SafeThreadLocal is very interesting. It'll be good not only for Lucene, but also other projects.

Could you please post it?

-- 
Chris Lu
-------------------------
Instant Scalable Full-Text Search On Any Database/Application
site: http://www.dbsight.net
demo: http://search.dbsight.com
Lucene Database Search in 3 minutes: http://wiki.dbsight.com/index.php?title=Create_Lucene_Database_Search_in_3_minutes
DBSight customer, a shopping comparison site, (anonymous per request) got 2.6 Million Euro funding!


On Wed, Sep 10, 2008 at 9:41 AM, robert engels <rengels@ix.netcom.com> wrote:
The other thing Lucene can do is create a SafeThreadLocal - it is rather trivial, and have that integrate at a higher-level, allowing for manual clean-up across all threads.

It MIGHT  be a bit slower than the JDK version (since that uses heuristics to clear stale entries), and so doesn't always clear.

But it will be far more deterministic.

If someone is interested I can post the class, but I think it is well within the understanding of the core Lucene developers.


On Sep 10, 2008, at 11:10 AM, robert engels wrote:

You do not need to create a new RAMDirectory - just write to the existing one, and then reopen() the IndexReader using it.

This will prevent lots of big objects being created. This may be the source of your problem.

Even if the Segment is closed, the ThreadLocal will no longer be referenced, but there will still be a reference to the SegmentTermEnum (which will be cleared when the thread dies, or "most likely" when new thread locals on that thread a created, so here is a potential problem.

Thread 1 does a search, creates a thread local that references the RAMDir (A).
Thread 2 does a search, creates a thread local that references the RAMDir (A).

All readers, are closed on RAMDir (A).

A new RAMDir (B) is opened.

There may still be references in the thread local maps to RAMDir A (since no new thread local have been created yet).

So you may get OOM depending on the size of the RAMDir (since you would need room for more than 1).  If you extend this out with lots of threads that don't run very often, you can see how you could easily run out of memory.  "I think" that ThreadLocal should use a ReferenceQueue so stale object slots can be reclaimed as soon as the key is dereferenced - but that is an issue for SUN.

This is why you don't want to create new RAMDirs.

A good rule of thumb - don't keep references to large objects in ThreadLocal (especially indirectly).  If needed, use a "key", and then read the cache using a the "key".
This would be something for the Lucene folks to change.

On Sep 10, 2008, at 10:44 AM, Chris Lu wrote:

I am really want to find out where I am doing wrong, if that's the case.

Yes. I have made certain that I closed all Readers/Searchers, and verified that through memory profiler.

Yes. I am creating new RAMDirectory. But that's the problem. I need to update the content. Sure, if no content update and everything the same, of course no OOM.

Yes. No guarantee of the thread schedule. But that's the problem. If Lucene is using ThreadLocal to cache lots of things by the Thread as the key, and no idea when it'll be released. Of course ThreadLocal is not Lucene's problem...

Chris

On Wed, Sep 10, 2008 at 8:34 AM, robert engels <rengels@ix.netcom.com> wrote:
It is basic Java. Threads are not guaranteed to run on any sort of schedule. If you create lots of large objects in one thread, releasing them in another, there is a good chance you will get an OOM (since the releasing thread may not run before the OOM occurs)...  This is not Lucene specific by any means.

It is a misunderstanding on your part about how GC works.

I assume you must at some point be creating new RAMDirectories - otherwise the memory would never really increase, since the IndexReader/enums/etc are not very large...

When you create a new RAMDirectories, you need to BE CERTAIN !!! that the other IndexReaders/Searchers using the old RAMDirectory are ALL CLOSED, otherwise their memory will still be in use, which leads to your OOM...


On Sep 10, 2008, at 10:16 AM, Chris Lu wrote:

I do not believe I am making any mistake. Actually I just got an email from another user, complaining about the same thing. And I am having the same usage pattern.

After the reader is opened, the RAMDirectory is shared by several objects.
There is one instance of RAMDirectory in the memory, and it is holding lots of memory, which is expected. 

If I close the reader in the same thread that has opened it, the RAMDirectory is gone from the memory.
If I close the reader in other threads, the RAMDirectory is left in the memory, referenced along the tree I draw in the first email.

I do not think the usage is wrong. Period.

-------------------------------------
Hi,

   i found a forum post from you here [1] where you mention that you
have a memory leak using the lucene ram directory. I'd like to ask you
if you already have resolved the problem and how you did it or maybe
you know where i can read about the solution. We are using
RAMDirectory too and figured out, that over time the memory
consumption raises and raises until the system breaks down but only
when we performing much index updates. if we only create the index and
don't do nothing except searching it, it work fine.

maybe you can give me a hint or a link,
greetz,
-------------------------------------

-- 
Chris Lu
-------------------------
Instant Scalable Full-Text Search On Any Database/Application
site: http://www.dbsight.net
demo: http://search.dbsight.com
Lucene Database Search in 3 minutes: http://wiki.dbsight.com/index.php?title=Create_Lucene_Database_Search_in_3_minutes
DBSight customer, a shopping comparison site, (anonymous per request) got 2.6 Million Euro funding!

