So is IndexWriter::optimize() non-blocking, even with SerialMergeScheduler?
That might explain our problem in trying to use optimize() to make maxDoc()
match between the two indexes before adding readers to ParallelReader. I see
that we could call optimize(true).
----- Original Message ----
From: Justin <cry...@yahoo.com>
To: java-user@lucene.apache.org
Sent: Thu, June 24, 2010 12:12:57 PM
Subject: Re: Problems with homebrew ParallelWriter
Hi Shai,
> Is it synchronized
public synchronized void addDocument(Document document)
throws CorruptIndexException, IOException {
Document document2 = new Document();
document2.add(...);
writer1.addDocument(document);
writer2.addDocument(document2);
}
> did you encounter any exceptions
I haven't seen the machines firsthand, but I assume my colleague looked for
obvious exceptions that would lead to an imbalance. All exceptions appear to
be logged, so we would see something.
> merges could happen on some slices not when you intended
public synchronized ParallelReader getParallelReader()
throws IOException, CorruptIndexException {
IndexReader reader1 = writer1.getReader();
IndexReader reader2 = writer2.getReader();
if (reader1.maxDoc() != reader2.maxDoc()) {
reader1.close();
reader2.close();
writer1.optimize(); // force merge for consistent maxDoc
writer2.optimize(); // force merge for consistent maxDoc
reader1 = writer1.getReader();
reader2 = writer2.getReader();
}
ParallelReader reader = new ParallelReader();
reader.add(reader1);
reader.add(reader2);
return reader;
}
As you can see above, my colleague optimizes the indexes to account for merges
that have occurred out-of-sync.
> if you've made progress, upload another patch?
If we make a revelation with regards to ParallelWriter, I'll be happy to share.
Thanks for giving us some places to look.
Justin
----- Original Message ----
From: Shai Erera <ser...@gmail.com>
To: java-user@lucene.apache.org
Sent: Wed, June 23, 2010 10:48:22 PM
Subject: Re: Problems with homebrew ParallelWriter
How do you add documents to the index? Is it synchronized (such that
basically only one thread can add documents at a time)?
The same goes for removing documents as well.
Also, did you encounter any exceptions during the run - if say an addDoc
fails on one of the slices, then you need to revert that addDoc in all
previous slices ...
I remember running into such exception when working on the Parallel Index
stuff, but I don't remember what caused it ...
About merging, note that if you use LogDocMP, then you can guarantee that
all slices will be in sync, but still some merges could happen on some
slices not when you intended them to happen. For example, during a flush of
one addDoc on one of the slices, before the others addDoc finished. But if
you didn't see any exceptions and didn't terminate the process mid-action,
then this should not happen ...
I hope this helps. Unfortunately I had to shift focus from LUCENE-1879.
Perhaps I'll get back to it one day. But if you advanced on PI somehow,
perhaps you can diff the patch that's there and your code, and if you've
made progress, upload another patch?
Shai
On Thu, Jun 24, 2010 at 1:44 AM, Justin <cry...@yahoo.com> wrote:
> Hi all,
>
> We've been waiting for LUCENE-1879 and LUCENE-2425 and have written our own
> ParallelWriter class in the meantime. Apparently our indexes are falling
> out of sync (I suspect my colleague is seeing error messages come from
> ParallelReader stating the the number of documents must be the same).
>
> Here's a code snippet from our ParallelWriter which extends Object:
>
> writer1 = new IndexWriter(dir, analyzer,
> create,
>
> new IndexWriter.MaxFieldLength(MFL));
>
> writer1.setMergePolicy(new LogDocMergePolicy());
>
> writer1.setMergeScheduler(new SerialMergeScheduler());
>
> writer1.setMaxBufferedDocs(MBD);
>
> writer1.setRAMBufferSizeMB(IndexWriter.DISABLE_AUTO_FLUSH);
>
> My colleague suspects that merging or flushing is being triggered on
> something other than the doc count which leads to the writers' different
> behaviors. I suspect our next step is to scatter breakpoints around Lucene
> source (we've got tr...@926791 to take advantage of latest NRT readers).
>
> Does anyone have ideas on how the indexes would get out of sync? Process
> close, committing, optimizing,... they all should work okay?
>
> Thanks,
> Justin
>
>
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org
> For additional commands, e-mail: java-user-h...@lucene.apache.org
>
>
---------------------------------------------------------------------
To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org
For additional commands, e-mail: java-user-h...@lucene.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org
For additional commands, e-mail: java-user-h...@lucene.apache.org
