsigned long long) ()616 #50x0000000000609f23 in log_slow_statement(THD*) ()617 #60x00000000006099d1 in dispatch_command(enum_server_command, THD*, char*, unsigned int) ()618 #70x0000000000606e02 in do_command(THD*) ()619 #80x00000000006f070f in do_handle_one_connection(THD*) ()620 #90x00000000006f020d in handle_one_connection ()621 #10 0x0000003c6e4077f1 in start_thread () from /lib64/libpthread.so.0622 #11 0x0000003c6e0e570d in clone () from /lib64/libc.so.6623 Thread 4 (Thread 0x2afdd0080700 (LWP 30763)):624 #00x0000003c6e40a7d6 in pthread_rwlock_rdlock () from /lib64/libpthread.so.0625 #10x0000000000825135 in inline_mysql_rwlock_rdlock ()626 #20x0000000000838004 in LOGGER::lock_shared() ()627 #30x00000000008283bf in LOGGER::slow_log_print(THD*, char const*, unsigned int, unsigned long long) ()628 #40x0000000000832b30 in slow_log_print(THD*, char const*, unsigned int, unsigned long long) ()629 #50x0000000000609f23 in log_slow_statement(THD*) ()630 #60x00000000006099d1 in dispatch_command(enum_server_command, THD*, char*, unsigned int) ()631 #70x0000000000606e02 in do_command(THD*) ()632 #80x00000000006f070f in do_handle_one_connection(THD*) ()633 #90x00000000006f020d in handle_one_connection ()634 #10 0x0000003c6e4077f1 in start_thread () from /lib64/libpthread.so.0635 #11 0x0000003c6e0e570d in clone () from /lib64/libc.so.6636 Thread 3 (Thread 0x2afdd0101700 (LWP 30764)):637 #00x0000003c6e40a7d6 in pthread_rwlock_rdlock () from /lib64/libpthread.so.0638 #10x0000000000825135 in inline_mysql_rwlock_rdlock ()639 #20x0000000000838004 in LOGGER::lock_shared() ()640 #30x00000000008283bf in LOGGER::slow_log_print(THD*, char const*, unsigned int, unsigned long long) ()641 #40x0000000000832b30 in slow_log_print(THD*, char const*, unsigned int, unsigned long long) ()642 #50x0000000000609f23 in log_slow_statement(THD*) ()643 #60x00000000006099d1 in dispatch_command(enum_server_command, THD*, char*, unsigned int) ()644 #70x0000000000606e02 in do_command(THD*) ()645 #80x00000000006f070f in do_handle_one_connection(THD*) ()646 #90x00000000006f020d in handle_one_connection ()647 #10 0x0000003c6e4077f1 in start_thread () from /lib64/libpthread.so.0648 #11 0x0000003c6e0e570d in clone () from /lib64/libc.so.6649 Thread 2 (Thread 0x2afe18080700 (LWP 30855)):650 #00x0000003c6e40e54d in close () from /lib64/libpthread.so.0651 #10x00000000008f56ed in my_close ()652 #20x0000000000825c16 in inline_mysql_file_close ()653 #30x000000000082b305 in MYSQL_LOG::close(unsigned int) ()654 #40x000000000082b634 in MYSQL_QUERY_LOG::reopen_file() ()655 #50x0000000000828283 in LOGGER::flush_slow_log() ()656 #60x000000000071d8fc in reload_acl_and_cache(THD*, unsigned long, TABLE_LIST*, int*) ()657 #70x0000000000610200 in mysql_execute_command(THD*) ()658 #80x000000000061534d in mysql_parse(THD*, char*, unsigned int, Parser_state*) ()659 #90x00000000006086a0 in dispatch_command(enum_server_command, THD*, char*, unsigned int) ()660 #10 0x0000000000606e02 in do_command(THD*) ()661 #11 0x00000000006f070f in do_handle_one_connection(THD*) ()662 #12 0x00000000006f020d in handle_one_connection ()663 #13 0x0000003c6e4077f1 in start_thread () from /lib64/libpthread.so.0664 #14 0x0000003c6e0e570d in clone () from /lib64/libc.so.6
会发现Thread 2在执行flush slow logs操作,其他的线程都在等待锁LOCK_log上边。
背后的原因其实很简单,在shell中执行rm slow log操作时,由于mysqld仍有文件句柄打开此文件,所以实际上此时文件并未删除。执行flush logs操作,其实际执行的是1)close;2)open 操作(logger.flush_slow_log -> mysql_slow_log.reopen_file),在close操作执行时,文件系统真正删除文件,此时该线程占用着LOCK_log锁。
删除时会执行刷脏(当然我构造这个场景很极端,基本所有slow log文件的内容都在文件系统缓存中),这个会很耗时间,比如我执行这个语句耗了3s。此时间段内,如果连接发来的语句需要记log(server层的log:slow log/binlog/general log共有LOCK_log这把锁)就会处于等待状态,那么系统对外的反应就是hang住了。
flush slow logs中调用执行的close所需时间和文件大小、以及文件系统缓存中该文件脏页比例都有关系,比如我在执行flush slow logs前使用sysctl vm.drop_caches=3清空
了文件系统缓存的话,同样大小的flush slow logs操作执行时间是0.42s,相应的阻塞时间也会减少不少。
可以考虑在slow logs的文件句柄上执行posix_fadvise调用,促使不会缓存很大的log文件内容(slow log也没啥需要缓存的),这有篇霸爷的文章,可以参考下 posix_fadvise清除缓存的误解和改进措施 。
另外,peter在07年就讨论过这个问题, Be careful rotating MySQL logs 其给出的建议是先mv file,然后flush logs,再执行删除文件的操作,让真正的删除行为由自己而不是mysqld完成。比较遗憾的是,五年过去了,LOCK_log这把锁的问题还没有完整的解决掉。
PS:
文章结尾记一点备忘,通过close/rm操作删除一个10G大小的文件,在执行sysctl vm.drop_caches=3清空缓存后,此操作的耗时仍在百毫秒量级(我的机器上是200ms+),其背后做了什么事情还需要找内核组的同事了解下。
上一篇: c# 自定义异常类
下一篇: ajax.call是什么意思?解决方案