Eviction and buddy backup regions

[brian.stansberry]

I don't think we handle evicting stuff from the buddy backup region very well. Looking at the code, all nodes under _BUDDY_BACKUP_ will be evicted based on the _default_ region settings, unless the user was really savvy (and anal) and configured parallel eviction regions, one for the main tree and one for each possible buddy.

Logically, if I say I want maxNodes=10000 in the /foo region, I'd want the main tree and all the buddy backup trees factored into the calculation.

Haven't looked at this in too much depth, but imagine this could be handled in the RegionManager.getRegion() impl. Do the search as it is now; if it only finds the _default_ region and the Fqn is under /_BUDDY_BACKUP_/xxxx, repeat the search with the child fqn.

[manik]

Makes sense. Simplifies configuration for reuse across all nodes in the cluster.

I presume this would mean that any *explicit* eviction configuration of /_BUDDY_BACKUP_/xxx should be ignored?

[manik]

JBCACHE-1190

[brian.stansberry]

You could allow explict configuration of /_BUDDY_BACKUP_/xxx. The getRegion impl

1) Tries to find a region using the passed in Fqn /_BUDDY_BACKUP_/xxx/foo.
2) If _default_ is found, check if Fqn is under /_BUDDY_BACKUP_/xxx; if so try to find a region using /foo.

That's overhead for an obscure use case. We could just skip that for now though and disallow explicit configuration. If requested, add it.

[brian.stansberry]

LOL. For an unrelated reason I was thinking about this yesterday and started this thread; today I'm chasing down an intermittent Branch_4_2 testsuite failure with EJB3 SFSB buddy replication and I find this is the cause. :)

[brian.stansberry]

Unfortunately, there's another major problem with eviction and buddy replication. If a node is only available via a cache loader, data gravitation fails. http://jira.jboss.com/jira/browse/JBCACHE-1192

[manik]

Unfortunately, there's another major problem with eviction and buddy replication. If a node is only available via a cache loader, data gravitation fails.

You mean if the node has been evicted in memory on the buddy, but lives on in the buddy's cache loader, the DG call doesn't force the buddy to load from cache loader?

[brian.stansberry]

Correct. CacheImpl.gravitateData() uses findNode() to look for the data; doesn't go through the interceptors.

If the call is to go through the interceptors, and auto-gravitation is turned on, there needs to be a mechanism to suppress gravitation for the call. We have the Option to force gravitation, but not the opposite.

[manik]

Crap. That would be a pretty serious bug then.

[brian.stansberry]

Yes. :( This would need to be resolved for AS 5 and for the next EAP 4.2 CP. JBCACHE-1190 too, if we can, else we have to do a lot of hacks with region configuration that will be hard to back out in later 5.x releases.