Re: [Mysqlfs-general] mysqlfs patch

[Andrew R. <ros...@gm...>]

On 21/08/07, Michal Ludvig <mi...@lo...> wrote:
> Hi Andrew,
>
> > I've recently created a patch for mysqlfs.  The idea of the patch is
> > to check for the existence of an entry by key in the data_blocks table
> > before attempting to write (query_write_one_block).  The reason? to
> > stop mysqlfs killing MySQL replication when attempting to insert an
> > data_block with the same inode/seq as one that already exists.
>
> Interesting. What version of MySQL are you running on both master and
> slave? Replication should never break when you attempt to insert a
> record with a primary key that already exists. In fact *any* user action
> or query should never break MySQL replication. If it does in your case
> and if it's reproducible I'd consider it a MySQL server bug worth filing
> a bugreport.
>

I am currently running a chained master->master setup. It is a common
MySQL replication problem with master->master setups, and crops up
when the PK's are not auto incrementing.

> > Now I'm not that familiar with the mysqlfs code, and was hoping
> > someone with more knowledge than myself could let me know if I've gone
> > in the right direction.
> >
> > http://dev.iris.ac/~ody/mysqlfs-x86_64/mysqlfs-0.4.0-rc1.patch
>
> Problem is that your query_inode_key_exists() and the subsequent insert
> are two independent transactions. I.e. even if your check returns false
> the inode can appear in another thread before the first thread decides
> to run the INSERT. You'd have to make it atomic to be safe, which in
> case of MYISAM tables means locking the whole table and therefore
> resulting to a worse performance.

I figured there would be a problem with threading.  Have you thought
about maybe providing some shared memory locking functionality? ;)

Many thanks for the help.  For now I think I'll keep my fingers
crossed and hope that what I have done has lowered the odds enough
that I will rarely, if ever, see a problem, considering there is
minimal writes going on to my mounted filesystems anyway.

cheers
Andrew