X-MimeOLE: Produced By Microsoft Exchange V6.5
Received: by onstor-exch02.onstor.net 
	id <01C7C1A4.7D1B2373@onstor-exch02.onstor.net>; Sun, 8 Jul 2007 13:11:02 -0800
MIME-Version: 1.0
Content-Type: multipart/alternative;
	boundary="----_=_NextPart_001_01C7C1A4.7D1B2373"
References: <BB375AF679D4A34E9CA8DFA650E2B04E01955CA6@onstor-exch02.onstor.net> <20070708134758.7e99f65f@ripper.onstor.net>
Content-class: urn:content-classes:message
Subject: RE: Linux clustering software
Date: Sun, 8 Jul 2007 13:11:02 -0800
Message-ID: <BB375AF679D4A34E9CA8DFA650E2B04E01955CA8@onstor-exch02.onstor.net>
X-MS-Has-Attach: 
X-MS-TNEF-Correlator: 
Thread-Topic: Linux clustering software
Thread-Index: AcfBoUT4VNc42O4JT/GaKeZvhzOozgAAgnKn
From: "Jonathan Goldick" <jonathan.goldick@onstor.com>
To: "Andy Sharp" <andy.sharp@onstor.com>
Cc: "Ian Brown" <ian.brown@onstor.com>

This is a multi-part message in MIME format.

------_=_NextPart_001_01C7C1A4.7D1B2373
Content-Type: text/plain;
	charset="iso-8859-1"
Content-Transfer-Encoding: quoted-printable

Good to know.  As part of our software roadmap I would like to replace =
the AFS Ubik code with the following:
=20
. HeartBeat
. This, or similar, Cluster Manager
. MySQL Cluster on top of DRDB, which works on top of HeartBeat and has =
the desired shared-nothing storage architecture.  While we do have a SAN =
and could store the DB there at some point, the ability to use the =
compact flash allows us to avoid solving the boot-strapping problems up =
front.
=20
We have little IP in our clustering code, but huge amounts of actual =
code written.  It's a maintainence problem we can ditch going forward.
=20
BTW, at some point soon I want to start having some brainstorming =
conversations/lunches on my random collection of thoughts on these =
topics.  I have a pretty good idea what I want us to do in a variety of =
areas, but it's enormously valuable to have people around who can say, =
"Oh yeah, we tried that and it sucked!"

________________________________

From: Andy Sharp
Sent: Sun 7/8/2007 1:47 PM
To: Jonathan Goldick
Cc: Ian Brown
Subject: Re: Linux clustering software



I'm quite familiar with this space as we developed a heartbeat-like
program as part of our HA suite at Continuous Computing called UpBeat.
Ours was very much like this paper is proposing, which is basically just
adding to the heartbeat package the things that it is missing compared
to UpBeat.

Getting heartbeat programs and algo's to work is actually rather tricky
and fraught with corner cases -- the only way to solve all the corner
cases is to force the user to choose (and understand?) particular
policies -- default policies always end up being "not what we wanted"
right at some incovenient moment.  Especially when it comes to
replicating data, which is really all that we are trying to do:
replicate and serve data across a cluster.  Should be simple, right?
~:^)

On Sun, 8 Jul 2007 13:05:37 -0700 "Jonathan Goldick"
<jonathan.goldick@onstor.com> wrote:

> Do either of you have any experience with Linux clustering software
> like
> =
http://www.linux-ha.org/_cache/TechnicalPapers__UKUUG-WinterConf-2004-SCR=
AT-Paper.pdf ?
> I'm spending some time on alternatives to our current software layers
> and where open source is as good or better for the SSC.  At the
> moment I'm looking into ways to split our cluster DB into a
> clustering software layer like the one above based on HeartBeat, and
> a replicated DB, perhaps MySQL. Note that this is not some imminent
> thing but more of roadmap exercise. Thanks, Jonathan



------_=_NextPart_001_01C7C1A4.7D1B2373
Content-Type: text/html;
	charset="iso-8859-1"
Content-Transfer-Encoding: quoted-printable

