作者:戴骏贤
网易游戏 技术部资深数据库系统工程师。
本文来源:原创投稿
*爱可生开源社区出品,原创内容未经授权不得随意使用,转载请联系小编并注明来源。
文章阅读时间约 15 分钟,文中著名参考文献在文末处有参考链接。
1. 问题背景
部门当前的数据库架构是 双主模式,既线上由 2 台互为主从的数据库搭建而成的集群。高可用通过 vip 和 headbeat 来做保证。通常情况下,vip 挂在主(本文称之为 da)上,当 da 发生了异常比如宕机等问题的时候,vip 自动漂移至从(本文称之为 dp)。架构如下图所示:
可以看出,所有的写入和读取操作都在 da 上进行。da 产生的 binlog 会被 dp 拉取重放。同时 dp 也会产生 binlog 并被 da 拉取重放(注:这里不会产生循环复制(既 da 产生的 binlog 事件在 dp 执行过了之后,再被 da 读取到再执行),因为 binlog 中每个 event 都有 serverid 标记,标记是哪个 server 产生的事物。当 da 读取到 binlog 的时候,发现 serverid 和本机的 serverid 一致,便会跳过当前的 binlog event)。正常情况下,da 的 Seconds_behind_Master 参数,应该保持在 0 或者很小的数值才对。但是监控系统发现,da 的 Seconds_behind_Master 参数经常发生跳变,上一秒为 0 下一秒就可能为一个很大值(比如 10000)。例如下面的数据,是通过脚本抓取到的线上 Seconds_behind_Master 情况:
2019-06-27 13:26:00
Read_Master_Log_Pos: 472851480
Relay_Log_File: da01-relay-bin.002323
Relay_Log_Pos: 371
Relay_Master_Log_File: dp01-bin.004267
Exec_Master_Log_Pos: 472851480
Seconds_Behind_Master: 0
2019-06-27 13:26:01
Read_Master_Log_Pos: 473594815
Relay_Log_File: da01-relay-bin.002323
Relay_Log_Pos: 371
Relay_Master_Log_File: dp01-bin.004267
Exec_Master_Log_Pos: 473594649
Seconds_Behind_Master: 622
2019-06-27 13:26:02
Read_Master_Log_Pos: 474422859
Relay_Log_File: da01-relay-bin.002323
Relay_Log_Pos: 371
Relay_Master_Log_File: dp01-bin.004267
Exec_Master_Log_Pos: 474422859
Seconds_Behind_Master: 0
2. 调查结论
前文讲到了本文调查的问题背景,在这里先给出下整个问题的调查结论。在 sql/rpl_slave.cc 中计算 Seconds_Behind_Master 逻辑如下【文献 2】
if ((mi->get_master_log_pos() == mi->rli->get_group_master_log_pos()) && (!strcmp(mi->get_master_log_name(), mi->rli->get_group_master_log_name()))) {
if (mi->slave_running == MYSQL_SLAVE_RUN_CONNECT)
protocol->store(0LL);
else
protocol->store_null();
} else {
long time_diff= ((long)(time(0) - mi->rli->last_master_timestamp) - mi->clock_diff_with_master);
protocol->store((longlong)(mi->rli->last_master_timestamp ? max(0L, time_diff) : 0));
}
当 Exec_Master_Log_Pos<Read_Master_Log_Pos 的时候进入 else 里面。在 da 的 IO 进程拉取 dp 的 binlog 的时候,会发现 dp 的 binlog 的 serverid 是自身产生的,所以不会去更新 last_master_timestamp。如果 dp 的 binlog 发生过轮转,会产生一个 ROTATE_EVENT 事件更新这个 last_master_timestamp。因此当 SQL 进程比 IO 进程慢的时候,计算出来的 time_diff 便是当前时间与上次 ROTATE_EVENT 事件的差值。复现方法:
- dp 上 flush logs ;
- da 上 stop slave sql_thread ;
- da 上 DML 数据。此时观察 show slave status; 可以看到 Read_Master_Log_Pos 很超前
- da 上 start slave ; show slave status G 可以看到 Seconds_Behind_Master 变大
3. 调查与分析过程
从前文的结果中可以看出,跟 Seconds_Behind_Master(下文简称 SBM)值有关的 MySQL 变量就是 last_master_timestamp。因此需要追踪下在收到的 binlog 的 serverid 和 MySQL 进程的 serverid 一致的情况下,last_master_timestamp 是如何更新的。根据前文的复现方法,需要去确认如下的 2 个点:
- dp 上 flush logs 的时候,da 获取到 dp 的 binglog 做的操作;
- 收到的 serverid 和本机一致的 binlog 的时候,da 的 MySQL 进程的操作。
为了便于观察 MySQL 的运行逻辑,本文参考【文献 9】的办法,编译了一个 debug 版本。在 MySQL 运行的过程中,tail -f mysqld.trace
文件观察 MySQL 命令执行的过程。
3.1 轮转 binlog 时的运行逻辑
3.1.1 写入 ROTATE_EVENT
在 dp 执行 flush logs 或者轮转 binlog 的时候,会产生一个 ROTATE_EVENT 事件(具体的解析可以参考【文献 11】)。在 mysqld.trace 文件中,可以观察到如下的信息。标志着 da 此时接收到了 dp 的 binlog 轮转事件的 ROTATE_EVENT。
handle_slave_io: info: IO thread received event of type Rotate
相应的代码在 sql/rpl_slave.cc 文件的 5819 行中:
5817 THD_STAGE_INFO(thd, stage_queueing_master_event_to_the_relay_log);
5818 event_buf= (const char*)mysql->net.read_pos + 1;
5819 DBUG_PRINT("info", ("IO thread received event of type %s",
5820 Log_event::get_type_str(
5821 (Log_event_type)event_buf[EVENT_TYPE_OFFSET])));
... ...
