Detailed explanation of parallel replication examples in MySQL replication

Catalog

Traditional single thread replication instructions

summary

MySQL5.6 Parallel replication based on Library level

MySQL5.7 Parallel replication based on group commit

Group submission instructions

MySQL8.0 be based on writeset Parallel replication of

View key parameters

Parameter configuration item description

References ：

Traditional single thread replication instructions

as everyone knows ,MySQL stay 5.6 Before the release , There are two threads on the slave node of master-slave replication , Namely I/O Threads and SQL Threads .

I/O The thread is responsible for receiving binary logs Event write in Relay Log.SQL Thread reads Relay Log And play it back in the database .

The above methods occasionally cause delay , Then, what are the possible conditions that may cause the delay of the master and slave nodes ?

1. Main database performs large transactions （ Such as ： Large table structure change operation ）.2. The main database is changed in large quantities （ Such as ： A large number of insert 、 to update 、 Delete operation ）.3.ROW In synchronous mode , The primary database large table has no primary key and is frequently updated .4. The database parameter configuration is unreasonable , There is a bottleneck in the slave node performance （ Such as ： The setting from the node transaction log is too small , Cause frequent disk brushing ）.5. The network environment is unstable , From the node IO Thread reads binlog There is a delay 、 Reconnection .6. The master-slave hardware configuration is different , The hardware resource usage of the slave node has reached the upper limit .( such as ： Master node SSD disc , From the node SAS disc )

The above reasons for delay can be roughly classified .

1. Hardware problems （ Include disk IO、 The Internet IO etc. ）2. Configuration problems .3. Database design issues .4. The main database is changed in large quantities , From the node SQL Single thread processing is not timely . summary

By analyzing the above reasons, we can see that to reduce the master-slave delay , In addition to improving the hardware conditions , The other is the need DBA Focus on database design and configuration issues , Finally, it is necessary to improve the concurrent processing capability of the slave node , Changing from single thread playback to multi thread parallel playback is a better method , The key point is how to solve the problem of data conflict and confirmation of fault recovery location point under the premise of multi-threaded recovery .

MySQL5.6 Parallel replication based on Library level

In the case of multiple databases in the instance , Can open multiple threads , Each thread corresponds to a database . In this mode, the slave node will start multiple threads . Threads fall into two categories  Coordinator  and  WorkThread.

Thread division executes logic

Coordinator Threads are responsible for determining whether transactions can be executed in parallel , If it can be done in parallel, the transaction will be distributed to WorkThread Threads execute , If the judgment cannot be executed , Such as DDL, Across the library operation etc. , Just wait for all worker After thread execution , Again by Coordinator perform .

Key configuration information slave-parallel-type=DATABASE The scheme is insufficient

This parallel replication mode , Only in instances where there are multiple DB And DB Only when the transactions of are relatively busy will there be a high degree of parallelism , However, in daily maintenance, the transaction processing of a single instance is relatively concentrated in one DB On . By observing the delay, it can be found that the delay is mainly based on the hotspot table . It is a good way to provide table based parallelism .

MySQL5.7 Parallel replication based on group commit Group submission instructions

Simply put, it is in the double 1 Set up , After the transaction is committed, the disk flushing operation is changed to merge multiple transactions into a group of transactions, and then perform Unified Disk flushing , This process lowers the disk IO The pressure of the . Details reference Lao Ye teahouse Tweets about group submissions https://mp.weixin.qq.com/s/rcPkrutiLc93aTblEZ7sFg

A group of transactions are committed at the same time, which means that there is no conflict between transactions in the group , Therefore, the transactions in the group can be executed concurrently on the slave node , The problem is how to distinguish whether transactions are in the same group , So in binlog Two new parameter messages appear in last_committed  and  sequence_number

How to judge whether a transaction is in a group ?

analysis binlog You can see there's more in it last_committed and sequence_number Two parameter information , among last_committed There is duplication .

