Ed -
If I understand it looks like you are recommending a method for implementing an asynchronous replication solution as a possible alternative to the current synchronous replication method?
Thanks.
--
Craig Carl
Gluster, Inc.
Cell - (408) 829-9953 (California, USA)
Gtalk - address@hidden
From: "Ed W" <address@hidden>
To: address@hidden
Sent: Thursday, September 23, 2010 1:32:32 PM
Subject: [Gluster-devel] Can I bring a development idea to Dev's attention?
Hi Glusterfs Dev's,
I might be trying to teach you to suck eggs, but I was reading through
the publications on http://www.xtreemfs.org/publications.php and one of
the ideas there seemed very relevant to gluster
Please have a look at their FatLease papers and the Paxos lease
negotiation algorithm. It was new to me, but seems like an elegant and
robust (google use it) way of managing a distributed lock scenario?
In particular it would seem to be a very interesting way to introduce
the idea of optimistic locking to the gluster type infrastructure. What
I'm thinking here is the situation where you have a client talking to a
single brick in a replicated set, that it can effectively optimistically
lock a localised set of files/directories on that brick and so further
reads/modifications can be lazily written out to other servers (without
compromising coherency). If a read/write goes to one of the other
replicas then of course it must first break the other servers optimistic
lock before continuing and effectively you temporarily revert back to
the current situation where every file access causes a network access to
all other servers.
The win is clearly where you end up with clients localising their access
for certain files to certain servers and that server will then benefit
from holding an optimistic lock, since no network activity is generated
for further read access and lazy writes become possible without split
brain risk.
eg consider a master/master setup with serveral webservers. Especially
where each physical machine touches only a subset of the files (eg each
machine usually only serves a subset of data), then that machine can
start to access the gluster share at basically native disk speeds,
having acquired an optimistic lock on that subset of data (no need to
check with other bricks since we now have a lock on the data). This
would seem to be a huge win for a large class of problems?
(Thinking about it another way, it's a bit like a standard optimistic
locking architecture, where one server gets promoted to become the
master reader/writer at a time (for a subset of files) and no other
server can read/write those files without first breaking the lock by
requesting to become the master. The difference is that Paxos allows
this to be done robustly and efficiently without a centralised lock server)
The clever part seems to be the fairly stateless method that is used to
bootstrap a system with stateful leases (read the publications). I
hadn't come across the Fatlease/Paxos idea before and it seems extremely
clever
Cheers
Ed W
_______________________________________________
Gluster-devel mailing list
address@hidden
http://lists.nongnu.org/mailman/listinfo/gluster-devel