5831
5832 /* XXX: 'synced' should be updated by queue_event to indicate
5833 whether event has been synced to disk */
5834 bool synced= 0;
5835 if (queue_event(mi, event_buf, event_len))
5836 {
5837 mi->report(ERROR_LEVEL, ER_SLAVE_RELAY_LOG_WRITE_FAILURE,
5838 ER(ER_SLAVE_RELAY_LOG_WRITE_FAILURE),
5839 "could not queue event from master");
5840 goto err;
5841 }
可以看到在 SQL 进程接收到一个 binlog 之后,会调用 queue_event 进行写入 relaylog。queue_event 处理 ROTATE_EVENT 关键代码如下:
8236 switch (event_type) {
... ...
8251 case binary_log::ROTATE_EVENT:
8252 {
8253 Rotate_log_event rev(buf, checksum_alg != binary_log::BINLOG_CHECKSUM_ALG_OFF ?
8254 event_len - BINLOG_CHECKSUM_LEN : event_len,
8255 mi->get_mi_description_event());
8256
8257 if (unlikely(process_io_rotate(mi, &rev)))
8258 {
8259 mi->report(ERROR_LEVEL, ER_SLAVE_RELAY_LOG_WRITE_FAILURE,
8260 ER(ER_SLAVE_RELAY_LOG_WRITE_FAILURE),
8261 "could not queue event from master");
8262 goto err;
8263 }
在 switch 中判断 binlog 如果是 ROTATE_EVENT,则调用 process_io_rotate 进行处理。其中 buf 是 SQL 进行拉取到的 binlog 的缓存,mi 是代表了 masterinfo。process_io_rotate 主要的工作是更新 master_info 的相关信息,例如下一个 binglog 的位置,master 的 binlog 位置信息等。可以在 mysqld.trace 文件中观察到如下的信息:
process_io_rotate: info: new (master_log_name, master_log_pos): ('dp-bin.000029', 4)
process_io_rotate 函数注释信息如下:
7767 /**
7768 Used by the slave IO thread when it receives a rotate event from the
7769 master.
7770
7771 Updates the master info with the place in the next binary log where
7772 we should start reading. Rotate the relay log to avoid mixed-format
7773 relay logs.
7774
7775 @param mi master_info for the slave
7776 @param rev The rotate log event read from the master
7777
7778 @note The caller must hold mi->data_lock before invoking this function.
7779
7780 @retval 0 ok
7781 @retval 1 error
<!--7782 */-->
7783 static int process_io_rotate(Master_info *mi, Rotate_log_event *rev)
... ...
7830 }
在完成相关的工作之后,来到 queue_event 函数的最后,写入 ROTATE_EVENT,在 mysqld.trace 中可以看到如下的信息:
queue_event: info: master_log_pos: 4
harvest_bytes_written: info: counter: 769 bytes_written: 44
queue_event: info: error: 0
相关代码如下:
8656 {
8657 bool is_error= false;
8658 /* write the event to the relay log */
8659 if (likely(rli->relay_log.append_buffer(buf, event_len, mi) == 0))
8660 {
8661 mi->set_master_log_pos(mi->get_master_log_pos() + inc_pos);
8662 DBUG_PRINT("info", ("master_log_pos: %lu", (ulong) mi->get_master_log_pos()));
8663 rli->relay_log.harvest_bytes_written(rli, true/*need_log_space_lock=true*/);
... ...
