IDP Cluster going forward (JDK 7)

[Marc Thornton]

> (When I last checked Infinispan, replication did only take
>place each time you put an object into the map, but not on intermediate
>changes to the object itself, and there was no replication of transient
>fields; the same applies for Ehcache or Hazelcast.)

Based on my reading, EhCache with BigMemory will keep the object consistent
outside of cache management operations, which sounds (at least in theory)
to provide more than Infinispan's solution.  The transient field issue is
one I know very little about in the context of distributed caching or the
Shibboleth IDP, but it sounds like any replication solution based on
(default) Java serialization would have issues adapting to transient
fields.  Is the expectation that transient fields would need to be
replicated as long as a reference to the object exists outside the
memcache? I probably need to dig into this more myself to understand why
it's an issue and the role of the transient fields in the Shibboleth IDP.

>I think I could easily adopt our Memcache plugin to support Infinispan,
>Ehcache, Hazelcast or even JDBC or what you want, but due to the lack of
>transient field persistence, it would never reach the same compatibility
>as the Terracotta way.

Understanding the "compatibility" issues or trade-off's, as in what will or
won't work in the distributed environment might become key in our case.  We
have had some pressure to upgrade our solution to JDK 7.  The better we
understand issues we might face in extending the storage service, the
better recommendation (ie "it can be done" or "it shouldn't be done") we
can provide if JDK 7 becomes a more solid push by our client.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://shibboleth.net/pipermail/users/attachments/20130529/eb7dfabf/attachment.html