[infinispan-dev] eager locking and async replication

[Mircea Markus]

Let me detail on this a little more: the optimisation refers to cache.lock() not to perform remote locks on ALL data owners, but only on the main data owner.
This way , if session affinity is used for enforcing key locality then cache.lock() would only acquire lock within the same JVM - i.e. very good performance without loosing eager's locking semantics. If the cluster changes, and the key is rehashed on a different node, than eager locking would do an RPC - but for many clusters the topology changes are infrequent. 
Another problem that raises is that of consistency during node failures: if K is on node A and it was locked by a tx originated on node B. If A fails then we can invalidate the transaction on B, so that it would rollback.
Another interesting race condition Sanne raised is with re-hashing: "it needs to be atomic to know who is the owner and aquire the lock, or the owner might be moved and you're locking on the wrong node (only)"
I think this is not related to this optimisation in particular, but stands for eager locking in general - any idea how this is handled btw? 
Cheers,
Mircea

On 20 Aug 2010, at 11:04, Manik Surtani wrote:

> Not sure I understand - are you proposing that the RPC for LockControlCommands are always async if the keys are generated using the key affinity service?
> 
> On 19 Aug 2010, at 16:29, Mircea Markus wrote:
> 
>> Hi,
>> 
>> Here is the scenario,  N=numOwners
>> 
>> TM.start()
>> c.lock(a); //this makes N (might me less) RPCs to acquire locks  
>> c.get(a)
>> ...
>> TM.commit(); // this would do an N calls for prepare/commit. Might happen async.
>> 
>> By using the key affinity service, one might enforce a tx to operate on "local" keys (i.e. keys that are hashed on the same node where the tx was started). 
>> Now, if we would be able to *only* eager lock the main data owner (v.s. N RPCs for lock acquisition  locks) than eager locking would be as fast as the non-eager locking for this scenario. 
>> What happens with the TX if the data owner crashes and only one copy is locked? We would need to invalidate the transaction at originator's side, which I think is possible.
>> For async repl with N >= 2 and key affinity the performance benefit for eager locking would be close to local puts: which is huge. 
>> 
>> This use case was brought by Erik ( cc)- please add you comments if something is missing. What do you think about this optimisation?
>> 
>> Cheers,
>> Mircea 
>> _______________________________________________
>> infinispan-dev mailing list
>> infinispan-dev at lists.jboss.org
>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
> 
> --
> Manik Surtani
> manik at jboss.org
> Lead, Infinispan
> Lead, JBoss Cache
> http://www.infinispan.org
> http://www.jbosscache.org
> 
> 
> 
> 
> 
> _______________________________________________
> infinispan-dev mailing list
> infinispan-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/infinispan-dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.jboss.org/pipermail/infinispan-dev/attachments/20100820/b13dee89/attachment.html