lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Otis Gospodnetic <>
Subject ThreadLocal leak (was Re: Leaking org.apache.lucene.index.* objects)
Date Sat, 16 Dec 2006 07:15:28 GMT
Moving to java-dev, I think this belongs here.
I've been looking at this problem some more today and reading about ThreadLocals.  It's easy
to misuse them and end up with memory leaks, apparently... and I think we may have this problem

The problem here is that ThreadLocals are tied to Threads, and I think the assumption in TermInfosReader
and SegmentReader is that (search) Threads are short-lived: they come in, scan the index,
do the search, return and die.  In this scenario, their ThreadLocals go to heaven with them,
too, and memory is freed up.

But when Threads are long-lived, as they are in thread pools (e.g. those in servlet containers),
those ThreadLocals stay alive even after a single search request is done.  Moreover, the Thread
is reused, and the new TermInfosReader and SegmentReader put some new values in that ThreadLocal
on top of the old values (I think) from the previous search request.  Because the Thread still
has references to ThreadLocals and the values in them, the values never get GCed.

I tried making ThreadLocals in TIR and SR static, I tried wrapping values saved in TLs in
WeakReference, I've tried using WeakHashMap like in Robert Engel's FixedThreadLocal class
from LUCENE-436, but nothing helped.  I thought about adding a public static method to TIR
and SR, so one could call it at the end of a search request (think servlet filter) and clear
the TL for the current thread, but that would require making TIR and SR public and I'm not
100% sure if it would work, plus that exposes the implementation details too much.
I don't have a solution yet.
But do we *really* need ThreadLocal in TIR and SR?  The only thing that TL is doing there
is acting as a per-thread storage of some cloned value (in TIR we clone SegmentTermEnum and
in SR we clone TermVectorsReader).  Why can't we just store those cloned values in instance
variables?  Isn't whoever is calling TIR and SR going to be calling the same instance of TIR
and SR anyway, and thus get access to those cloned values?

I'm really amazed that we haven't heard any reports about this before.  I am not sure why
my application started showing this leak only about 3 weeks ago.  It is getting more pounded
on than before, so maybe that made the leak more obvious.  My guess is that more common Lucene
usage is with a single index or a small number of them, and with short-lived threads, where
this problem isn't easily visible.  In my case I deal with a few tens of thousands of indices
and several parallel search threads that live forever in the thread pool.

Any thoughts about this or possible suggestions for a fix?

----- Original Message ----
From: Otis Gospodnetic <>
Sent: Friday, December 15, 2006 12:28:29 PM
Subject: Leaking org.apache.lucene.index.* objects


About 2-3 weeks ago I emailed about a memory leak in my application.  I then found some problems
in my code (I wasn't closing IndexSearchers explicitly) and took care of those.  Now I see
my app is still leaking memory - jconsole clearly shows the "Tenured Gen" memory pool getting
filled up until I hit the OOM, but I can't seem to pin-point the source.

I found that a bunch or o.a.l.index.* objects are not getting GCed, even though they should.
 For example:

$ jmap -histo:live 7825 | grep apache.lucene.index | head -20 | sort -k2 -nr
num   #instances    #bytes  class name
  4:   1764840    98831040  org.apache.lucene.index.CompoundFileReader$CSIndexInput
  5:   2119215    67814880  org.apache.lucene.index.TermInfo
  7:   1112459    35598688  org.apache.lucene.index.SegmentReader$Norm
  9:   2132311    34116976  org.apache.lucene.index.Term
 12:   1117897    26829528  org.apache.lucene.index.FieldInfo
 13:    225340    18027200  org.apache.lucene.index.SegmentTermEnum
 15:    589727    14153448  org.apache.lucene.index.TermBuffer
 21:     86033     8718504  [Lorg.apache.lucene.index.TermInfo;
 20:     86033     8718504  [Lorg.apache.lucene.index.Term;
 23:     86120     7578560  org.apache.lucene.index.SegmentReader
 26:     90501     5068056
 27:     86120     4822720  org.apache.lucene.index.TermInfosReader
 33:     86130     3445200  org.apache.lucene.index.SegmentInfo
 36:     87355     2795360$Descriptor
 38:     86120     2755840  org.apache.lucene.index.FieldsReader
 39:     86050     2753600  org.apache.lucene.index.CompoundFileReader
 42:     46903     2251344  org.apache.lucene.index.SegmentInfos
 43:     93778     2250672$Entry
 45:     93778     1500448$CreationPlaceholder
 47:     86510     1384160  org.apache.lucene.index.FieldInfos

I'm running my app in search-only mode - no adds or deletes.
The counts of these objects just keeps going up, even though I am explicitly closing the IndexSearcher.
 I can see that file descriptors _are_ freed up after searcher.close(), because lsof no longer
shows them, but the above objects just linger and accumulate, even when I force GC via jconsole
or via the profiler.

I thought maybe various *Readers are not getting close()d, but I've double-checked all *Readers
above, and they all seem to close their IndexInput references.  The static nested class CompoundFileReader.CSIndexInput
has a close() without any implementation.  At first I thought that was an omission, but adding
a close of the inner IndexInput there resulted in a search-time error.  I've added the lovely
print debugging to various close() methods and see those methods being called.  I've added
finalize() with some print debugging to SegmentReader, TermInfosReader, SegmentTermEnum, FieldsReader,
and CompoundFileReader.  All but CFReader get finalized after a while.

My application is running as a webapp and has thousands of separate indices.  This means it's
very multi-threaded and the servlet container has a pool of threads that handle requests,
and each request may be for a different index.  I cache IndexSearchers for a while, and purge/close
them every N minutes if they have been idle more than M minutes.
It occurred to me last night that things like TermInfosReader and SegmentReader are using
ThreadLocal, and since threads are used in a thread pool, and thus shared with requests handling
searches against different indices, it's not clear to me what happens with object instances
that are put in those ThreadLocals in such scenario.  Aren't things going to step on each
others' toes?
TIR has close() and SR has doClose(), so I put <TL inst>.set(null) there.  This immediately
got rid of those instances of CompoundFileReader.CSIndexInput in my dev environment!!!! Yeeees!

But in my dev environment I tested my additions by slamming my app against a *single* index.
 I took my modified Lucene to production, and quickly saw all those o.a.l.index.* objects
accumulate again.  I also see a lot of ThreadLocal's kids:

 16:    419387    13420384  java.lang.ThreadLocal$ThreadLocalMap$Entry

I *think* that points out to some issues with how that ThreadLocal is used there, in a multi-threaded,
multi-index environments.
I'm running JDK 6, and while this problem sounds a bit like LUCENE-436, I'm not yet sure if
it's the same thing.  
Because my IndexSearchers (and thus all those o.a.l.index.* objects) are long-lived, and threads
are shared and reused for searching of other indices, those close() and doClose() methods
are not called at the end of the request life-cycle, so at the end of the request those TL
instances will *still* have something in them.  When their thread is later reused for searching
of another index, new data will be put in them, but the old data will never be cleaned out!
It seems a bit odd, but with this ThreadLocals, shouldn't a multi-threaded, multi-index webapp
really have to "clean" those ThreadLocal instances either before or at the end of the request?

I'm running out of ideas, and was wondering if anyone has any thoughts about what could still
be holding references to the above classes.  I have some 20-30MB memory snapshots (via YourKit)
and heap dumps (via jmap), if anyone is interested.


To unsubscribe, e-mail:
For additional commands, e-mail:

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message