On Wed, Sep 10, 2008 at 7:12 AM, robert engels <rengels@ix.netcom.com> wrote:
Sorry, but I am fairly certain you are mistaken.

If you only have a single IndexReader, the RAMDirectory will be shared in all cases.

The only memory growth is any buffer space allocated by an IndexInput (used in many places and cached).

Normally the IndexInput created by a RAMDirectory do not have any buffer allocated, since the underlying store is already in memory.

You have some other problem in your code...

On Sep 10, 2008, at 1:10 AM, Chris Lu wrote:

Actually, even I only use one IndexReader, some resources are cached via the ThreadLocal cache, and can not be released unless all threads do the close action.

SegmentTermEnum itself is small, but it holds RAMDirectory along the path, which is big.

-- 
Chris Lu
-------------------------
Instant Scalable Full-Text Search On Any Database/Application
site: http://www.dbsight.net
demo: http://search.dbsight.com
Lucene Database Search in 3 minutes: http://wiki.dbsight.com/index.php?title=Create_Lucene_Database_Search_in_3_minutes
DBSight customer, a shopping comparison site, (anonymous per request) got 2.6 Million Euro funding!

On Tue, Sep 9, 2008 at 10:43 PM, robert engels <rengels@ix.netcom.com> wrote:
You do not need a pool of IndexReaders...

It does not matter what class it is, what matters is the class that ultimately holds the reference.

If the IndexReader is never closed, the SegmentReader(s) is never closed, so the thread local in TermInfosReader is not cleared (because the thread never dies). So you will get one SegmentTermEnum, per thread * per segment.

The SegmentTermEnum is not a large object, so even if you had 100 threads, and 100 segments, for 10k instances, seems hard to believe that is the source of your memory issue.

The SegmentTermEnum is cached by thread since it needs to enumerate the terms, not having a per thread cache, would lead to lots of random access when multiple threads read the index - very slow.

You need to keep in mind, what if every thread was executing a search simultaneously - you would still have 100x100 SegmentTermEnum instances anyway !  The only way to prevent that would be to create and destroy the SegmentTermEnum on each call (opening and seeking to the proper spot) - which would be SLOW SLOW SLOW.

On Sep 10, 2008, at 12:19 AM, Chris Lu wrote:

I have tried to create an IndexReader pool and dynamically create searcher. But the memory leak is the same. It's not related to the Searcher class specifically, but the SegmentTermEnum in TermInfosReader.

-- 
Chris Lu
-------------------------
Instant Scalable Full-Text Search On Any Database/Application
site: http://www.dbsight.net
demo: http://search.dbsight.com
Lucene Database Search in 3 minutes: http://wiki.dbsight.com/index.php?title=Create_Lucene_Database_Search_in_3_minutes
DBSight customer, a shopping comparison site, (anonymous per request) got 2.6 Million Euro funding!

On Tue, Sep 9, 2008 at 10:14 PM, robert engels <rengels@ix.netcom.com> wrote:
A searcher uses an IndexReader - the IndexReader is slow to open, not a Searcher. And searchers can share an IndexReader.

You want to create a single shared (across all threads/users) IndexReader (usually), and create an Searcher as needed and dispose.  It is VERY CHEAP to create the Searcher.

I am fairly certain the javadoc on Searcher is incorrect.  The warning "For performance reasons it is recommended to open only one IndexSearcher and use it for all of your searches" is not true in the case where an IndexReader is passed to the ctor.

Any caching should USUALLY be performed at the IndexReader level.

You are most likely using the "path" ctor, and that is the source of your problems, as multiple IndexReader instances are being created, and thus the memory use.


On Sep 9, 2008, at 11:44 PM, Chris Lu wrote:

On J2EE environment, usually there is a searcher pool with several searchers open.
The speed to opening a large index for every user is not acceptable.

-- 
Chris Lu
-------------------------
Instant Scalable Full-Text Search On Any Database/Application
site: http://www.dbsight.net
demo: http://search.dbsight.com
Lucene Database Search in 3 minutes: http://wiki.dbsight.com/index.php?title=Create_Lucene_Database_Search_in_3_minutes
DBSight customer, a shopping comparison site, (anonymous per request) got 2.6 Million Euro funding!

On Tue, Sep 9, 2008 at 9:03 PM, robert engels <rengels@ix.netcom.com> wrote:
You need to close the searcher within the thread that is using it, in order to have it cleaned up quickly... usually right after you display the page of results.

If you are keeping multiple searcher refs across multiple threads for paging/whatever, you have not coded it correctly.

Imagine 10,000 users - storing a searcher for each one is not going to work...

On Sep 9, 2008, at 10:21 PM, Chris Lu wrote:

Right, in a sense I can not release it from another thread. But that's the problem.

It's a J2EE environment, all threads are kind of equal. It's simply not possible to iterate through all threads to close the searcher, thus releasing the ThreadLocal cache.
Unless Lucene is not recommended for J2EE environment, this has to be fixed.

