Recovery?

[Mark Smith]

> what to do if a content server explodes? Is there a script to restore the
> poor system using the other hosts?

Well, MogileFS is designed to have multiple content servers.  Every file
has a minimum replica count which specifies how many servers it has to be
on, and so it's replicated across that many machines.

If a machine implodes, all files should be available from another machine.
So, you'd then mark a machine as 'dead' in the database, allocate a new
machine, new device ids, and files would be automatically replicated to the
new machine.

We're working on a MogileFS admin tool (mogadm) which will enable you to
mark hosts as down/dead in order to do this process without having to mess
with the database directly, but haven't gotten to it yet.

> Is there a (theoretical) way to replace the mysql dbase by something less,
> well, dbase like? What do you use with livejournal for the tracker? Some master-master
> setup?

Right now, we point the MySQL host at one machine, and then we have a spare
machine as a slave doing nothing but replicating and keeping a copy of the
SQL database.  We also frequently back this data up to our backup array.

There really isn't any way to use something else, not at this point.
You'd have to convert all of the mogilefsd queries to use some other
storage method, which would be a good bit of effort (but not impossible, of
course).

> Where are the database table definitions anyway? :)

Heh, look in the devnotes directory, I think it's called.  Either way, the
file is sql.txt.

The setup process is basically to pipe this at your MySQL database so it
sets up the tables.  If the scheme changes later, we modify mogilefsd to
perform the changes for you (with a --upgrade flag it'll tell you about).

--
Mark Smith
junior at danga.com