On 19 Feb 2010, at 15:21, Marcel Reutegger wrote: > On Thu, Feb 18, 2010 at 13:55, Jukka Zitting <[email protected]> wrote: >> Agreed. Ideally (not sure if that's feasible) we'd push all caching >> down below the unified persistence layer. > > IMO that's exactly where a distributed cache belongs to. > > the current design already implements some aspects but lacks other > important things. currently each cluster node has it's own bundlecache > with invalidation triggered on external changes. but filling the cache > is always done through shared storage. I think this can become a > bottleneck when the number of cluster nodes is more than just a few. > e.g. consider that each of the cluster nodes has event listener > registered that are interested in current changes. this will cause a > rush on the shared storage whenever a change is done. ideally the > cluster nodes (or more specifically the micro-kernels) would share > their cached items with other cluster nodes. shared storage is only > accessed when none of the cluster nodes can provide the request > version of an item.
Agreed, Although distributed caches appear like a fix all solution replicated, and to some extents invalidated caches always seem to choke on traffic levels when a cache really starts to matter. Having used ehcache (which is good as a cache impl with jmx) reasonably extensively some observations. Under the Cache layer (as opposed to a layer that knows something about the app) Replication is best totally avoided as a way of making a cluster scale to more than a few nodes. Invalidation can be effective, provide its not used to trigger reload and the invalidation traffic isn't massive. Over in Shindig, there is a simple Cache API with memcached and ehcache default implementations. Most deployers tune those or replace with their own local implementations, putting cluster wide code below the Cache API and above the cache implementation. Putting the cache below the persistence api, or using a cache API would, IMHO address most of these issue. BTW, JSR-107 looks rather dead, and I found it lacking is certain areas. Ian > > regards > marcel