-- 
Chris Lu
-------------------------
Instant Scalable Full-Text Search On Any Database/Application
site: http://www.dbsight.net
demo: http://search.dbsight.com
Lucene Database Search in 3 minutes: http://wiki.dbsight.com/index.php?title=Create_Lucene_Database_Search_in_3_minutes
DBSight customer, a shopping comparison site, (anonymous per request) got 2.6 Million Euro funding!


On Tue, Sep 9, 2008 at 8:14 PM, robert engels <rengels@ix.netcom.com> wrote:
Your code is not correct. You cannot release it on another thread - the first thread may creating hundreds/thousands of instances before the other thread ever runs...

On Sep 9, 2008, at 10:10 PM, Chris Lu wrote:

If I release it on the thread that's creating the searcher, by setting searcher=null, everything is fine, the memory is released very cleanly.
My load test was to repeatedly create a searcher on a RAMDirectory and release it on another thread. The test will quickly go to OOM after several runs. I set the heap size to be 1024M, and the RAMDirectory is of size 250M. Using some profiling tool, the used size simply stepped up pretty obviously by 250M.

I think we should not rely on something that's a "maybe" behavior, especially for a general purpose library.

Since it's a multi-threaded env, the thread that's creating the entries in the LRU cache may not go away quickly(actually most, if not all, application servers will try to reuse threads), so the LRU cache, which uses thread as the key, can not be released, so the SegmentTermEnum which is in the same class can not be released.

And yes, I close the RAMDirectory, and the fileMap is released. I verified that through the profiler by directly checking the values in the snapshot.

Pretty sure the reference tree wasn't like this using code before this commit, because after close the searcher in another thread, the RAMDirectory totally disappeared from the memory snapshot.

-- 
Chris Lu
-------------------------
Instant Scalable Full-Text Search On Any Database/Application
site: http://www.dbsight.net
demo: http://search.dbsight.com
Lucene Database Search in 3 minutes: http://wiki.dbsight.com/index.php?title=Create_Lucene_Database_Search_in_3_minutes
DBSight customer, a shopping comparison site, (anonymous per request) got 2.6 Million Euro funding!

On Tue, Sep 9, 2008 at 5:03 PM, Michael McCandless <lucene@mikemccandless.com> wrote:

Chris Lu wrote:

The problem should be similar to what's talked about on this discussion.
http://lucene.markmail.org/message/keosgz2c2yjc7qre?q=ThreadLocal

The "rough" conclusion of that thread is that, technically, this isn't a memory leak but rather a "delayed freeing" problem.  Ie, it may take longer, possibly much longer, than you want for the memory to be freed.


There is a memory leak for Lucene search from Lucene-1195.(svn r659602, May23,2008)

This patch brings in a ThreadLocal cache to TermInfosReader.

One thing that confuses me: TermInfosReader was already using a ThreadLocal to cache the SegmentTermEnum instance.  What was added in this commit (for LUCENE-1195) was an LRU cache storing Term -> TermInfo instances.  But it seems like it's the SegmentTermEnum instance that you're tracing below.


It's usually recommended to keep the reader open, and reuse it when
possible. In a common J2EE application, the http requests are usually
handled by different threads. But since the cache is ThreadLocal, the cache
are not really usable by other threads. What's worse, the cache can not be
cleared by another thread!

This leak is not so obvious usually. But my case is using RAMDirectory,
having several hundred megabytes. So one un-released resource is obvious to
me.

Here is the reference tree:
org.apache.lucene.store.RAMDirectory
 |- directory of org.apache.lucene.store.RAMFile
    |- file of org.apache.lucene.store.RAMInputStream
        |- base of org.apache.lucene.index.CompoundFileReader$CSIndexInput
            |- input of org.apache.lucene.index.SegmentTermEnum
                |- value of java.lang.ThreadLocal$ThreadLocalMap$Entry

So you have a RAMDir that has several hundred MB stored in it, that you're done with yet through this path Lucene is keeping it alive?

Did you close the RAMDir?  (which will null its fileMap and should also free your memory).

Also, that reference tree doesn't show the ThreadResources class that was added in that commit -- are you sure this reference tree wasn't before the commit?

Mike

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




--
Chris Lu
-------------------------
Instant Scalable Full-Text Search On Any Database/Application
site: http://www.dbsight.net
demo: http://search.dbsight.com
Lucene Database Search in 3 minutes: http://wiki.dbsight.com/index.php?title=Create_Lucene_Database_Search_in_3_minutes
DBSight customer, a shopping comparison site, (anonymous per request) got 2.6 Million Euro funding!





















--
Chris Lu
-------------------------
Instant Scalable Full-Text Search On Any Database/Application
site: http://www.dbsight.net
demo: http://search.dbsight.com
Lucene Database Search in 3 minutes: http://wiki.dbsight.com/index.php?title=Create_Lucene_Database_Search_in_3_minutes
DBSight customer, a shopping comparison site, (anonymous per request) got 2.6 Million Euro funding!