hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Lars George <lars.geo...@gmail.com>
Subject Re: Replication terminology "master-master"
Date Fri, 28 Nov 2014 07:20:07 GMT
Since we already use this terminology inside HBase, wouldn't "peer to peer" be the right term?

Lars 

Sent from my iPhone

> On 28 Nov 2014, at 02:16, Misty Stanley-Jones <mstanleyjones@cloudera.com> wrote:
> 
> I think of multi-master more in the case of load-balancing, where multiple
> machines are all fulfilling the role of master at the same time on the same
> cluster. I can't picture how it would apply to multiple clusters. I'm not
> sure, maybe i'm not getting it, but multi-master seems like a subtlely
> different thing.
> 
>> On Thu, Nov 27, 2014 at 1:23 AM, Nick Dimiduk <ndimiduk@gmail.com> wrote:
>> 
>> I guess "multi-master" is the more common term for this kind of
>> replication.
>> 
>>> On Tue, Nov 25, 2014 at 11:05 AM, Demai Ni <nidmgg@gmail.com> wrote:
>>> 
>>> Agree with you guys. I also experienced difficulty to explain about
>>> Master-Master and  cyclical to my users before, until I began to talk
>> about
>>> M-S only as building-block, which can be applied as one please.
>>> 
>>> Just like to chime in another point, maybe the document should emphasize
>>> that the Master or Slave term is from *the perspective of
>>> 'table:columnFamly'*. It is often confused when using term as
>>> 'Master-cluster' or 'Slave-Cluster' as if the cluster as a whole is a
>>> Master/Slave. Unless the clusters are setup strictly as one-way
>>> replication. the Cluster often serves for both roles.
>>> 
>>> Demai
>>> 
>>> On Tue, Nov 25, 2014 at 9:21 AM, Andrew Purtell <
>> andrew.purtell@gmail.com>
>>> wrote:
>>> 
>>>> 
>>>>> On Nov 25, 2014, at 11:09 AM, Jean-Daniel Cryans <
>> jdcryans@apache.org>
>>>> wrote:
>>>>> 
>>>>> "A master can be
>>>>> slave too, and a slave a master, and cycles are allowed" since with
>>> just
>>>>> that you can build everything else.
>>>> 
>>>> +1
>>>> 
>>>> 
>>>> 
>>>>> On Nov 25, 2014, at 11:09 AM, Jean-Daniel Cryans <
>> jdcryans@apache.org>
>>>> wrote:
>>>>> 
>>>>>> On Mon, Nov 24, 2014 at 8:29 PM, lars hofhansl <larsh@apache.org>
>>>> wrote:
>>>>>> 
>>>>>> Hmm... From that angle all we really have is master-slave
>> replication,
>>>>>> without limit as how it can be setup.A master can be slave too, and
>> a
>>>> slave
>>>>>> a master, and cycles are allowed.
>>>>>> I find "cyclical replication" more confusing than "master-master"
>>>>>> (although it's true that master-master is strictly a subset of
>>> cyclical
>>>> -
>>>>>> just a cycle of two).
>>>>> 
>>>>> I agree with that.
>>>>> 
>>>>> 
>>>>>> We might want to scrap all the terms and just state (hopefully a
bit
>>>>>> nicer) what I tried to say in the first two sentences above.
>>>>> 
>>>>> One thing about "master-master" is that some audiences will get it
>>> right
>>>>> away, for example folks coming from the MySQL world, but maybe we
>> want
>>> to
>>>>> skip using it in our doc and just say what you wrote above "A master
>>> can
>>>> be
>>>>> slave too, and a slave a master, and cycles are allowed" since with
>>> just
>>>>> that you can build everything else.
>>>>> 
>>>>> 
>>>>>> -- Lars
>>>>>> 
>>>>>>     From: Misty Stanley-Jones <mstanleyjones@cloudera.com>
>>>>>> To: dev@hbase.apache.org
>>>>>> Sent: Monday, November 24, 2014 7:26 PM
>>>>>> Subject: Replication terminology "master-master"
>>>>>> 
>>>>>> Hi all,
>>>>>> 
>>>>>> I think "master-master" is a term that should be re-thought. It is
>> not
>>>>>> really a "type" of replication, but refers to a characteristic of
 a
>>>>>> cluster, specifically a cluster which participates in multiple
>>> clusters
>>>>>> with different roles -- it is a slave in one cluster and a master
in
>>>>>> another cluster. I think with the current terminology, people
>> confuse
>>> it
>>>>>> with "cyclical" replication, in which two clusters replicate to each
>>>> other,
>>>>>> and eventually each has all the data from both.
>>>>>> 
>>>>>> Since master-master in this sense is really not a type of
>>> replication, I
>>>>>> think we should just scrap it. You can have master-slave replication
>>> or
>>>>>> cyclical replication, or a combination. With master-slave
>>> replication, a
>>>>>> cluster can fulfill both roles at the same time, as long as it is
in
>>>>>> different clusters. This is easy to understand as a sort of
>> recursive
>>>>>> cascade.
>>>>>> 
>>>>>> Am I explaining it right, and what do you guys think about changing
>>> our
>>>>>> terminology?
>>>>>> 
>>>>>> Thanks,
>>>>>> Misty
>> 

Mime
View raw message