Hi Michael,
> Does that mean you no longer see the original problem (changes not
> being reflected)?
Yes. The deleted documents do not appear in search results any more. I am
not sure that if they are flushed to disk
at that time yet but at least there is a sign that they are "deleted". I
have stopped and started the servlet engine to ensure
that deleted document is no longer there. I think that Lucene requires the
previously opened IndexReader be closed before changes
can be reflected.
> You get the entrySet from the Map, you then iterate over its
> Map.Entry, then you replace in your original map some entries (the
> ones that are opened). So, you are modifying a Java collection while
> iterating over elements from its Set view... I just don't know if
> that's safe (anyone?)
I am a bit skeptical about my approach. Because the IndexSearchers can be
used by other threads (requests) at the same time.
So when I close them, some users can be affected. I will find a better way
to do it.
Also, because reload() is synchronized so there is a single thread accessing
it only. So I think that there will be no ConcurrentModificationException
> Would be good to instrument/debug and confirm
> that the precise IndexReader that's searching the Directory your
> IndexWriter just committed to, is in fact reopened.
Do you think that these code are enough
IndexReader oldReader = searcher.getIndexReader();
IndexReader newReader = oldReader.reopen(true);
if (newReader != oldReader) {
oldReader.close();
searcher.close();
searchers.put(entry.getKey(), new IndexSearcher(newReader));
}
Thanks,
Dinh
On Tue, Nov 3, 2009 at 4:50 PM, Michael McCandless <
[email protected]> wrote:
> On Tue, Nov 3, 2009 at 4:24 AM, Dinh <[email protected]> wrote:
> > Hi Michael,
> >
> > Thank a lot for your advice
> >
> >> Can you verify you are in fact reopening the reader that's reading the
> >> same Directory the writer is writing to?
> >
> > Yes. I have a single and configurable index path. So I can not make a
> > mistake here
>
> OK.
>
> >> Also, you are failing to close the old reader after opening a new one.
> >> This shouldn't cause the issue you're seeing, but, will lead
> >> eventually to OOME or file descriptor exhaustion.
> >
> > I have rewritten the method as follows
> >
> > /**
> > * Reloads searchers after index is changed (added, deleted or
> updated).
> > */
> > public static synchronized void reload() {
> >
> > Set<Map.Entry<String, IndexSearcher>> set = searchers.entrySet();
> >
> > for (Map.Entry<String, IndexSearcher> entry : set) {
> > try {
> > IndexSearcher searcher = entry.getValue();
> > IndexReader oldReader = searcher.getIndexReader();
> > IndexReader newReader = oldReader.reopen(true);
> >
> > if (newReader != oldReader) {
> > oldReader.close();
> > searcher.close();
> > searchers.put(entry.getKey(), new
> > IndexSearcher(newReader));
> > }
> > } catch (Exception e) {
> > log.warn(e.getMessage(), e);
> > }
> > }
> > }
>
> Your reload method looks better now! (You are now closing the old reader).
>
> > And it works now.
>
> Does that mean you no longer see the original problem (changes not
> being reflected)?
>
> >> Finally, are you sure the iteration over the Map entries, that
> >> overwrites each entry, is safe?
> >
> > Do you think that my iteration is safe now? At least I have closed the
> > previous searcher and oldReader before creating new ones. However, I
> don't
> > know if it is a good practice to do so.
>
> You get the entrySet from the Map, you then iterate over its
> Map.Entry, then you replace in your original map some entries (the
> ones that are opened). So, you are modifying a Java collection while
> iterating over elements from its Set view... I just don't know if
> that's safe (anyone?). Would be good to instrument/debug and confirm
> that the precise IndexReader that's searching the Directory your
> IndexWriter just committed to, is in fact reopened.
>
> Mike
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: [email protected]
> For additional commands, e-mail: [email protected]
>
>