}
}
到此,ROTATE_EVENT 事件被完全写入到了 relaylog 中。借用一张经典 MySQL 复制原理图【文献12】说明这个过程,即完成了图中从 BinaryLog 读取到 RelayLog 的写入过程。
3.1.2 SQL 进程读取和执行 ROTATE_EVENT
在前文的描述中,可以观察到在 RelayLog 的写入过程中并没有去更新 last_master_timestamp 相关信息。因此更新 last_master_timestamp 的工作,只可能在 SQL 进程中去完成。接下来本节将探索下ROTATE_EVENT在 SQL 进程中的执行过程。
MySQL 复制代码中,SQL 进程的入口函数是 handle_slave_sql:
7158 /**
7159 Slave SQL thread entry point.
7160
7161 @param arg Pointer to Relay_log_info object that holds information
7162 for the SQL thread.
7163
7164 @return Always 0.
7165 */
7166 extern "C" void *handle_slave_sql(void *arg)
在 handle_slave_sql 中,关键的代码是 while (!sql_slave_killed(thd,rli))
这个循环:
7438 while (!sql_slave_killed(thd,rli))
7439 {
7440 THD_STAGE_INFO(thd, stage_reading_event_from_the_relay_log);
7441 DBUG_ASSERT(rli->info_thd == thd);
7442 THD_CHECK_SENTRY(thd);
7443
7444 if (saved_skip && rli->slave_skip_counter == 0)
7445 {
... ...
7456 }
7457
7458 if (exec_relay_log_event(thd,rli))
7459 {
7460 ... ...
7525 }
7526 goto err;
7527 }
7528 }
可以发现,每次循环的时候,都会去执行 exec_relay_log_event 这个函数:
5098 Top-level function for executing the next event in the relay log.
5099 This is called from the SQL thread.
5100
5101 This function reads the event from the relay log, executes it, and
5102 advances the relay log position. It also handles errors, etc.
5103
5104 This function may fail to apply the event for the following reasons:
5105
5106 - The position specfied by the UNTIL condition of the START SLAVE
5107 command is reached.
5108
5109 - It was not possible to read the event from the log.
5110
5111 - The slave is killed.
5112
5113 - An error occurred when applying the event, and the event has been
5114 tried slave_trans_retries times. If the event has been retried
5115 fewer times, 0 is returned.
5116
5117 - init_info or init_relay_log_pos failed. (These are called
5118 if a failure occurs when applying the event.)
5119
5120 - An error occurred when updating the binlog position.
5121
5122 @retval 0 The event was applied.
5123
5124 @retval 1 The event was not applied.
5125 */
5126 static int exec_relay_log_event(THD* thd, Relay_log_info* rli)
在 exec_relay_log_event 中,执行对应的 relaylog 之后,会相应的更新last_master_timestamp。在这里并行复制模式和非并行复制模式下,更新last_master_timestamp的方式是不同的。在这里先介绍下非并行复制模式下更新last_master_timestamp的步骤。
3.1.2.1 非并行复制模式下更新 last_master_timestamp
在 exec_relay_log_event 中判断是否是并行复制是通过 is_parallel_exec 函数实现的。如果是并行复制模式则返回 True,否则返回 False。在非并行复制的模式下,会进入如下的代码执行:
5175 /*
5176 Even if we don't execute this event, we keep the master timestamp,
5177 so that seconds behind master shows correct delta (there are events
5178 that are not replayed, so we keep falling behind).
5179
5180 If it is an artificial event, or a relay log event (IO thread generated
5181 event) or ev->when is set to 0, or a FD from master, or a heartbeat
5182 event with server_id '0' then we don't update the last_master_timestamp.
5183
5184 In case of parallel execution last_master_timestamp is only updated when
5185 a job is taken out of GAQ. Thus when last_master_timestamp is 0 (which
5186 indicates that GAQ is empty, all slave workers are waiting for events from
5187 the Coordinator), we need to initialize it with a timestamp from the first
5188 event to be executed in parallel.
