[aur-general] TU Application -Thomas Hatch

Thomas S Hatch thatch45 at gmail.com
Wed Jan 5 15:44:37 EST 2011


On Wed, Jan 5, 2011 at 1:35 PM, Kaiting Chen <kaitocracy at gmail.com> wrote:

> On Wed, Jan 5, 2011 at 2:50 PM, Thomas S Hatch <thatch45 at gmail.com> wrote:
>
> > The difference is that Gluster is a nightmare!
> >
> > The problem with gluster is that the replication is tiered, and that
> there
> > is no metadata. The client is then the master, this means that if you
> > connect to gluster with a mis-configured client you can have large scale
> > data corruption.
> >
> > Next since the replication of data is tiered you don't have true
> > replication, so only the gluster server you connect to to save the data
> has
> > the correct data, if that server goes down the replications are old and
> you
> > have data corruption.
> >
> > The gluster devs actually had to recall gluster 3.1 because the data
> > corruption was rampant.
> >
> > The difference between gluster and MooseFS is that MooseFS works!
> >
> > MooseFS also has a cool web frontend :)
> >
> > We were using gluster and the business cost became catastrophic, picking
> up
> > the peices was a nightmare.
> >
> > MooseFS saves data to replication nodes in paralell! MooseFS maintains a
> > master metalogger so client connections are agnostic.
> > MooseFS maintains metadata replication so you can restore is something
> > happens to the master.
> >
> > I take it you don't like FUSE? EVERYBODY is doing it ;)
> >
> > I am looking forward to Ceph, which does not require fuse, but I don't
> > think
> > it is going to be production ready for at least a year, and MooseFS
> easily
> > compete with Ceph IMHO.
> >
> > If there are GlusterFS devs in the room, please disregard the previous
> rant
> > :)
> >
>
> Thanks for the very thorough answer. And yes I hate the idea of a
> filesystem
> in userspace. Everyone knows the FS's should be in kernel space! Mostly
> it's
> the fact that in my opinion bypassing the kernel's caching mechanism is
> entirely impractical for a high performance FS. Feel free to correct me if
> I'm wrong.
>
> Anyways your application looks really good. Good luck! --Kaiting.
>
> --
> Kiwis and Limes: http://kaitocracy.blogspot.com/
>

Thanks Kaiting!

I was looking back at my post and worrying that it was too much of a rant :)
I spend a month full time testing and trying distributed filesystems and my
conclusion was that MooseFS is the best one out there.

I agree that fuse is something that should be used with caution, and that a
filesystem does belong in kernel space. But in the situation of a
distributed filesystem I thought that the benefits of fuse in allowing for
higher flexibility made it a permissible option.  All in all I think that
fuse for network filesystems can be a huge advantage, on the other hand I am
much more cautious about local filesystems that operate behind fuse.

So to sum it up, I feel good with moosefs, sshfs, but am cautious when
looking at things like zfs-fuse.

But with that said, I don't believe in disparaging someone's project on the
grounds of what it is, the fact that people are creating new things and
sharing them is a wonderful thing!

On the other hand, I will speak my mind when someone's project corrupts my
data :)


More information about the aur-general mailing list