about ctdb data replication

[ronnie sahlberg]

No CTDB does not need any shared storage at all.

However,
Often you will need a tiny shared storage filesystem so that you can
put a 'recovery lock file' on it.
This file is 0 bytes in size  but is used to prevent split brain if
the cluster becomes partitioned.
Unless your filesystem support coherent fcntl() locking, i.e. unless
you use GPFS,   you probably have to disable that
by setting the reclock file to "".

As far as i know,  only GPFS has correct clusterwide fcntl() locking.
Please correct me if i am wrong.


It would be nice to change this to instead use Persistent Reservations
on a dedicated SCSI devices for this.
I think in all places where CTDB+Samba goes  it should be quite easy
to find a dedicated LUN that could be sued for that.

regards
ronnie sahlberg


On Thu, Mar 14, 2013 at 4:55 PM, Andrew Bartlett <abartlet at samba.org> wrote:
> On Thu, 2013-03-14 at 21:59 +1100, Amitay Isaacs wrote:
>> On Thu, Mar 14, 2013 at 8:32 PM, Liujun (A) <liujun09 at huawei.com> wrote:
>>
>> > I know ctdb has a ltdb(local tdb) on each node , is the global db must
>> > store on share filesystem（like gfs）,which all node can write and read from
>> > the global db.
>> >
>>
>> CTDB does not need the databases to be stored in the shared storage. In
>> fact it is designed such that the databases on each node are independent.
>>
>>
>> > Another questions is if the same data are stored in each node for data
>> > recovery?
>> >
>>
>> Recovery process will merge all the databases and then push them to all the
>> nodes. So after recovery all the databases will be exact replica on all the
>> nodes.  Persistent databases are always in sync (that is they are always
>> the same across all the nodes). Non-persistent databases can change
>> independently on all the nodes and do not need to be synchronized.
>
> Does CTDB need shared storage at all, except for the actual data files?
>
> Could you use CTDB for a HA solution that (electronically) unplugged the
> disks from one node, and plugged them into another node, avoiding the
> need for a cluster FS?  (Naturally with logic to ensure that clients
> could not talk to the node without the disk attached)
>
> Andrew Bartlett
>
> --
> Andrew Bartlett                                http://samba.org/~abartlet/
> Authentication Developer, Samba Team           http://samba.org
>
>