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 rcoli@digg.com designates
 209.85.222.174 as permitted sender)
Message-ID: <4BF6C61E.3080903@digg.com>
Date: Fri, 21 May 2010 10:42:54 -0700
From: Rob Coli <rcoli@digg.com>
User-Agent: Mozilla/5.0 (Macintosh; U; Intel Mac OS X 10.6; en-US;
 rv:1.9.1.9) Gecko/20100317 Thunderbird/3.0.4
MIME-Version: 1.0
To: user@cassandra.apache.org
Subject: Re: Multiple hard disks configuration
References: <AANLkTimx95chN8nUOVXsDYo9wF9Mq0ASVL487faiycBv@mail.gmail.com>
 <AANLkTilC_-vqTm67Y2Zt_O7yBahQIgzbRxHOkPKmAPhK@mail.gmail.com>
In-Reply-To: <AANLkTilC_-vqTm67Y2Zt_O7yBahQIgzbRxHOkPKmAPhK@mail.gmail.com>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

> On Mon, May 17, 2010 at 10:39 PM, Ma Xiao<maxiao@robowork.com>  wrote:
>> Hi all,
>>      Recently we have a 5 nodes running cassandra, 4 X 1.5TB drives for each,
>> ...then I put 4 paths with DataFileDirectory,
>> my question is what's going to happen when one of the disk fail, especialy
>> the one has os installed which also holds the commit log? can we simply
>> replace the disk, and cassandra will get the replica write back? or it
>> should be deal  with as entire node faile(wipe all date on the node and
>> rejoin the ring), that will involve mass data copy from other node but we
>> have just one disk fail. Note, we dont have hardware level raid installed.
>> Any suggestion?

On 5/18/10 11:23 PM, Jonathan Ellis wrote:
 > Yes, you can rely on replication for this (run nodetool repair).

@Ma Xiao :

Nodes which have lost data due to hardware failure and are "repair"ing 
can serve "no value" reads for data they think they have but do not 
actually currently have, when read with ConsistencyLevel.ONE. Each one 
of these no value reads triggers a read repair, however, which should 
result in your hottest data being repaired relatively quickly.

As I understand it, using your 750gb of data case as an example, you are 
likely to serve a non-trivial number of these no value reads if you read 
with CL.ONE in a disk failure-then-repair scenario. Reading with 
CL.QUORUM avoids this risk.

=Rob