5189 */
5190 if ((!rli->is_parallel_exec() || rli->last_master_timestamp == 0) &&
5191 !(ev->is_artificial_event() || ev->is_relay_log_event() ||
5192 (ev->common_header->when.tv_sec == 0) ||
5193 ev->get_type_code() == binary_log::FORMAT_DESCRIPTION_EVENT ||
5194 ev->server_id == 0))
5195 {
5196 rli->last_master_timestamp= ev->common_header->when.tv_sec +
5197 (time_t) ev->exec_time;
5198 DBUG_ASSERT(rli->last_master_timestamp >= 0);
5199 }
可以看到,last_master_timestamp 的值便是 ev->common_header->when.tv_sec 和 (time_t) ev->exec_time 的和。在代码中加上点调试信息,重新编译:
DBUG_PRINT("info", ("before rli->last_master_timestamp = %lu", rli->last_master_timestamp));
DBUG_PRINT("info", ("before rli->is_parallel_exec() = %d", int(rli->is_parallel_exec())));
if ((!rli->is_parallel_exec() || rli->last_master_timestamp == 0) &&
!(ev->is_artificial_event() || ev->is_relay_log_event() ||
(ev->common_header->when.tv_sec == 0) ||
ev->get_type_code() == binary_log::FORMAT_DESCRIPTION_EVENT ||
ev->server_id == 0))
{
rli->last_master_timestamp= ev->common_header->when.tv_sec +
(time_t) ev->exec_time;
DBUG_PRINT("info", ("after rli->last_master_timestamp = %lu", rli->last_master_timestamp));
DBUG_PRINT("info", ("after rli->is_parallel_exec() = %d", int(rli->is_parallel_exec())));
DBUG_ASSERT(rli->last_master_timestamp >= 0);
}
可以在 mysqld.trace 中观察到如下的调试信息,印证了前文关于非并行复制的结论。同时也可以看出,非并行复制下 last_master_timestamp 的更新是在 binlogevent 执行之前。
exec_relay_log_event: info: before rli->last_master_timestamp = 0
exec_relay_log_event: info: before rli->is_parallel_exec() = 0
exec_relay_log_event: info: after rli->last_master_timestamp = 1562201989
exec_relay_log_event: info: after rli->is_parallel_exec() = 0
3.1.2.2 并行复制模式下更新 last_master_timestamp
前文介绍了非并行复制模式下的 last_master_timestamp 的更新逻辑。本节将介绍下在并行复制模式下,last_master_timestamp 是如何更新的。
在并行复制模式下 last_master_timestamp 的更新逻辑更为复杂。同时对于 ROTATE_EVENT 事件和普通的 binlog 事件更新模式也有所不同。接下来首先介绍下普通的 binlog 事件下 last_master_timestamp 的更新逻辑:
3.1.2.2.1 普通的 binlog 事件下 last_master_timestamp 的更新逻辑
在【文献1】和【文献2】中对于并行复制逻辑的分析如下:
并行复制有一个分发队列 gaq,SQL 线程将 binlog 事务读取到 gaq,然后再分发给 worker 线程执行。并行复制时,binlog 事件是并发穿插执行的,gaq 中有一个 checkpoint 点称为 lwm,lwm 之前的 binlog 都已经执行,而 lwm 之后的 binlog 有些执行有些没有执行。 假设 worker 线程数为 2,gap 有 1,2,3,4,5,6,7,8 个事务。worker 1 已执行的事务为 1 4 6, woker 2 执行的事务为 2 3 ,那么 lwm 为 4。并行复制更新 gap checkpiont 时,会推进 lwm 点,同时更新
last_master_timestamp
为 lwm 所在事务结束的 event 的时间。因此,并行复制是在事务执行完成后才更新last_master_timestamp
,更新是以事务为单位。同时更新 gap checkpiont 还受slave_checkpoint_period
参数的影响。coordinator 会调用函数 mts_checkpoint_routune 定期做“checkpoint”操作,将 Relay_log_info.gaq 中已经执行结束的事务移出队列,移除方式为从队列头开始检查,如果 done 为真则移出,否则停止扫描,并标记 Low Water Mark 为移出的事务中最大的 sequence_number(lock-interval 终止点)
可以看出,并行复制下 last_master_timestamp 的更新逻辑与 lwm 和 mts_checkpoint_routune 函数有关。mts_checkpoint_routune 函数头如下:
/**
Processing rli->gaq to find out the low-water-mark (lwm) coordinates
which is stored into the cental recovery table.
@param rli pointer to Relay-log-info of Coordinator
@param period period of processing GAQ, normally derived from
@c mts_checkpoint_period
@param force if TRUE then hang in a loop till some progress
@param need_data_lock False if rli->data_lock mutex is aquired by
the caller.
@return FALSE success, TRUE otherwise
*/
bool mts_checkpoint_routine(Relay_log_info *rli, ulonglong period,
bool force, bool need_data_lock)
在 coordinator 中会定期调用 mts_checkpoint_routine 将 Relay_log_info.gaq 中已经执行结束的事务移出队列,并且更新 Low Water Mark 为移出的事务中最大的 sequence_number(参见 mts_checkpoint_routine 中如下):
do
{
if (!is_mts_db_partitioned(rli))
mysql_mutex_lock(&rli->mts_gaq_LOCK);
cnt= rli->gaq->move_queue_head(&rli->workers);
if (!is_mts_db_partitioned(rli))
mysql_mutex_unlock(&rli->mts_gaq_LOCK);
#ifndef DBUG_OFF
if (DBUG_EVALUATE_IF("check_slave_debug_group", 1, 0) &&
cnt != opt_mts_checkpoint_period)
sql_print_error("This an error cnt != mts_checkpoint_period");
#endif
} while (!sql_slave_killed(rli->info_thd, rli) &&
cnt == 0 && force &&
!DBUG_EVALUATE_IF("check_slave_debug_group", 1, 0) &&
(my_sleep(rli->mts_coordinator_basic_nap), 1));
在函数的最后,会执行 last_master_timestamp 的更新逻辑:
/*
Update the rli->last_master_timestamp for reporting correct Seconds_behind_master.
