Gluster users - Sep 2016 - EC clarification

On 21/09/16 14:36, Gandalf Corvotempesta wrote:
> Il 01 set 2016 10:18 AM, "Xavier Hernandez" <xhernandez at
datalab.es
> <mailto:xhernandez at datalab.es>> ha scritto:
>> If you put more than one fragment into the same server, you will lose
> all the fragments if the server goes down. If there are more than 4
> fragments on that server, the file will be unrecoverable until the
> server is brought up again.
>>
>> Putting more than one fragment into a single server only makes sense
> to account for disk failures, since the protection against server
> failures is lower.
>>
>
> Exactly, what i would like to ensure is that the 4 segments needed for
> recovery are placed automatically on at least 4 servers (and not on 4
> different bricks that could be on the same server)
>
You *must* ensure that *all* bricks forming a single disperse set are 
placed in a different server. There are no 4 special fragments. All 
fragments have the same importance. The way to do that is ordering them 
when the volume is created:

gluster volume create test disperse 16 redundancy 4 
server{1..20}:/bricks/test1 server{1..20}:/bricks/test2 
server{1..20}:/bricks/test3

This way all 20 fragments from each disperse set will be placed in a 
different server. However each server will have 3 bricks and no fragment 
from a single file will be stored in more than one brick of each server.

Xavi

2016-09-21 14:42 GMT+02:00 Xavier Hernandez <xhernandez at
datalab.es>:
> You *must* ensure that *all* bricks forming a single disperse set are
placed
> in a different server. There are no 4 special fragments. All fragments have
> the same importance. The way to do that is ordering them when the volume is
> created:
>
> gluster volume create test disperse 16 redundancy 4
> server{1..20}:/bricks/test1 server{1..20}:/bricks/test2
> server{1..20}:/bricks/test3
>
> This way all 20 fragments from each disperse set will be placed in a
> different server. However each server will have 3 bricks and no fragment
> from a single file will be stored in more than one brick of each server.
Now it's clear.
So, at very minimum, EC is good starting from 7+3 (10 servers with 1
brick each) because: 512*(7-3) = 2048
Any smaller combinations would mean less redundancy (6+2) or not
optimized stripe size like 512*(5-3)=1024

Is this correct ? And what if I have to add some brick to the current
servers or I have to add new servers?
Can I add them freely or I have to follow some rules ?