[infinispan-dev] StressTest: why removals are not working as expected

[Galder Zamarreno]

I've replicated this in 
https://svn.jboss.org/repos/infinispan/trunk/core/src/test/java/org/infinispan/distribution/DisabledL1Test.java

When we call getFirstNonOwner("k1").put("k1", "value"); apart from 
replicating k1 to the owners of k1, it also resides in the non owner. 
So, when we call getOwners("k1")[0].remove("k1");, it will only remove 
it from those nodes where k1 is owner. The end result is that a 
non-owner ends up with k1.

The fix for this does not seem trivial: You either, avoid non-owners 
storing data in the cache if DIST and !L1, or when !L1, do a cluster 
wide removal.

The first option looks to be more correct one. If I don't have an L1 and 
I'm not the owner of k1, why should I keep it at all? Maybe this could 
be implemented via a Flag, i.e. Flag.SKIP_CACHE (you'd probably need to 
use Flag.SKIP_CACHE_STORE too to avoid putting in the cache store).

With the second option, you'd have to start figuring out which other ops 
would need to be executed cluster wide regardless of ownership when L1 
is disabled. More prone to errors IMO.

Thoughts?

On 10/29/2009 11:41 AM, Galder Zamarreno wrote:
> Bela,
>
> https://jira.jboss.org/jira/browse/ISPN-240 explains why the removals
> are not working as expected in the stress test you wrote.
>
> Cheers,

-- 
Galder Zamarreño
Sr. Software Engineer
Infinispan, JBoss Cache