[sheepdog-users] Some question from new user

[Vasiliy Tolstov]

2017-11-16 16:00 GMT+03:00 Raymond Burkholder <ray at oneunified.net>:
> If you don't mind some questions:
>
> On 11/16/2017 03:42 AM, Vasiliy Tolstov wrote:
>>
>> 2017-11-15 22:49 GMT+03:00 Gandalf Corvotempesta
>> <gandalf.corvotempesta at gmail.com>:
>>>
>>> 1) is sheepdog stable and ready for production use?
>>
>>
>> It contains rough edges firstly with cluster recovery when nodes up/down.
>
>
> Even with the rough edges, do you think it is good enough to run?  You are
> still using it?

I'm not using it. I'm test sheepdog, ceph and scst. When i'm poweroff
cycle sheepdog node (1 node with 64 disks) sheepdog started and write
to error log
do_epoch_log_read(178) invalid epoch 1 log, also my tests (i'm use
infiniband interconnect) displays that scst based solution much faster
then sheepdog or ceph


>
>>
>>> 2) do I need any metadata servers (like Ceph) or the only needed
>>> servers are the storage ones (like Gluster) ?
>>
>>
>> No metadata needed, but you need cluster manager (zookeeper, corosync).
>> I'm try corosync and it not very stable (not sheepdog, but corosync)
>> Sometimes i have split bran when cluster
>> divided on parts and only restart corosync solved this issue. (i'm try
>> corosync 2.x)
>
>
> Do you lose data when you encounter these situations or are you able to
> recover successfully and return to normal operation?
>


Data is safe in this case, but you need manual steps to fixup cluster.


>>
>>> 3) can I use multiple disks in a server, or each server should
>>> "expose" a single filesystem (thus, I have to use RAID)
>>
>>
>> Sheepdog supports multidisk but you need metastore to save epoch and
>> config data so raid is preferrable.
>
>
> So, just to confirm, you recommend a raid system of some sort on each host,
> in addition to the replication across hosts which Sheepdog supplies?  Is
> there some reference information I could read about the metastore, epoch,
> and config data?
>

https://github.com/sheepdog/sheepdog/wiki/Multi-disk-on-Single-Node-Support

>>
>>> 7) any "scrub" feature to be sure that all object a replicated
>>> properly and not subject to "bit-rot" ?
>>
>>
>> No, but sheepdog have cluster check (in my case works sometimes with
>> errors)
>
>
> What sort of errors do you see?  Is there something to fix or do you ignore
> them?
>
>>

I don't know can you ignore it or not, but check sometimes displays
that i don't have some objects.


-- 
Vasiliy Tolstov,
e-mail: v.tolstov at selfip.ru