Updated readwritesplit documentation.

This commit is contained in:
Markus Makela
2015-05-12 04:57:18 +03:00
parent 8f83845f34
commit 943c9f5f65
2 changed files with 17 additions and 24 deletions

View File

@ -37,9 +37,8 @@ In master-slave replication cluster also read-only queries are routed to master
* statement includes a stored procedure, or an UDF call
### Limitations in client session handling
Some of the queries that client sends are routed to all backends instead of sending them just to one of server. These queries include "USE <db name>" and “SET autocommit=0” among many others. Read/Write Splitter sends a copy of these queries to each backend server and forwards the first reply it receives to the client. Below is a list of MySQL commands which we call session commands :
Some of the queries that client sends are routed to all backends instead of sending them just to one of server. These queries include `USE <db name>` and `SET autocommit=0` among many others. Readwritesplit sends a copy of these queries to each backend server and forwards the master's reply to the client. Below is a list of MySQL commands which are classified as session commands :
```
COM_INIT_DB (USE <db name> creates this)
COM_CHANGE_USER
@ -52,9 +51,7 @@ COM_STMT_RESET
COM_STMT_PREPARE
Also these are session commands:
COM_QUIT (no response)
COM_QUIT (no response, session is closed)
COM_REFRESH
@ -62,9 +59,7 @@ COM_DEBUG
COM_PING
In addition there are query types which belong to the same group:
SQLCOM_CHANGE_DB
SQLCOM_CHANGE_DB (USE ... statements)
SQLCOM_DEALLOCATE_PREPARE
@ -74,25 +69,24 @@ SQLCOM_SET_OPTION
SELECT ..INTO variable|OUTFILE|DUMPFILE
Then there are queries which modify session characteristics, listed as derived, internal RWSplit types:
SET autocommit=1|0
```
QUERY_TYPE_ENABLE_AUTOCOMMIT
There is a possibility for misbehavior; if `USE mytable` was executed in one of the slaves and it failed, it may be due to replication lag rather than the fact it didn’t exist. Thus the same command may end up with different result among backend servers. The slaves which fail to execute a session command will be dropped from the active list of slaves for this session to guarantee a consistent session state across all the servers that are in use by the session.
QUERY_TYPE_DISABLE_AUTOCOMMIT
The above-mentioned behavior can be partially controller with the `use_sql_variables_in` configuration parameter.
There is a possibility for misbehavior; if "USE mytable" was executed in one of the slaves and it failed, it may be due to replication lag rather than the fact it didn’t exist. Thus the same command may end up with different result among backend servers. This disparity is missed.
The above-mentioned behavior can be partially controller with RWSplit configuration parameter called
use_sql_variables_in=[master|all] (master)
```
use_sql_variables_in=[master|all] (master)
```
Server-side session variables are called as SQL variables. If "master" or no value is set, SQL variables are read and written in master only. Autocommit values and prepared statements are routed to all nodes always.
NOTE: If variable is written as a part of write query, it is treated like write query and not routed to all servers. For example, INSERT INTO test.t1 VALUES (@myvar:= 7) .
**NOTE**: If variable is written as a part of write query, it is treated like write query and not routed to all servers. For example, `INSERT INTO test.t1 VALUES (@myvar:= 7)` will be routed to the master and an error in the error log will be written.
Examples:
#### Examples of session command limitations
If new database "db" was created and client executes “USE db” and it is routed to slave before the CREATE DATABASE clause is replicated to all slaves there is a risk of executing query in wrong database. Similarly, if any response that RWSplit sends back to the client differ from that of the master, there is a risk for misbehavior.
If new database "db" was created and client executes “USE db” and it is routed to slave before the CREATE DATABASE clause is replicated to all slaves there is a risk of executing query in wrong database. Similarly, if any response that RWSplit sends back to the client differ from that of the master, there is a risk for misbehavior. To prevent this, any failures in session command execution are treated as fatal errors and all connections by the session to that particular slave server will be closed. In addition, the server will not used again for routing for the duration of the session.
Most imaginable reasons are related to replication lag but it could be possible that a slave fails to execute something because of some non-fatal, temporary failure while execution of same command succeeds in other backends.

View File

@ -103,7 +103,7 @@ In Master-Slave replication cluster also read-only queries are routed to master
### Limitations in client session handling
Some of the queries that client sends are routed to all backends instead of sending them just to one of server. These queries include "USE <db name>" and SET autocommit=0 among many others. Readwritesplit sends a copy of these queries to each backend server and forwards the master's reply to the client. Below is a list of MySQL commands which are classified as session commands :
Some of the queries that client sends are routed to all backends instead of sending them just to one of server. These queries include `USE <db name>` and `SET autocommit=0` among many others. Readwritesplit sends a copy of these queries to each backend server and forwards the master's reply to the client. Below is a list of MySQL commands which are classified as session commands :
```
COM_INIT_DB (USE <db name> creates this)
@ -138,7 +138,7 @@ SELECT ..INTO variable|OUTFILE|DUMPFILE
SET autocommit=1|0
```
There is a possibility for misbehavior; if `USE mytable` was executed in one of the slaves and it failed, it may be due to replication lag rather than the fact it didn’t exist. Thus the same command may end up with different result among backend servers. This disparity is missed.
There is a possibility for misbehavior; if `USE mytable` was executed in one of the slaves and it failed, it may be due to replication lag rather than the fact it didn’t exist. Thus the same command may end up with different result among backend servers. The slaves which fail to execute a session command will be dropped from the active list of slaves for this session to guarantee a consistent session state across all the servers that are in use by the session.
The above-mentioned behavior can be partially controller with the `use_sql_variables_in` configuration parameter.
@ -152,11 +152,10 @@ Server-side session variables are called as SQL variables. If "master" or no val
### Examples of limitations
If new database "db" was created and client executes “USE db” and it is routed to slave before the CREATE DATABASE clause is replicated to all slaves there is a risk of executing query in wrong database. Similarly, if any response that RWSplit sends back to the client differ from that of the master, there is a risk for misbehavior.
If new database "db" was created and client executes “USE db” and it is routed to slave before the CREATE DATABASE clause is replicated to all slaves there is a risk of executing query in wrong database. Similarly, if any response that RWSplit sends back to the client differ from that of the master, there is a risk for misbehavior. To prevent this, any failures in session command execution are treated as fatal errors and all connections by the session to that particular slave server will be closed. In addition, the server will not used again for routing for the duration of the session.
Most imaginable reasons are related to replication lag but it could be possible that a slave fails to execute something because of some non-fatal, temporary failure while execution of same command succeeds in other backends.
## Examples
Examples of the readwritesplit router in use can be found in the [Tutorials](../Tutorials) folder.