<HTML dir=3Dltr><HEAD><TITLE>Re: Linux clustering software</TITLE>=0A=
<META http-equiv=3DContent-Type content=3D"text/html; charset=3Dunicode">=0A=
<META content=3D"MSHTML 6.00.6000.16481" name=3DGENERATOR></HEAD>=0A=
<BODY>=0A=
<DIV id=3DidOWAReplyText52983 dir=3Dltr>=0A=
<DIV dir=3Dltr><FONT face=3DArial color=3D#000000 size=3D2>Good to =
know.&nbsp; As part of our software roadmap I would like to&nbsp;replace =
the AFS Ubik code with&nbsp;the following:</FONT></DIV>=0A=
<DIV dir=3Dltr><FONT face=3DArial size=3D2></FONT>&nbsp;</DIV>=0A=
<DIV dir=3Dltr><FONT face=3DArial color=3D#000000 size=3D2>. =
HeartBeat</FONT></DIV>=0A=
<DIV dir=3Dltr><FONT face=3DArial color=3D#000000 size=3D2>. This, or =
similar,&nbsp;Cluster Manager</FONT></DIV>=0A=
<DIV dir=3Dltr><FONT face=3DArial size=3D2>. MySQL Cluster on top of =
DRDB, which works on top of HeartBeat and has the desired shared-nothing =
storage architecture.&nbsp; While we do have a SAN and could store the =
DB there at some point, the ability to use the compact flash allows us =
to avoid solving the boot-strapping problems up front.</FONT></DIV></DIV>=0A=
<DIV dir=3Dltr>&nbsp;</DIV>=0A=
<DIV dir=3Dltr>We have little IP in our clustering code, but huge =
amounts of actual code written.&nbsp; It's a maintainence problem we can =
ditch going forward.</DIV>=0A=
<DIV dir=3Dltr>&nbsp;</DIV>=0A=
<DIV dir=3Dltr>BTW, at some point soon I want to start having some =
brainstorming conversations/lunches on my random collection of thoughts =
on these topics.&nbsp; I have a pretty good idea what I want us to do in =
a variety of areas, but it's enormously valuable to have people around =
who can say, "Oh yeah, we tried that and it sucked!"</DIV>=0A=
<DIV dir=3Dltr><BR>=0A=
<HR tabIndex=3D-1>=0A=
<FONT face=3DTahoma size=3D2><B>From:</B> Andy Sharp<BR><B>Sent:</B> Sun =
7/8/2007 1:47 PM<BR><B>To:</B> Jonathan Goldick<BR><B>Cc:</B> Ian =
Brown<BR><B>Subject:</B> Re: Linux clustering =
software<BR></FONT><BR></DIV>=0A=
<DIV>=0A=
<P><FONT size=3D2>I'm quite familiar with this space as we developed a =
heartbeat-like<BR>program as part of our HA suite at Continuous =
Computing called UpBeat.<BR>Ours was very much like this paper is =
proposing, which is basically just<BR>adding to the heartbeat package =
the things that it is missing compared<BR>to UpBeat.<BR><BR>Getting =
heartbeat programs and algo's to work is actually rather tricky<BR>and =
fraught with corner cases -- the only way to solve all the =
corner<BR>cases is to force the user to choose (and understand?) =
particular<BR>policies -- default policies always end up being "not what =
we wanted"<BR>right at some incovenient moment.&nbsp; Especially when it =
comes to<BR>replicating data, which is really all that we are trying to =
do:<BR>replicate and serve data across a cluster.&nbsp; Should be =
simple, right?<BR>~:^)<BR><BR>On Sun, 8 Jul 2007 13:05:37 -0700 =
"Jonathan Goldick"<BR>&lt;jonathan.goldick@onstor.com&gt; =
wrote:<BR><BR>&gt; Do either of you have any experience with Linux =
clustering software<BR>&gt; like<BR>&gt; <A =
href=3D"http://www.linux-ha.org/_cache/TechnicalPapers__UKUUG-WinterConf-=
2004-SCRAT-Paper.pdf">http://www.linux-ha.org/_cache/TechnicalPapers__UKU=
UG-WinterConf-2004-SCRAT-Paper.pdf</A> ?<BR>&gt; I'm spending some time =
on alternatives to our current software layers<BR>&gt; and where open =
source is as good or better for the SSC.&nbsp; At the<BR>&gt; moment I'm =
looking into ways to split our cluster DB into a<BR>&gt; clustering =
software layer like the one above based on HeartBeat, and<BR>&gt; a =
replicated DB, perhaps MySQL. Note that this is not some =
imminent<BR>&gt; thing but more of roadmap exercise. Thanks, =
Jonathan<BR></FONT></P></DIV></BODY></HTML>
------_=_NextPart_001_01C7C1A4.7D1B2373--
