[infinispan-dev] Rehashing on view changes

[Bela Ban]

On 11/26/10 2:59 PM, Vladimir Blagojevic wrote:
> On 10-11-26 10:50 AM, Bela Ban wrote:
>> Quick question: when I have a view {A,B,C,D,E,F} at time T and then
>> receive 2 subsequent views within 3 seconds:
>>
>> - {A,B,C}
>> - {A,B,C,E,F}
>>
>> will there be 2 rehashes / rebalancings, or 1 ? Is there a delay that
>> can be configured for rebalancing on view changes ?
>
> 2 rehashes.


OK, good to know. I'll add some code to RELAY which hides the first view 
{A,B,C} and installs {A,B,C,E,F} directly. Unless, of course, D, E and F 
have failed, then {A,B,C} *will* get installed.

The simple solution I have in mind is to wait a little before removing 
all of the remote nodes when D crashes. The wait of course onyl needs to 
happen when we know we have more than 1 node in the remote cluster. This 
is essentially optimistically assuming that E will take over, and won't 
help on a network partition.

Once ISPN-493 is done, I can remove that code again.

> That idea about delay in initiating rehash is excellent! However, we
> have to be careful to include delay only for high view instability. This
> is actually on my plate https://jira.jboss.org/browse/ISPN-493

OK. I think the delay should be configurable (0 turns it off)...

-- 
Bela Ban
Lead JGroups / Clustering Team
JBoss