lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Michael McCandless <luc...@mikemccandless.com>
Subject Re: Parallel Index Search
Date Mon, 16 Oct 2006 14:45:44 GMT
Supriya Kumar Shyamal wrote:

>>> If I am not mistaken the process of locking the Index by different 
>>> objects like IndexReader or Indexwriter, theoratically only one 
>>> Thread can access the index at a time.
>>
>> Actually, only one writer can write to the index at once.  Multiple
>> readers can read from the index.
>>
>> On top of that, multiple threads may share one writer and one reader.
>> Ie, these classes are thread safe.
>>
>>> When we do search on the index it creates a commit lock so the other 
>>> thread does not modify the index, so other thread waits until the 
>>> previosu therads release the lock, is it right?
>>>
>>> So in this case I should say index accessed one by one not parallel?
>>
>> The commit lock is only held while a reader is loading the index and
>> while a writer is "committing" its changes to the index.  These times
>> should be brief.  Whereas, the write lock is held for the entire time
>> that a writer is open.
> But IndexSearcher open index using INdexReader right?

Yes, IndexSearcher will open an IndexReader.  So every time you open
it, the commit lock is acquired, the IndexReader reads all segments
details from the index, and then the commit lock is released.  But,
this lock needs to be shared across all your machines, so that the
writer is prevented from committing while a reader is reading the
index.

Aside: note that you should not be re-opening an IndexSearcher with
every search.  It's much more efficient to open it (and reopen it
every so often) and then share that one instance across many searches.

>>
>>> Its just my speculation, please don't get me wrong.
>>>
>>> Because I try to share the same index by 6 instances and since the 
>>> lock for 5 instances are disabled and only once instance can modify 
>>> the index, at this case I achive the parallel read of the index. Only 
>>> disadvantage is that when the index modified then I get 
>>> FileNotFoundException, ao I do some kind of  respawn the search again.
>>
>> You should not have to disable locking to do this sharing; in fact,
>> disabling locking will lead to the "FileNotFoundException" on
>> instantiating a reader when a writer is committing.
>>
>>> If I implement the lock mechanism in the DB using the custom locking 
>>> then I am afraid the index performance will be reduced but the only 
>>> advantage is that I can avoid FNFE.
>>
>> There are also open issues at least over NFS eg:
>>
>>     http://issues.apache.org/jira/browse/LUCENE-673
>>
>> Whereby even with proper (native) locking it's still possible to
>> hit exceptions due to caching in the NFS client.  Are you using NFS?
>>
> Yes I am using NFS and for sharing I use the READ-ONLY FS mount so that 
> the search instance cannot do any modification by mistake , otherwise it 
> can corrupt the index.

OK, Lucene currently hits intermittant FileNotFoundException over NFS
(see that issue I listed for the gory details).  Furthermore (worse),
correct locking (native locking) does not in fact fix the issue.

We are working on small changes to how Lucene stores the index in the
filesystem ("lockless commits") that do away with the commit lock
entirely.  This not only enables readers to actually be entirely
read-only, but it also allows Lucene to work properly over NFS.

Unfortunately it's going to be a while before this is actually
available (it's still being developed).  In the mean time, one
workaround is to do something similar to the Solr project (built on
top of Lucene) which takes snapshots of the index at "known safe"
times and then readers read from these snapshots instead of the live
index.

Mike

---------------------------------------------------------------------
To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-user-help@lucene.apache.org


Mime
View raw message