MogileFS status update
brad at danga.com
Mon Nov 13 02:06:46 UTC 2006
A little MogileFS status update ...
LiveJournal has switched to using the current svn version. As a result, a
bunch of little bugs were found really quickly and patched in the past
The deleter job was also rewritten to be a lot smarter and handle all the
failure cases. (Like Dormando, we also got bit by the dumb deleter...
which has been dumb since 1.x....) Basically it can now know ahead of
time when something will fail and reschedule it smartly for later, or
reschedule it once a delete from a device fails. (this was all done this
evening, and has been tested, but we're not running it in production
Also, mogstored has been updated to use (and require) the new features in
Perlbal 1.52 to isolate the Linux::AIO/IO::AIO aio threads from being
starved by one slow/dying device. Now, when mogstored starts, it starts
10 aio threads per device, capped at 100 (they're very light), and each
device can only have 10 outstanding aios at a time. You can now telnet
into the mogstoreds and type "aio" on the management port to get current
and historical statistics: current outstanding, total ever, current
queued (if 10 are outstanding), total ever queued.
Also, the mogadm command is undergoing major love as of this weekend, so
it's now much more interactively-helpful, telling you the command syntaxes
as you go, if you leave parts off. It's a work in progress, but I'm happy
with the guts of it and the first few commands I've converted.
Also, Dormando's committed to hacking on this periodically with me
somewhere between San Jose and San Francisco, if anybody knows a good
coffee shop and wants to join.
More information about the mogilefs