If GAQ is empty, set it to zero.
Else, update it with the timestamp of the first job of the Slave_job_queue
which was assigned in the Log_event::get_slave_worker() function.
*/
ts= rli->gaq->empty()
? 0
: reinterpret_cast<Slave_job_group*>(rli->gaq->head_queue())->ts;
rli->reset_notified_checkpoint(cnt, ts, need_data_lock, true);
可以看出如果 GAP 为空(既 work 的队列为空)则更新 last_master_timestamp 为 0,否则更新为队列中第一个是元素的 timestamp。
3.1.2.2.2 ROTATE_EVENT 事件下 last_master_timestamp 的更新模式
前文提到在并行复制模式下,普通的 binlog 事件会被 coordinator 进程分发到 work 中去执行。并且在 mts_checkpoint_routine 中去推进 Low Water Mark 和更新 last_master_timestamp。但是如果执行的 binlog 是一个 ROTATE_EVENT 事件,则 coordinator 进程不会将事件分发到 work 中,而是在 coordinator 进程中自己执行并更新 last_master_timestamp。
对于 ROTATE_EVENT 事件的执行逻辑,入口依然是 exec_relay_log_event 函数。在函数中调用了 apply_event_and_update_pos 函数执行 binlog 事件。
5247 /* ptr_ev can change to NULL indicating MTS coorinator passed to a Worker */
5248 exec_res= apply_event_and_update_pos(ptr_ev, thd, rli);
5249 /*
5250 Note: the above call to apply_event_and_update_pos executes
5251 mysql_mutex_unlock(&rli->data_lock);
5252 */
5253
5254 /* For deferred events, the ptr_ev is set to NULL
5255 in Deferred_log_events::add() function.
5256 Hence deferred events wont be deleted here.
5257 They will be deleted in Deferred_log_events::rewind() funciton.
5258 */
在 apply_event_and_update_pos 函数中,会首先调用 apply_event 去判断当前的 binlog 事件是否可以被分发到 work 进程执行。在这里收到的 binlog 事件是 ROTATE_EVENT,因此 apply_event 返回的 exec_res 为 0,ev->worker==rli(rli 代表了 coordinator 进程),因此不会进入下面的 if 逻辑中:
4709 if (reason == Log_event::EVENT_SKIP_NOT)
4710 {
4711 // Sleeps if needed, and unlocks rli->data_lock.
4712 if (sql_delay_event(ev, thd, rli))
4713 DBUG_RETURN(SLAVE_APPLY_EVENT_AND_UPDATE_POS_OK);
4714
4715 exec_res= ev->apply_event(rli);
4717 if (!exec_res && (ev->worker != rli))
4718 {
... ...
}
在 mysqld.trace 中可以观察到如下的输出:
Log_event::shall_skip: info: skip reason=0=NOT
LOG_EVENT:apply_event: info: event_type=Rotate
apply_event_and_update_pos: info: apply_event error = 0
apply_event 函数定义和注释如下:
/**
Scheduling event to execute in parallel or execute it directly.
In MTS case the event gets associated with either Coordinator or a
Worker. A special case of the association is NULL when the Worker
can't be decided yet. In the single threaded sequential mode the
event maps to SQL thread rli.
@note in case of MTS failure Coordinator destroys all gathered
deferred events.
@return 0 as success, otherwise a failure.
*/
int Log_event::apply_event(Relay_log_info *rli)
因为无法被分发到work执行,因此会进入如下的执行逻辑中:
4832 DBUG_PRINT("info", ("apply_event error = %d", exec_res));
4833 if (exec_res == 0)
4834 {
4835 /*
4836 Positions are not updated here when an XID is processed. To make
4837 a slave crash-safe, positions must be updated while processing a
4838 XID event and as such do not need to be updated here again.
4839
4840 However, if the event needs to be skipped, this means that it
4841 will not be processed and then positions need to be updated here.
4842
4843 See sql/rpl_rli.h for further details.
4844 */
4845 int error= 0;
4846 if (*ptr_ev &&
4847 (ev->get_type_code() != binary_log::XID_EVENT ||
4848 skip_event || (rli->is_mts_recovery() && !is_gtid_event(ev) &&
4849 (ev->ends_group() || !rli->mts_recovery_group_seen_begin) &&
4850 bitmap_is_set(&rli->recovery_groups, rli->mts_recovery_index))))
4851 {
4852 #ifndef DBUG_OFF
4853 /*
4854 This only prints information to the debug trace.
4855
4856 TODO: Print an informational message to the error log?
4857 */
... ...
