[hibernate-dev] hibernate search + indexing with transaction issue

[Yan Falken]

Hi, 

I have the following problem with indexing in transaction with hibernate
search: in case the transaction is opened and I want to perform multiple
work types - only one type is actually performed after the commit and the others are
ignored; example: 

TransactionManager tm = new DummyTransactionManager();
tm.begin();
SEntity se1 = new SEntity(10, "first", "second", true);

Work delete = new Work(se1, "100", WorkType.DELETE);
Work add = new Work(se1, "100", WorkType.ADD);

Ctx ctx = new Ctx(tm.getTransaction());
searchFactory.getWorker().performWork(delete, ctx);
searchFactory.getWorker().performWork(add, ctx);

tm.commit();

only DELETE is performed - ADD is ignored

log: 

2009-03-23 13:45:02,195 TRACE                 main|       MaskedProperty| found a match for key: [hibernate.search.default.indexBase] value: /tmp/c1idx
2009-03-23 13:45:02,195 TRACE                 main|       MaskedProperty| found a match for key: [default.indexBase] value: /tmp/c1idx
2009-03-23 13:45:02,341 DEBUG                 main| BuilderIndexedEntity| Field selection in projections is set to false for entity problems.SEntity.
2009-03-23 13:45:02,502 DEBUG      pool-1-thread-1|  PerDPQueueProcessor| Skipping usage of an IndexWriter for updates
2009-03-23 13:45:02,503 TRACE      pool-1-thread-1|  PerDPQueueProcessor| Locking Workspace (or waiting to...)
2009-03-23 13:45:02,503 TRACE      pool-1-thread-1|  PerDPQueueProcessor| Workspace lock aquired.
// opened
2009-03-23 13:45:02,503 DEBUG      pool-1-thread-1|  PerDPQueueProcessor| Opening an IndexReader for update
2009-03-23 13:45:02,503 TRACE      pool-1-thread-1|            Workspace| IndexReader opened
// only remove 
2009-03-23 13:45:02,503 TRACE      pool-1-thread-1| eleteExtWorkDelegate| Removing class problems.SEntity#100 by id using an IndexReader.
2009-03-23 13:45:02,539 TRACE      pool-1-thread-1|            Workspace| IndexReader closed
2009-03-23 13:45:02,539 TRACE      pool-1-thread-1|  PerDPQueueProcessor| Unlocking Workspace

I digged into the code a little bit and there is class method: 

org.hibernate.search.engine.DocumentBuilderIndexedEntity.addWorkToQueue() which is in it's beginning iterating the existing queue to avoid unecessary duplicated works

if ( workType == WorkType.DELETE ) { //TODO add PURGE?
	//DELETE should have precedence over any update before (HSEARCH-257)
	//if an Add work is here, remove it
	//if an other delete is here remember but still search for Add
	if ( luceneWork instanceof AddLuceneWork ) {
		toDelete.add( luceneWork );
	}
	else if ( luceneWork instanceof DeleteLuceneWork ) {
		duplicateDelete = true;
	}
}
else {
	//we can safely say we are out, the other work is an ADD
	return;
}

I believe that return should be changed to continue. I did several tests and after the change it can handle ADD/DELETE works in one queue. 

Could you confirm this and tell me in case it's correct how fix should be raised? Please ignore this email if its design limitation and the multiple WTs cannot be fixed easily. 

Thanks
YF