lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Adrian Tarau <>
Subject Re: Lucene & Transactional semantics
Date Tue, 17 Jun 2008 20:02:47 GMT

I had the same problem, so one year ago I implemented transactions on top of
Lucene(I had an idea how to do it, but I also peeked a little bit in Compass
sources). Basically I create a new index every time when a new transaction
is started and this new index is made visible if commit is successful. Of
course, you will still need to refresh your IndexReader(s). I planned also
for a form of two-phase commit but right now it works pretty well doing
everything manually - controlling when to commit/rollback/restore.

How it works :

1. start 2 transactions, one with MySQL one with Lucene(I have an abstract
layer on top of both engines so the API calls are the same), commit first
2. if Lucene fails, rollback MySQL.
3. if Lucene commit(precommit actually - everything is stored, I have a new
optimized index, ready to be used) is successful, commit MySQL.
4. if MySQL fails, "cancel" Lucene transaction(remove index). There is a
slightly chance to not being able to cancel the Lucene transaction(which is
a new index). This index will hang there on disk/memory until the session is
closed or the application is restarted.
5. if MySQL commit is successful, the newly created Lucene index is
"activated". There is a slightly chance to not being able to "activate"
Lucene transaction(application crashes, disk full so cannot write state on
disk) - so in this case the global transaction is "in-doubt". In this case a
system administrator must act and activate the index(the index is fine, but
not active) - after he fixes any existing problems - or decided to clear the
database(you need eventually a unique global transaction id stored in the
database too, as an additional column). Your data is in the database, the
application works fine but some entries are "invisible" when searching.

So is kind of a two-phase commit. You can read more about two-phase commit
here : or
just google about "two-phase commit". 


Beto Siless wrote:
> Hi, I'm with the transaction problem too: I have Documents which are 
> represented by a Business Object (persisted in a DB with an ORM), 
> indexed with Lucene and finally stored in the file system. So it's very 
> difficult to maintain the consistency in an error scenario.
> The main problem is that if you implement some ad-hoc transaction with 
> Lucene (working in a RAMDirectory or keeping the commands to apply until 
> the end), you still have to coordinate the lucene transaction with the 
> others. Cause if lucene transaction rollbacks you can abort the db 
> transaction, but if lucene transaction commits you can't do anything if 
> the DB transaction fails with out a 3pc transaction manager.
> Does Anybody have an idea about how to reduce the error time window? 
> Could this problem be solved storing the index in a database?
> Thanks
> Beto
> Marios Skounakis wrote:
>> Hi all,
>> I am interested in developing a system which will use Lucene to implement
>> the search functionality. A key characteristic of this system is that
>> certain information about the indexed documents will be editable by the
>> user administrators. For instance, the user administrators can manually
>> create "document collections" and assign some of the indexed documents to
>> them. One way to implement document collections would by having documents
>> have a dedicated field for storing the document collection id, and
>> storing the document collection information in a database.
>> Ideally, such an operation as the above should have transactional
>> semantics, i.e. if a user wants to assign documents x, y and z to
>> collection C, then either all three documents should be assigned to the
>> collection or, in case of error, none of the documents should be assigned
>> to the collection. Also, if the operation were to be followed by an SQL
>> query to update the database with the number of documents assigned to
>> collection C, that should be included in the "transaction" as well.
>> Is there a straightforward way to do this with Lucene? Or are
>> "transactions" a no-no for a system like Lucene and I should just go
>> ahead without having transactional semantics?
>> Thanks in advance,
>> Marios Skounakis
>> ------------------------------------------------------------------------
>> No virus found in this incoming message.
>> Checked by AVG Free Edition.
>> Version: 7.1.362 / Virus Database: 267.13.1/169 - Release Date:
>> 11/15/2005
> ---------------------------------------------------------------------
> To unsubscribe, e-mail:
> For additional commands, e-mail:

View this message in context:
Sent from the Lucene - Java Users mailing list archive at

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message