[OpenAFS] OpenAFS with RAID

Horst Birthelmer horst@riback.net
Thu, 29 Dec 2005 15:02:01 +0100


On Dec 29, 2005, at 12:28 PM, Christof Hanke wrote:
> Jeffrey Altman wrote:
>> Stephan Wiesand wrote:
>>> Wouldn't it be an option to not take over the IP address, but  
>>> just the
>>> vice partition? Once failure of the peer is recognized and confirmed
>>> (which is a problem, I agree, but not at all AFS-specific):
>>>
>>>  1) stonith
>>>  2) mount the new vice partition and salvage it
>>>     [ 2a) is there a need to restart the fileserver? ]
>>>  3) vos syncvldb
>> You would also require a
>>     4) vos syncserv
>>> There must be some thinko in all this, or people would be doing  
>>> this a
>>> lot. What is it I'm overlooking?
>> I believe this scenario will not work because the VLDB entries for  
>> all
>> of the volumes that are being mounted by Server B are listed as being
>> on Server A.  Since Server A is unreachable, the volume server when
>> performing the "vos syncvldb" and "vos syncserv" steps will not be  
>> able
>> to verify that the volume is no longer on Server A.  Therefore, there
>> is now an "irreconcilable conflict" that will cause the vos  
>> command to
>> "write a message to the standard error stream."  The vos "command  
>> never
>> removes volumes from file server machines."  The quotes are from  
>> the man
>> pages for "vos syncvldb" and "vos sycnserv".
>
> I guess, you con circumvent this, by moving the sysid as well.
> The hotswap may then work as follows:
> all /vicep* and /usr/afs (tranarc paths assumed) are on an external  
> storage.
> If/When fileserver A breaks, you shut it down.
> A standby node B, with the same OS than fileserver A  (which is up  
> to this place _no_ fileserver) mounts the stuff on the external  
> storage. Then you startup the fileserver binary on that.

I did something like this years ago and it worked, but somehow I  
still have a bad feeling about doing it ;-)

It was using FibreChannel devices, which are very easy to 'move'.
It included moving the IP address of the fileserver as well, to avoid  
the syncvldb and syncserv headache.

Horst