4873 error= ev->update_pos(rli);
... ...
}
}
可以发现,在此调用了 update_pos 进行更新。因为当前的 binlog 事件是 ROTATE_EVENT,因此会调用如下的代码:
/*
Got a rotate log event from the master.
This is mainly used so that we can later figure out the logname and
position for the master.
We can't rotate the slave's BINlog as this will cause infinitive rotations
in a A -> B -> A setup.
The NOTES below is a wrong comment which will disappear when 4.1 is merged.
This must only be called from the Slave SQL thread, since it calls
flush_relay_log_info().
@retval
0 ok
*/
int Rotate_log_event::do_update_pos(Relay_log_info *rli)
在 do_update_pos 更新 last_master_timestamp 的关键逻辑如下:
if (rli->is_parallel_exec())
{
bool real_event= server_id && !is_artificial_event();
rli->reset_notified_checkpoint(0,
real_event ?
common_header->when.tv_sec +
(time_t) exec_time : 0,
true/*need_data_lock=true*/,
real_event? true : false);
}
可以在 mysqld.trace 中观察到如下的输出:
Rotate_log_event::do_update_pos: info: server_id=248; ::server_id=236
Rotate_log_event::do_update_pos: info: new_log_ident: dp-bin.000063
Rotate_log_event::do_update_pos: info: pos: 4
Rotate_log_event::do_update_pos: info: old group_master_log_name: 'dp-bin.000062' old group_master_log_pos: 154
Relay_log_info::inc_group_relay_log_pos: info: log_pos: 4 group_master_log_pos: 154
Rotate_log_event::do_update_pos: info: new group_master_log_name: 'dp-bin.000063' new group_master_log_pos: 4
Rotate_log_event::do_update_pos: info: ------> 1562684581(此处为本文加入的调试信息,在reset_notified_checkpoint更新last_master_timestamp的时候打印出。)
至此,对于并行复制和非并行复制下当收到 ROTATE_EVENT 事件时,更新 last_master_timestamp 的逻辑分析全部完成。
3.2 da 执行与本实例 serverid 一致的 binlog 时的运行逻辑
前文分析了当 MySQL 收到一个 ROTATE_EVENT 事件的时候所运行的逻辑。本节将分析下在 MySQL 收到与自己的 serverid 一致的 binlog 事件的时候的运行逻辑。
3.2.1 io thread 处理与本实例 serverid 一致的 binlog
首先来看下 io thread 接收到与本身 serverid 一致的 binlog 的时候所做的操作。在 handle_slave_io 函数中,会在 while 循环中不断的调用 queue_even 函数。
5740 while (!io_slave_killed(thd,mi))
5741 {
5742 ulong event_len;
... ...
5750 event_len= read_event(mysql, mi, &suppress_warnings);
... ...
5813 /* XXX: 'synced' should be updated by queue_event to indicate
5814 whether event has been synced to disk */
5815 bool synced= 0;
5816 if (queue_event(mi, event_buf, event_len))
5817 {
5818 mi->report(ERROR_LEVEL, ER_SLAVE_RELAY_LOG_WRITE_FAILURE,
5819 ER(ER_SLAVE_RELAY_LOG_WRITE_FAILURE),
5820 "could not queue event from master");
5821 goto err;
5822 }
... ...
}
在 queue_even 函数中,对于收到与自己 serverid 一致的 binlog 的处理逻辑如下:
8534 /*
8535 If this event is originating from this server, don't queue it.
8536 We don't check this for 3.23 events because it's simpler like this; 3.23
8537 will be filtered anyway by the SQL slave thread which also tests the
8538 server id (we must also keep this test in the SQL thread, in case somebody
8539 upgrades a 4.0 slave which has a not-filtered relay log).
8540
8541 ANY event coming from ourselves can be ignored: it is obvious for queries;
8542 for STOP_EVENT/ROTATE_EVENT/START_EVENT: these cannot come from ourselves
8543 (--log-slave-updates would not log that) unless this slave is also its
8544 direct master (an unsupported, useless setup!).
8545 */
... ...
8560 if ((s_id == ::server_id && !mi->rli->replicate_same_server_id) ||
8561 /*
8562 the following conjunction deals with IGNORE_SERVER_IDS, if set
8563 If the master is on the ignore list, execution of
8564 format description log events and rotate events is necessary.
8565 */
8566 (mi->ignore_server_ids->dynamic_ids.size() > 0 &&
8567 mi->shall_ignore_server_id(s_id) &&
8568 /* everything is filtered out from non-master */
8569 (s_id != mi->master_id ||
8570 /* for the master meta information is necessary */
8571 (event_type != binary_log::FORMAT_DESCRIPTION_EVENT &&
8572 event_type != binary_log::ROTATE_EVENT))))
8573 {
8574 /*
8575 Do not write it to the relay log.
8576 a) We still want to increment mi->get_master_log_pos(), so that we won't
8577 re-read this event from the master if the slave IO thread is now
8578 stopped/restarted (more efficient if the events we are ignoring are big
8579 LOAD DATA INFILE).
8580 b) We want to record that we are skipping events, for the information of
8581 the slave SQL thread, otherwise that thread may let
8582 rli->group_relay_log_pos stay too small if the last binlog's event is
8583 ignored.
