Riak on SAN
guido.medina at temetra.com
Thu Oct 3 10:22:47 EDT 2013
Some users might avoid the ZFS overheads, remember we are on a KV world
where to read/write many keys you will have to do so concurrently, say
there is less than 1% chances for things things going wrong with 1 a
server belonging to a Riak cluster, if building a Riak server is cheap,
would you pay the overhead price?
It is a choice, you might risk it and remove such overheads which in
theory is not risky at all because Riak will replicate your keys in
different nodes, for example, on a 5 servers cluster, 2 servers will
have to die simultaneously for you to lose data (Someone correct me if
I'm wrong in this part)
All I'm trying to say is that such over data protection might not be
necessary at all with a distributed and high available NoSQL DB like Riak.
On 03/10/13 15:11, Heinz Nikolaus Gies wrote:
> Hi Guido,
> I don’t see how snappy compression renders ZFS useless, you might do
> some things twice like crcing but it also protects on different
> layers. While the ZFS crc protects data on the disks the in app crc
> could protect the data ‘all’ the way up, compression wise you might
> not even turn on ZFS compression and even if you do, you could still
> get a higher ratio given that ZFS will use compression over the entire
> volume not ‘just’ the data in the DB.
> That said there is a lot more to ZFS then compression and CRC ;) like
> snapshots, cloning, ARC ^^
> On 03 Oct 2013, at 9:56, Guido Medina <guido.medina at temetra.com
> <mailto:guido.medina at temetra.com>> wrote:
>> If using LevelDB backend, LevelDB has a nice compression (snappy),
>> including CRC checks and all sort of data corruption checks, I have
>> read on this mail list people that has required to disable snappy
>> compression because it renders ZFS useless (not much to compress
>> after that)
>> Hence, it is kind of related to using ZFS or not, if you go for ZFS
>> whatever variant you will have to support two sub-systems, if you let
>> LevelDB snappy compression on, you won't have to worry about it.
>> As for backup, Basho provides a sort of cluster-to-cluster
>> replication tool, we built our own in Java, making backups per
>> storage on every node won't make much sense due to CAP/distributed
>> nature, replicating the keys to another cluster is what will make sense.
>> Hope that helps and is understandable,
>> On 03/10/13 13:54, Pedram Nimreezi wrote:
>>> Not sure what ZFS has to do with snappy compression, as it's a file
>>> system not a compression algorithm..
>>> feature wise, ZFS is quite possibly the most enterprise file system
>>> around, including advanced data corruption prevention and remote
>>> backing up..
>>> This would be a viable option in BSD/Solaris environments, at least
>>> for making snapshots.
>>> Might make a nice write up for the Basho blog..
>>> Backups for riak I think require a bit more consideration then file
>>> system snapshot send,
>>> and should include provisions for transferring data to
>>> smaller/larger clusters, transfer
>>> ring ownerships properly, etc.
>>> On Thu, Oct 3, 2013 at 7:15 AM, Guido Medina
>>> <guido.medina at temetra.com <mailto:guido.medina at temetra.com>> wrote:
>>> And for ZFS? I wouldn't recommend it, after Riak 1.4 snappy
>>> LevelDB compression does a nice job, why take the risk of yet
>>> another not so enterprise ready compression algorithms.
>>> I could be wrong though,
>>> On 03/10/13 12:11, Guido Medina wrote:
>>>> I have heard some SAN's horrors stories too, Riak nodes are so
>>>> cheap that I don't see the point in even having any mirror on
>>>> the node, here my points:
>>>> 1. Erlang interprocess communication brings some network
>>>> usage, why yet another network usage on replicating the
>>>> data? If the whole idea of Riak is have your data
>>>> replicated in different nodes.
>>>> 2. If a node goes down or die for whatever reason, bring up
>>>> another node and rebuild it.
>>>> 3. If you want to really replicate your cluster Riak offers
>>>> the enterprise replication which I'm quite sure will be
>>>> less expensive than a SAN and will warranty to have your
>>>> cluster ready to go somewhere else as a backup.
>>>> 4. I would even go further, SSDs are so cheap and Riak nodes
>>>> are so cheap now adays that I would even build a cluster
>>>> using RAID 0 or RAID 5 SSDs (yes, no mirror with RAID 1, if
>>>> too afraid, RAID 5), that will have a great impact on
>>>> performance. Again, if something goes wrong with 1 node,
>>>> refer to point 2.
>>>> SANs and all those "legacy" backup and replication IMHO are
>>>> meant for other products, like an Oracle money eater DB server.
>>>> On 03/10/13 12:00, Brian Akins wrote:
>>>>> So, call me naive, but couldn't ZFS be used as Heinze suggested?
>>>>> I have some SAN horror stories - both operationally and from
>>>>> an economic perspective.
>>>>> riak-users mailing list
>>>>> riak-users at lists.basho.com <mailto:riak-users at lists.basho.com>
>>> riak-users mailing list
>>> riak-users at lists.basho.com <mailto:riak-users at lists.basho.com>
>>> /* Sincerely
>>> Pedram Nimreezi - Chief Technology Officer */
>>> // The hardest part of design … is keeping features out. - Donald Norman
>> riak-users mailing list
>> riak-users at lists.basho.com <mailto:riak-users at lists.basho.com>
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the riak-users