sequence_number # This value refers to the sequence number of the transaction commit , Monotone increasing . last_committed # This value has two meanings ,1. The same value means that these transactions are in the same group ,2. This value is also the maximum number representing the last set of transactions . [[email protected] GreatSQL]# mysqlbinlog mysql-bin.0000002 | grep last_committedGTID last_committed=0 sequence_number=1GTID last_committed=0 sequence_number=2GTID last_committed=2 sequence_number=3GTID last_committed=2 sequence_number=4GTID last_committed=2 sequence_number=5GTID last_committed=2 sequence_number=6GTID last_committed=6 sequence_number=7GTID last_committed=6 sequence_number=8 Database configuration slave-parallel-type=LOGICAL_CLOCK The scheme is insufficient

There is a deficiency in group based synchronization , That is, when the transaction busy degree of the primary node is low , Results in group submission in the time period fsync The amount of disk flushing transactions is less , As a result, the parallelism of playback from the library is not high , There may even be only one transaction in a group , In this way, the multithreading of slave nodes is basically unnecessary , You can set the following two parameters , Let the master node delay the submission .

binlog_group_commit_sync_delay # Time to wait for late submission ,binlog Wait for a while after submitting fsync. Let each group More things to do , Artificially improve the parallelism .binlog_group_commit_sync_no_delay_count # Maximum number of transactions to be committed , If the waiting time is not up , The number of transactions reached , Right now fsync. Commit as soon as the desired parallelism is reached , Try to minimize the waiting delay . MySQL8.0 be based on writeset Parallel replication of

writeset A method to judge whether a transaction can be played back in parallel based on the conflict of transaction results , He is made up of binlog-transaction-dependency-tracking Parameter control , By default WRITESET Method .

View key parameters Command-Line Format--binlog-transaction-dependency-tracking=value

System Variable	binlog_transaction_dependency_tracking
Scope	Global
Dynamic	Yes
SET_VAR Hint Applies	No
Type	Enumeration
Default Value	COMMIT_ORDER
Valid Values	COMMIT_ORDER WRITESET WRITESET_SESSION

Parameter configuration item description COMMIT_ORDER # Use 5.7 Group commit The way in which the transaction depends . WRITESET     # Determine transaction dependencies by writing collections . WRITESET_SESSION # Use write set , But the same session Transactions in cannot have the same last_committed.

writeset It's a HASH An array of types , It records the updated information of the transaction , adopt transaction_write_set_extraction Judge whether the record updated by the current transaction conflicts with the record updated by the historical transaction , After judgment, the corresponding treatment method shall be adopted .writeset The maximum stored value is determined by binlog-transaction-dependency-history-size control .

It should be noted that , When set to WRITESET or WRITESET_SESSION When , Transaction commit is out of order , Can be set by  slave_preserve_commit_order=1  Force sequential submission .

binlog_transaction_dependency_history_size

Set the maximum number of row hashes saved in memory , Used to cache row information modified by previous transactions . Once this hash number is reached , Will clear the history .

Command-Line Format--binlog-transaction-dependency-history-size=#

System Variable	binlog_transaction_dependency_history_size
Scope	Global
Dynamic	Yes
SET_VAR Hint Applies	No
Type	Integer
Default Value	25000
Minimum Value	1
Minimum Value	1000000

transaction_write_set_extraction

This mode supports three algorithms , By default XXHASH64, When the slave node is configured writeset When copying , This configuration cannot be configured as OFF. This parameter is already in MySQL 8.0.26 Abandoned in , It will be deleted later .

Command-Line Format--transaction-write-set-extraction[=value]

Deprecated	8.0.26
System Variable	binlog_transaction_dependency_history_size
Scope	Global, Session
Dynamic	Yes
SET_VAR Hint Applies	No
Type	Enumeration
Default Value	XXHASH64
Valid Values	OFF MURMUR32 XXHASH64

Database configuration

slave_parallel_type = LOGICAL_CLOCKslave_parallel_workers = 8binlog_transaction_dependency_tracking = WRITESETslave_preserve_commit_order = 1 References ： Ali kernel monthly ： http://mysql.taobao.org/monthly/2018/06/04/ Official documents ： https://dev.mysql.com/doc/refman/8.0/en/replication-options-binary-log.html

This is about MySQL Replication This is the end of the article on parallel replication , More about MySQL  Please search the previous articles of SDN or continue to browse the relevant articles below. I hope you will support SDN more in the future ！