maria-developers team mailing list archive
Mailing list archive
Re: Transactional INFORMATION_SCHEMA / status variables?
For my personal opinion, I can go either way. What is most important on a
high query database is non-blocking. For example, if the hard drive is
acting funny, show master status, show slave status, and certain other show
commands can stall. Other show commands, like show variables and show global
status never stall.
When it comes to information schema, the top priority is no stalling. This
makes monitoring more reliable. We've had situations where a query detecting
the status of a machine would stall, and then every 5 minutes it would check
again, and stall, and over time it eats up the connections with dead
When I hear about transactions with information schema, I get a little antsy
because of the possibility of stalls. I'd rather it be inaccurate slightly,
or to be able to turn off and on a variable for the transactions.
On Wed, Oct 6, 2010 at 3:03 AM, Kristian Nielsen
> Would it make sense to have transactional behaviour for status variables,
> and/or information_schema tables? Or does this break the code and/or user
> expectations too much?
> The motivation is to follow up on MWL#116, group commit, which introduces
> consistent commit order between storage engines and binary log.
> I want to use this to get a consistent binlog position for START
> WITH CONSISTENT SNAPSHOT, without taking any additional locks. Currently, I
> believe it is common (eg. mysqlbinlog --master-data --single-transaction)
> do something like this:
> FLUSH TABLES WITH READ LOCK;
> START TRANSACTION WITH CONSISTENT SNAPSHOT;
> UNLOCK TABLES
> <take consistent backup by dumping tables using the consistent snapshot>
> and this takes a lock that can depending on circumstances severely affect
> One idea is to let the binlog storage engine participate in START
> WITH CONSISTENT SNAPSHOT, by installing a start_consistent_snapshot()
> in its handlerton. And then do something like this:
> START TRANSACTION WITH CONSISTENT SNAPSHOT;
> SELECT variable_value FROM INFORMATION_SCHEMA.SESSION_STATUS WHERE
> variable_name IN ('binlog_master_file', 'binlog_master_position');
> <dump other transactional tables>
> If the SELECT of binlog_master_file and binlog_master_position could be
> transactional, then the binlog engine could return the correct values
> associated with the consistent snapshot, without blocking any other
> I like the simplicity of this idea, but I do not understand server
> handling enough to be sure it will work well, some concerns:
> - Using SHOW STATUS / INFORMATION_SCHEMA.SESSION_STATUS like this could be
> surprising to people that do not expect different parts of the results to
> display different kinds of transactional behaviour. This could be helped
> using instead a separate INFORMATION_SCHEMA table for the binlog
> - But this raises another concern, if an INFORMATION_SCHEMA can be
> transactional in this respect?
> - Also, it seems to me that merely selecting from such transactional
> INFORMATION_SCHEMA table would then start a new transaction inside the
> binlog engine. I wonder if this would cause any unpleasant side effects?
> Any suggestions or comments?
> If this does not work, I have another idea, which I think is more general,
> also more complicated to implement.
> The idea is that every transaction has a local transaction ID, assigned at
> start (we already have this, in thd->transaction.xid_state.xid).
> Each engine will transactionally store the local transaction ID of the last
> transaction committed. The binary log will similarly store this ID along
> every transaction that is binlogged.
> Then START TRANSACTION WITH CONSISTENT SNAPSHOT could optionally return the
> local transaction ID of the last committed transaction visible in the
> obtained. This local transaction ID could then be mapped to binlog position
> (with eg. mysqldump), and more generally any binlog plugin could provide a
> to map such local transaction ID into its own global transaction ID.
> Similarly, after restore of InnoDB hot backup or LVM snapshot, one could
> the engine for last committed local transaction ID, and map this to binlog
> position / global transaction ID to be able to use the restored backup to
> provision a new slave.
> This would work with any storage engine and any binlog/replication
> implementation, without any need for FLUSH TABLES WITH READ LOCK.
> - Kristian.
> Mailing list: https://launchpad.net/~maria-developers
> Post to : maria-developers@xxxxxxxxxxxxxxxxxxx
> Unsubscribe : https://launchpad.net/~maria-developers
> More help : https://help.launchpad.net/ListHelp