[infinispan-dev] Feedback from Mobicents Cluster Framework on top of Infinispan 5.0 Alpha1

[Eduardo Martins]

Hi all, just completed first iteration of Mobicents Cluster Framework
2.x, which includes impl using first alpha release of Infinispan 5. We
have everything we had in jboss cache working, it's a dumb down but
higher level framework, with stuff like Fault Tolerant timers, which
is then reuse in whole Mobicents platform to provide cluster related
features. I believe we now already use a lot of stuff from Infinispan
rich feature set, so I guess it's good timing to give some feedback,
report some minor issues, and clear some doubts.

1) Marshallers

Generally I'm "just OK" with the current API, the fact it
Externalizers depends on Annotations makes impossible for a higher
level framework to allow it's clients to plug their own kind of
Externalizers, without exposing Infinispan classes. In concrete, our
framework exposes its own Externalizer model, but to use them we need
to wrap the instances of classes these handle, in classes bound to
Infinispan Externalizers, which when invoked by Infinispan then lookup
our "own" Externalizers, that is, 2 Externalizers lookups per call. If
there was no annotations required we could wrap instead our clients
Externalizers and plug them into Infinispan, this would mean a single
Externalizer lookup. By the way, since the Externalizer now uses
generics the Annotations values look bad, specially since it's
possible to introduce errors that can compile.

Another issue, in this case I consider it of minor importance due to
low level of Externalizers concept, is the lack of inheritance of
Externalizers, for instance we extend DefaultConsistentHash, instead
of extending its Externalizer I had to make an Externalizer from
scratch, and copy of DefaultConsistenHash's Externalizer code. This is
messy to manage, if the code on Infinispan change we will always have
to check if the code we copied is still valid.

2) CacheContainer vs CacheManager

The API is now deprecating the CacheManager, replacing with
CacheContainer, but some functionality is missing, for instance
CacheContainer is not a Listenable, thus no option to register
Listeners, unless unsafe casting of the instance exposed by the Cache
or a ref is available to the original CacheManager. Related matter,
IMHO a Cache listener should be able to listen to CacheManager events,
becoming global listeners.

3) Configuration stuff

Generally I think the Configuration and GlobalConfiguration could be
simplified a bit, I found myself several times looking at the impl
code to understand how to achieve some configurations. Better
documentation wouldn't hurt too, it's great to have a complete
reference, but the configuration samples are not ok, one is basically
empty, the other has all possible stuff, very unrealistic, would be
better to have reusable examples for each mode, with then
recommendations on how to improve these.

Infinispan 5 introduces a new global configuration setter to provide
an instance, with same method name as the one to provide the class
name. I believe one is enough, and to be more friendly with
Microcontainer and similar frameworks I would choose the one to set
the instance.

4) AtomicMap doubt

I read in the Infinispan blog that AtomicMap provides colocation of
all entries, is that idea outdated? If not we may need a way to turn
that off :) For instance would not that mean the Tree API does not
works well with distribution mode? I apologize in advance if I'm
missing something, but if AtomicMap defines colocation, AtomicMap is
good for the node's data map, but not for the node's childs fqns.
Shouldn't each child fqn be freely distributed, being colocated
instead with the related node cache entry and data (atomic)map? Our
impl is kind of an "hybrid" of the Tree API, allows cache entries
references (similar to childs) but no data map, and the storage of
references through AtomicMap in same way as Tree API worries me.
Please clarify.

5) Minor issues found

See these a lot, forgotten info logging?

03:39:06,603 INFO  [TransactionLoggerImpl] Starting transaction logging
03:39:06,623 INFO  [TransactionLoggerImpl] Stopping transaction logging

MBean registration tries twice the same MBean, the second time fails
and prints log (no harm besides that, the process continues without
failures):

03:39:06,395 INFO  [ComponentsJmxRegistration] Could not register
object with name:
org.infinispan:type=Cache,name="___defaultcache(dist_sync)",manager="DefaultCacheManager",component=Cache

6) Final thoughts

Kind of feel bad to provide all these negative stuff in a single mail,
took me an hour to write it, but don't get me wrong, I really enjoy
Infinispan, it's a great improvement. I'm really excited to have it
plugged in AS7 (any plan on this?) and then migrate our platform to
this new cluster framework. I expect a big performance improvement, on
something already pretty fast, and much less memory usage, our
Infinispan impl "feels" very fine grained and optimized in all points
of view. Of course, the distribution mode is the cherry on top of the
cake, hello true scalability.

I hope to find time to contribute back more, and in better ways, like
concrete enhancements or issues with test cases as happened with jboss
cache, but right now that's the best I could. By the way, I'm using
the nick mart1ns in the infinispan freenode irc channel, feel free to
ping me there.

Regards,

-- Eduardo
..............................................
http://emmartins.blogspot.com
http://redhat.com/solutions/telco