8584 But events which were generated by this slave and which do not exist in
8585 the master's binlog (i.e. Format_desc, Rotate & Stop) should not increment
8586 mi->get_master_log_pos().
8587 If the event is originated remotely and is being filtered out by
8588 IGNORE_SERVER_IDS it increments mi->get_master_log_pos()
8589 as well as rli->group_relay_log_pos.
8590 */
8591 if (!(s_id == ::server_id && !mi->rli->replicate_same_server_id) ||
8592 (event_type != binary_log::FORMAT_DESCRIPTION_EVENT &&
8593 event_type != binary_log::ROTATE_EVENT &&
8594 event_type != binary_log::STOP_EVENT))
8595 {
8596 mi->set_master_log_pos(mi->get_master_log_pos() + inc_pos);
8597 memcpy(rli->ign_master_log_name_end, mi->get_master_log_name(), FN_REFLEN);
8598 DBUG_ASSERT(rli->ign_master_log_name_end[0]);
8599 rli->ign_master_log_pos_end= mi->get_master_log_pos();
8600 }
8601 rli->relay_log.signal_update(); // the slave SQL thread needs to re-check
8602 DBUG_PRINT("info", ("master_log_pos: %lu, event originating from %u server, ignored",
8603 (ulong) mi->get_master_log_pos(), uint4korr(buf + SERVER_ID_OFFSET)));
8604 }
结合注释,可以看出当 MySQL 收到与本实例 serverid 一致的 binlog 的时候,不会将当前的 binlog 事件写入 relaylog 中。同时会完成如下的 2 个事情:(ps :这是一个关键的地方,在下文中将提到它的作用)
- 将 mi->get_master_log_name() 拷贝到 rli->ign_master_log_name_end 中。
- rli->relay_log.signal_update()。此处会新 signal_cnt 这个变量的值。
在 mysqld.trace 中可以观察到如下的日志信息:
queue_event: info: master_log_pos: 219, event originating from 236 server, ignored
queue_event: info: error: 0
3.2.2 sql thread 处理与本实例 serverid 一致的 binlog
接下里看看 sql 进程对 binlog 的处理过程。入口的函数还是 exec_relay_log_event,在 exec_relay_log_event 函数中会调用next_event函数获取一个可执行的 binlog 事件。在这里 next_event 对于非并行复制会有一个特殊的处理:
9175 if (!rli->is_parallel_exec())
9176 rli->last_master_timestamp= 0;
如果是非并行复制,则当读取一个 binlog 的时候,都会把 last_master_timestamp 设置成 0 。因此在非并行复制下,收到与本实例 serverid 一致的 binlog 的时候,mysqld.trace 中可以观察到 rli->last_master_timestamp 的值会一直为 0:
next_event: info: seeing an ignored end segment
handle_slave_io: info: IO thread received event of type Query
exec_relay_log_event: info: ================================before rli->last_master_timestamp = 0
exec_relay_log_event: info: ================================before rli->is_parallel_exec() = 0
apply_event_and_update_pos: info: thd->options: ; rli->last_event_start_time: 0
Log_event::shall_skip: info: ev->server_id=0, ::server_id=236, rli->replicate_same_server_id=0, rli->slave_skip_counter=0
Log_event::shall_skip: info: skip reason=0=NOT
LOG_EVENT:apply_event: info: event_type=Rotate
... ...
... ...
