[infinispan-dev] Flags ignored during batch operations

[Manik Surtani]

On 7 Jul 2010, at 23:20, Sanne Grinovero wrote:

> Hello,
> I was quite surprised to hit a timeout on lock acquire on a method
> flagged with SKIP_LOCKING.

This is almost certainly a bug.

> Debugging I've found that in case a batch operation is running and no
> transaction is active org.infinispan.interceptors.BatchingInterceptor
> creates a new InvocationContext, ignoring the existing one, and so the
> invocation proceeds but the flags I defined are cleared.

Hmm, it should only create a new InvocationContext if there *is* an active transaction (not if there is *not* one).  This is because batching is implemented using transactions, and to prevent interference with an existing transaction, a new context is used to isolate the batch from the tx.

> What is the reason to create a new InvocationContext? Could that be
> changed to at least carry over the defined flags?

Carrying over flags is the generic fix, since it covers both the case of a tx being present and not.  So this should be done anyway, but we should also look at why a new context is created if there is no ongoing tx.

> I've been playing with it by using setFlags(existingFlags) and have
> already experienced another little performance boost with Lucene,
> especially as it's now respecting SKIP_REMOTE_LOOKUP too, but I'm not
> going to change this without understanding.

Go ahead and fix by carrying over flags.  Like I said this should have happened anyway.

Cheers
Manik
--
Manik Surtani
manik at jboss.org
Lead, Infinispan
Lead, JBoss Cache
http://www.infinispan.org
http://www.jbosscache.org