Memcached Storage Service Code Review

[Marvin S. Addison]

I have thoughtfully reviewed the memcached implementation of 
StorageService that Scott pointed out:

https://wiki.shibboleth.net/confluence/display/SHIB2/Memcached+StorageService

The component does a fair bit of wrangling to provide special handling 
for certain types (e.g. login context, session, some special case 
objects that are not entirely serializable). That special handling looks 
overly complicated to me, but I trust it works. The use of a plain Map 
as a dual store was surprising, but I believe it's required as a simple 
solution to the problem of capturing all state changes. (In lieu of 
Terracotta's complex bytecode instrumentation that hooks setters/fields 
deep inside an object graph.) In any case the use of a JVM-local data 
structure limits the functionality of this solution, but the limitations 
and requirements for proper use are clearly documented.

The use of a servlet filter to push session state to memcached on the 
output side of the request processing pipeline was clever. Again this 
appears to be needed since IdPv2 components don't explicitly persist 
changes on every mutation, so the session is stored wholesale at the end 
of the request processing pipeline just to be safe.

The only real mechanical thing that jumped out at me was a spin lock 
that had no failsafe condition:

// Try to acquire a lock
while(!client.add("lock:"+name, 30, "").get().booleanValue()) {
   log.trace("LOCKED... trying again: {}", name);
   Thread.sleep(100);
}

Overall this looks like a viable if limited alternative to Terracotta 
for the documented purpose.

M