Re: startup of resource groups with one node down

Lars Marowsky-Bree wrote:

>On 2002-05-04T14:35:06,
>   Martin Bene <[email protected]> said:
>
>>When starting ha - services on the cluster with just one node, I do get a stonith reset for the absent node, and the status for the 2nd node goes to "down" as expected.
>>
>>However, I get erros trying to bring the resource groups online:
>>
>>Resource Group (web) Status = Internal Error : Cluster = webc : Owner =  : Error State = Node not available (exclusivity).
>>
I get the same problem, and I've been able to repeat it exactly.  If I 
have 2 nodes, and I bring both of them down, then only bring up the 
first node, resource groups never go online.

>
>Well, yes. The resource groups were already trying to go online before the
>STONITH was sent I think; you should force them offline and online them again.
>
I tried this before, but this was problematic(sometimes worked, 
sometimes didn't).

>
>>This definitely isn't as expected - any idea what I'm doing wrong or what I
>>have to do to avoid this situation? 
>>
>
>Do not do this. ;-) A cluster is based on the assumption that one node
>survives; both nodes crashing is not expected ;-)
>
Well, there has to be a special case when a cluster is first coming up. 
 I mean, if an entire cluster goes down, it is not good to have to bring 
up all nodes online.  What happens if I could only bring up 2 of the 3 
failed nodes back online?

However, if you look at my argument--it is kind of baseless, because you 
get lots of problems when all nodes in your cluster go down.  The main 
problem is, who has the most consistent-up2date copy of the 
cdb(configuration database)?  For ex, we have 3 nodes, and say 1 node 
goes down.  The remaining 2 nodes then later on make changes to the 
cdb(i.e. modified existing resource groups, added new resources, changed 
failover policy, changed cluster interconnect ip address, etc).  Now say 
if these 2 nodes fail almost simultaneously, which results in no nodes 
currently up. Now lets say the first node comes back up.  I don't know 
if it can form a quorum(<= 50% of nodes), but certainly the first node 
cannot just assume that it has the newest copy of cdb.  We know that the 
other 2 nodes have the up2date cdb.  If the first node is allowed to 
join the quorum, and later the other 2 nodes come up, which copy of the 
cdb is used? Problems....

My question is: do you have to bring up all nodes in the cluster just 
for FailSafe to continue operating normally?  A person maintaining a 
cluster will have to know this...


-- Kashif