Mailing-List: contact user-help@cassandra.apache.org; run by ezmlm
Precedence: bulk
Reply-To: user@cassandra.apache.org
Received-SPF: pass (nike.apache.org: domain of ssrameez@gmail.com designates
 209.85.192.176 as permitted sender)
MIME-Version: 1.0
In-Reply-To: 
 <CAEDUwd3kjrqtesDn3Cb5s=e=YFVr1KUu+FScU=zNkQrjefEpcg@mail.gmail.com>
References: 
 <CA+LRppgbt9+AfztyN6=oMtGrKZTSnRPac4YJf5kNaPKRnHiBiQ@mail.gmail.com>
	<CAEDUwd3kjrqtesDn3Cb5s=e=YFVr1KUu+FScU=zNkQrjefEpcg@mail.gmail.com>
Date: Wed, 6 Aug 2014 08:13:03 +0530
Message-ID: 
 <CAFPQ3xZoGUrp30FbXQG8CwFrLk4=yed3CfqLrUwp6HMGLn8tWQ@mail.gmail.com>
Subject: Re: Make an existing cluster multi data-center compatible.
From: Rameez Thonnakkal <ssrameez@gmail.com>
To: user@cassandra.apache.org
Content-Type: multipart/alternative; boundary=047d7bfe9e2064f4e804ffecee34

--047d7bfe9e2064f4e804ffecee34
Content-Type: text/plain; charset=UTF-8

I think the RAC placement of these 12 nodes will become important. As the
12 nodes are placed in SimpleSnitch, which is not RAC aware, it would be
good to retain them in single RAC in the property file snitch also
initially. node repair is a safe option. If you need to change the RAC
placement, my take would be to increase the Replication factor to atleast 3
and then distribute the nodes in different RAC.

This is not an expert opinion but a newbie thought.

Regards,
Rameez


On Tue, Aug 5, 2014 at 11:35 PM, Robert Coli <rcoli@eventbrite.com> wrote:

> On Tue, Aug 5, 2014 at 3:52 AM, Rene Kochen <rene.kochen@schange.com>
> wrote:
>
>> Do I have to run full repairs after this change? Because the yaml file
>> states: IF YOU CHANGE THE SNITCH AFTER DATA IS INSERTED INTO THE CLUSTER,
>> YOU MUST RUN A FULL REPAIR, SINCE THE SNITCH AFFECTS WHERE REPLICAS ARE
>> PLACED.
>>
>
> As long as you correctly configure the new snitch so that the replica sets
> do not change, no, you do not need to repair.
>
> Barring that, if you manage to transform the replica set in such a way
> that you always have one (fully repaired) replica from the old set, repair
> will help. I do not recommend this very risky practice. In practice the
> only transformation of snitch in a cluster with data which is likely to be
> safe is one whose result is a NOOP in terms of replica placement.
>
> In fact, the yaml file is stating something unreasonable there, because
> repair cannot protect against this case :
>
> - 6 node cluster, A B C D E F,  RF = 2
>
> 1) Start with SimpleSnitch so that A, B have the two replicas of row key X.
> 2) Write row key X, value Y, to nodes A and B.
> 2) Change to OtherSnitch so that now C,D are responsible for row key X.
> 3) Repair and notice that neither C nor D answer "Y" when asked for row X.
>
> =Rob
>
>

--047d7bfe9e2064f4e804ffecee34
Content-Type: text/html; charset=UTF-8
Content-Transfer-Encoding: quoted-printable

<div dir=3D"ltr"><div><div>I think the RAC placement of these 12 nodes will=
 become important. As the 12 nodes are placed in SimpleSnitch, which is not=
 RAC aware, it would be good to retain them in single RAC in the property f=
ile snitch also initially. node repair is a safe option. If you need to cha=
nge the RAC placement, my take would be to increase the Replication factor =
to atleast 3 and then distribute the nodes in different RAC. <br>
<br></div>This is not an expert opinion but a newbie thought. <br><br></div=
>Regards,<br>Rameez<br></div><div class=3D"gmail_extra"><br><br><div class=
=3D"gmail_quote">On Tue, Aug 5, 2014 at 11:35 PM, Robert Coli <span dir=3D"=
ltr">&lt;<a href=3D"mailto:rcoli@eventbrite.com" target=3D"_blank">rcoli@ev=
entbrite.com</a>&gt;</span> wrote:<br>
<blockquote class=3D"gmail_quote" style=3D"margin:0 0 0 .8ex;border-left:1p=
x #ccc solid;padding-left:1ex"><div dir=3D"ltr"><div class=3D"gmail_extra">=
<div class=3D"gmail_quote"><div class=3D"">On Tue, Aug 5, 2014 at 3:52 AM, =
Rene Kochen <span dir=3D"ltr">&lt;<a href=3D"mailto:rene.kochen@schange.com=
" target=3D"_blank">rene.kochen@schange.com</a>&gt;</span> wrote:<br>

<blockquote class=3D"gmail_quote" style=3D"margin:0 0 0 .8ex;border-left:1p=
x #ccc solid;padding-left:1ex"><div dir=3D"ltr">Do I have to run full repai=
rs after this change? Because the yaml file states: IF YOU CHANGE THE SNITC=
H AFTER DATA IS INSERTED INTO THE CLUSTER, YOU MUST RUN A FULL REPAIR, SINC=
E THE SNITCH AFFECTS WHERE REPLICAS ARE PLACED.<br>

</div></blockquote><div><br></div></div><div>As long as you correctly confi=
gure the new snitch so that the replica sets do not change, no, you do not =
need to repair.</div><div><br></div><div>Barring that, if you manage to tra=
nsform the replica set in such a way that you always have one (fully repair=
ed) replica from the old set, repair will help. I do not recommend this ver=
y risky practice. In practice the only transformation of snitch in a cluste=
r with data which is likely to be safe is one whose result is a NOOP in ter=
ms of replica placement.</div>

<div><br></div><div>In fact, the yaml file is stating something unreasonabl=
e there, because repair cannot protect against this case :</div><div>=C2=A0=
</div><div>- 6 node cluster, A B C D E F, =C2=A0RF =3D 2</div><div><br></di=
v><div>
1) Start with SimpleSnitch so that A, B have the two replicas of row key X.=
</div>
<div>2) Write row key X, value Y, to nodes A and B.</div><div>2) Change to =
OtherSnitch so that now C,D are responsible for row key X.</div><div>3) Rep=
air and notice that neither C nor D answer &quot;Y&quot; when asked for row=
 X.</div>

<div><br></div><div>=3DRob</div><div><br></div></div></div></div>
</blockquote></div><br></div>

--047d7bfe9e2064f4e804ffecee34--