Consistent Critical Read Filter
Consistent Critical Read Filter
This filter was introduced in MariaDB MaxScale 2.1.
The Consistent Critical Read (CCR) filter allows consistent critical reads to be done through MaxScale while still allowing scaleout of non-critical reads.
When the filter detects a statement that would modify the database, it attaches a routing hint to all following statements done by that connection. This routing hint guides the routing module to route the statement to the master server where data is guaranteed to be in an up-to-date state. Writes from one session do not, by default, propagate to other sessions.
Note: This filter does not work with prepared statements. Only text protocol queries are handled by this filter.
Controlling the Filter with SQL Comments
The triggering of the filter can be limited further by adding MaxScale supported comments to queries and/or by using regular expressions. The query comments take precedence: if a comment is found it is obeyed even if a regular expression parameter might give a different result. Even a comment cannot cause a SELECT-query to trigger the filter. Such a comment is considered an error and ignored.
The comments must follow the MaxScale hint syntax
and the HintFilter needs to be in the filter chain before the CCR-filter. If a
query has a MaxScale supported comment line which defines the parameter
that comment is caught by the CCR-filter. Parameter values
are supported, causing the filter to trigger (
match) or not trigger (
on receiving the write query. For example, the query
INSERT INTO departments VALUES ('d1234', 'NewDepartment'); -- maxscale ccr=ignore
would normally cause the filter to trigger, but does not because of the
match-comment typically has no effect, since write queries by
default trigger the filter anyway. It can be used to override an ignore-type
regular expression that would otherwise prevent triggering.
The CCR filter has no mandatory parameters.
The time window during which queries are routed to the master. The duration can be specified as documented here but the value will always be rounded to the nearest second. If no explicit unit has been specified, the value is interpreted as seconds in MaxScale 2.4. In subsequent versions a value without a unit may be rejected. The default value for this parameter is 60 seconds.
When a data modifying SQL statement is processed, a timer is set to the value of time. Once the timer has elapsed, all statements are routed normally. If a new data modifying SQL statement is processed within the time window, the timer is reset to the value of time.
Enabling this parameter in combination with the count parameter causes both the time window and number of queries to be inspected. If either of the two conditions are met, the query is re-routed to the master.
The number of SQL statements to route to master after detecting a data modifying SQL statement. This feature is disabled by default.
After processing a data modifying SQL statement, a counter is set to the value of count and all statements are routed to the master. Each executed statement after a data modifying SQL statement cause the counter to be decremented. Once the counter reaches zero, the statements are routed normally. If a new data modifying SQL statement is processed, the counter is reset to the value of count.
These regular expression settings control which statements trigger statement re-routing. Only non-SELECT statements are inspected. For CCRFilter, the exclude-parameter is instead named ignore, yet works similarly.
match=.*INSERT.* ignore=.*UPDATE.* options=case,extended
global is a boolean parameter that when enabled causes writes from one
connection to propagate to all other connections. This can be used to work
around cases where one connection writes data and another reads it, expecting
the write done by the other connection to be visible.
This parameter only works with the
time parameter. The use of
count at the same time is not allowed and will be treated as an error.
Here is a minimal filter configuration for the CCRFilter which should solve most problems with critical reads after writes.
[CCRFilter] type=filter module=ccrfilter time=5
With this configuration, whenever a connection does a write, all subsequent reads done by that connection will be forced to the master for 5 seconds.
This prevents read scaling until the modifications have been replicated to the slaves. For best performance, the value of time should be slightly greater than the actual replication lag between the master and its slaves. If the number of critical read statements is known, the count parameter could be used to control the number reads that are sent to the master.