exec_relay_log_event: info: ================================before rli->last_master_timestamp = 0
exec_relay_log_event: info: ================================before rli->is_parallel_exec() = 0
apply_event_and_update_pos: info: thd->options: ; rli->last_event_start_time: 0
Log_event::shall_skip: info: ev->server_id=0, ::server_id=236, rli->replicate_same_server_id=0, rli->slave_skip_counter=0
Log_event::shall_skip: info: skip reason=0=NOT
LOG_EVENT:apply_event: info: event_type=Rotate
apply_event_and_update_pos: info: apply_event error = 0
apply_event_and_update_pos: info: OPTION_BEGIN: 0; IN_STMT: 0
Rotate_log_event::do_update_pos: info: server_id=0; ::server_id=236
Rotate_log_event::do_update_pos: info: new_log_ident: dp-bin.000069
Rotate_log_event::do_update_pos: info: pos: 282
MYSQL_BIN_LOG::signal_update: info: signal_cnt : 13
queue_event: info: master_log_pos: 326, event originating from 236 server, ignored
queue_event: info: error: 0
之后的逻辑中,如果发现 rli->ign_master_log_name_end[0] 不为空(对应了上一个小节中的第一项:将 mi->get_master_log_name() 拷贝到 rli->ign_master_log_name_end 中),则构造一个 serverid 为 0 的 Rotate event 并返回:
9181 if (rli->ign_master_log_name_end[0])
9182 {
9183 /* We generate and return a Rotate, to make our positions advance */
9184 DBUG_PRINT("info",("seeing an ignored end segment"));
9185 ev= new Rotate_log_event(rli->ign_master_log_name_end,
9186 0, rli->ign_master_log_pos_end,
9187 Rotate_log_event::DUP_NAME);
9188 rli->ign_master_log_name_end[0]= 0;
9189 mysql_mutex_unlock(log_lock);
9190 if (unlikely(!ev))
9191 {
9192 errmsg= "Slave SQL thread failed to create a Rotate event "
9193 "(out of memory?), SHOW SLAVE STATUS may be inaccurate";
9194 goto err;
9195 }
9196 ev->server_id= 0; // don't be ignored by slave SQL thread
9197 DBUG_RETURN(ev);
9198 }
紧接着调用 apply_event_and_update_pos函数。在apply_event_and_update_pos 函数中,如前所述如果是 ROTATE EVENT 则返回 0,不会被 work 进程并行执行,并进入 update_pos 逻辑中。在 update_pos 逻辑中关键的代码如下(sql/log_event.cc 文件):
6649 if (rli->is_parallel_exec())
6650 {
6651 bool real_event= server_id && !is_artificial_event();
6652 rli->reset_notified_checkpoint(0,
6653 real_event ?
6654 common_header->when.tv_sec +
6655 (time_t) exec_time : 0,
6656 true/*need_data_lock=true*/,
6657 real_event? true : false);
6658 }
可以见得 real_event 在 server_id 是 0 的时候为 false。因此当进入 reset_notified_checkpoint 函数后,因为 update_timestamp 条件(传入的值便是 real_event 的值)为 false 便不会更新 last_master_timestamp。
262 /**
263 This method is called in mts_checkpoint_routine() to mark that each
264 worker is required to adapt to a new checkpoint data whose coordinates
265 are passed to it through GAQ index.
266
267 Worker notices the new checkpoint value at the group commit to reset
268 the current bitmap and starts using the clean bitmap indexed from zero
269 of being reset checkpoint_seqno.
270
271 New seconds_behind_master timestamp is installed.
272
273 @param shift number of bits to shift by Worker due to the
274 current checkpoint change.
275 @param new_ts new seconds_behind_master timestamp value
276 unless zero. Zero could be due to FD event
277 or fake rotate event.
278 @param need_data_lock False if caller has locked @c data_lock
279 @param update_timestamp if true, this function will update the
280 rli->last_master_timestamp.
281 */
282 void Relay_log_info::reset_notified_checkpoint(ulong shift, time_t new_ts,
283 bool need_data_lock,
284 bool update_timestamp)
285 {
... ...
330 if (update_timestamp)
331 {
332 if (need_data_lock)
333 mysql_mutex_lock(&data_lock);
334 else
335 mysql_mutex_assert_owner(&data_lock);
336 last_master_timestamp= new_ts;
337 if (need_data_lock)
338 mysql_mutex_unlock(&data_lock);
339 }
340 }
3.3 小结
从上文的分析可以得出如下的 2 个结论 :
- 并行复制下会发现 last_master_timestamp 的值,一直是 ROTATE EVENT 发生的时间。当 Exec_Master_Log_Pos < Read_Master_Log_Pos 的时候,便会发现 Seconds_Behind_Master 的值突变(突变的值是当前时间和 ROTATE EVENT 发生的时间差的秒数)
- 在非并行复制下,因为 rli->last_master_timestamp 会在 next_event 被置为 0。同时之后的代码也不会去更新这个参数的值,因此 rli->last_master_timestamp 一直为 0 。show slave status 的时候,当 Exec_Master_Log_Pos < Read_Master_Log_Pos 的时候,Seconds_Behind_Master 的值不会突变。
4. 总结
本文分析了并行复制和非并行复制下,Seconds_Behind_Master 参数值更新的相关逻辑。限于本文的作者水平有限,文中的错误在所难免,恳请大家批评指正。
5. 参考文献
ps:如上的参考资料并非完全在本文中被引用,其他没有被引用的资料在本文的形成的工作中,提供了思路上和其他方面的参考,因此一并列入参考文献。感谢如上的文献作者提供的参考。
**粗体** _斜体_ [链接](http://example.com) `代码` - 列表 > 引用
。你还可以使用@
来通知其他用户。