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: NFS and Lucene 2.0 status - still troublesome ?
Date Mon, 13 Nov 2006 17:02:34 GMT

The quick answer is: NFS is still problematic in Lucene 2.0.

The longer answer is: we'd like to fix this, but it's not fully fixed
yet.  You can see here:

     http://issues.apache.org/jira/browse/LUCENE-673

for gory details.

There are at least two different problems with NFS (spelled out in the
above issue):

   * Intermittant IOException on instantiating a reader.

     This is in fact [surprisingly] not due to locking, at least in my
     testing.  The unreleased version of Lucene now supports native
     locks (through java.nio.*), but even when using native locks I can
     still reproduce this error in my testing.

     The good news is: the lockless commits patch (which is not yet
     committed but I think close):

         http://issues.apache.org/jira/browse/LUCENE-701

     resolves this issue.  Lockless commits also makes readers entirely
     read only, so your read-only NFS mount for readers becomes
     possible.

   * "Stale NFS handle" IOException when searching.

     Lucene's readers provide "point in time" searching: once open,
     a reader searches the snapshot of the index as of the point it was
     open.

     Unfortunately, the implementation of this feature currently relies
     on the filesystem to provide access to files even after they are
     deleted.  NFS makes no such guarantee.

     This means on searching you have to catch this exception and then
     close & open a new searcher.

     I think it would make sense to change how Lucene implements point
     in time searching so we don't rely on filesystem semantics.  But
     this is a ways off.

I'm hopeful that with lockless commits, and then with the caveat of
closing/opening your searchers on hitting "Stale NFS handle" during
searching (until we can change how "point in time" searching is
implemented), that Lucene will work fine over NFS.

Anyway, in the meantime, one good workaround is to either use Solr:

     http://incubator.apache.org/solr/

directly, or, borrow its approach.  With Solr, a writer writes to the
index and periodically (at a known safe time) takes a snapshot, and
then readers only read from the current snapshot.

Mike

Peter A. Friend wrote:
> 
> On Nov 13, 2006, at 8:10 AM, √ėyvind Stegard wrote:
> 
>> I've searched the list and have found many references to problems when
>> using Lucene over NFS. Mostly because of file-based locking, which
>> doesn't work all that well for many NFS installations. I'm under the
>> impression that the core locking logic between writers and/or readers
>> hasn't changed in a significant way between Lucene 1.4 and 2.0 (?). I
>> guess this means NFS is still problematic ?
> 
> Unfortunately it all depends on the reliability of the NFS drivers in 
> the OS, and the kind of filers you are using. If the environment isn't 
> too busy, NFS lockd *may* work on some systems, but it usually ends up 
> collapsing under load.
> 
>  From there you have to hand craft some C code to create lock files, and 
> what works again depends on your system. On some systems doing an 
> exclusive create will work (can only be expected to work on version 3 
> mounts), but then local caches will bite you, so you end up having to 
> disable the directory cache, assuming your system supports such an 
> option. That failing, creating locks as symlinks to unique temporary 
> files that don't exist will usually blow through the cache and work ok. 
> This of course doesn't rule out problems in the NFS implementation that 
> show up under heavy load, and allow more than one machine to think it 
> has the lock. You also have to include some code to sensibly expire 
> locks left from crashes.
> 
>> We are considering a model where a single node updates the search index
>> according to changes in the repository (only one physical index for the
>> entire cluster) while multiple other nodes can search the very same
>> index over NFS (read-only). But I guess there is a need for a single
>> lock-directory shared and writable between all nodes, and that this
>> makes NFS-usage difficult ?
> 
> The fact that only a single node will be doing writes greatly improves 
> the chances of this working. I don't recall whether readers ever check 
> for locks, it's best if that can be avoided. I know that it's safe to 
> write the new indexes since they aren't being referred to by the 
> segments file, but I'm not sure what sequence of operations are used 
> when re-writing the segments file. I think unlinking the old segments 
> file and using a rename to put the new one in place is probably the 
> safest bet.
> 
> Peter
> 
> 
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
> For additional commands, e-mail: java-user-help@lucene.apache.org
> 


---------------------------------------------------------------------
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