Just a few examples of "problems" using ThreadLocals.

http://opensource.atlassian.com/projects/hibernate/browse/HHH-2481
http://www.theserverside.com/news/thread.tss?thread_id=41473

Once again, I'm not pointing to Lucene SegmentReader as a "bad" implementation, and maybe the current "problems" of ThreadLocals are not a problem for SegmentReader but it seems safer to use ThreadLocals to pass context information which is cleared when the call exits instead of storing long-lived objects.


robert engels wrote:
Aside from the pre-1.5 thread local "perceived leak", there are no issues with ThreadLocals if used properly.

There is no need for try/finally blocks, unless you MUST release resources immediately, usually this is not the case, which is why a ThreadLocal is used in the first place.

From the ThreadLocalMap javadoc...

 /**
     * ThreadLocalMap is a customized hash map suitable only for
     * maintaining thread local values. No operations are exported
     * outside of the ThreadLocal class. The class is package private to
     * allow declaration of fields in class Thread.  To help deal with
     * very large and long-lived usages, the hash table entries use
     * WeakReferences for keys. However, since reference queues are not
     * used, stale entries are guaranteed to be removed only when
     * the table starts running out of space.
     */

/**
         * Heuristically scan some cells looking for stale entries.
         * This is invoked when either a new element is added, or
         * another stale one has been expunged. It performs a
         * logarithmic number of scans, as a balance between no
         * scanning (fast but retains garbage) and a number of scans
         * proportional to number of elements, that would find all
         * garbage but would cause some insertions to take O(n) time.
         *
         * @param i a position known NOT to hold a stale entry. The
         * scan starts at the element after i.
         *
         * @param n scan control: <tt>log2(n)</tt> cells are scanned,
         * unless a stale entry one is found, in which case
         * <tt>log2(table.length)-1</tt> additional cells are scanned.
         * When called from insertions, this parameter is the number
         * of elements, but when from replaceStaleEntry, it is the
         * table length. (Note: all this could be changed to be either
         * more or less aggressive by weighting n instead of just
         * using straight log n. But this version is simple, fast, and
         * seems to work well.)
         *
         * @return true if any stale entries have been removed.
         */


The instance ThreadLocals (and what the refer to) will be GC'd when the containing Object is GC'd.

There IS NO MEMORY LEAK in ThreadLocal. If the ThreadLocal refers to an object that has native resources (e.g. file handles), it may not be released until other thread locals are created by the thread (or the thread terminates).

You can avoid this "delay" by calling remove(), but in most applications it should never be necessary - unless a very strange usage...

On Jul 9, 2008, at 2:37 PM, Adrian Tarau wrote:

From what I know, storing objects in ThreadLocal is safe as long as you release the object within a try {} finall {} block or store objects which are independent of the rest of the code(no dependencies).Otherwise it can get pretty tricky(memory leaks, classloader problems) after awhile.

It is pretty convenient to pass HTTP request information with a ThreadLocal in a servlet(but you should cleanup the variable before leaving the servlet) but I'm not sure how safe it is in this case.

robert engels wrote:
Using synchronization is a poor/invalid substitute for thread locals in many cases.

The point of the thread local in these referenced cases is too allow streaming reads on a file descriptor. if you use a shared file descriptor/buffer you are going to continually invalidate the buffer.

On Jul 8, 2008, at 5:12 AM, Michael McCandless wrote:


Well ... SegmentReader uses ThreadLocal to hold a thread-private instance of TermVectorsReader, to avoid synchronizing per-document when loading term vectors.

Clearing this ThreadLocal value per call to SegmentReader's methods that load term vectors would defeat its purpose.

Though, of course, we then synchronize on the underlying file (when using FSDirectory), so perhaps we are really not saving much by using ThreadLocal here.  But we are looking to relax that low level synchronization with LUCENE-753.

Maybe we could make our own ThreadLocal that just uses a HashMap, which we'd have to synchronize on when getting the per-thread instances.  Or, go back to sharing a single TermVectorsReader and synchronize per-document.

Jason has suggested moving to a model where you ask the IndexReader for an object that can return term vectors / stored fields / etc, and then you interact with that many times to retrieve each doc.  We could then synchronize only on retrieving that object, and provide a thread-private instance.

It seems like we should move away from using ThreadLocal in Lucene and do "normal" synchronization instead.

Mike

Adrian Tarau wrote:

Usually ThreadLocal.remove() should be called at the end(in a finally block), before the current call leaves your code.

Ex : if during searching ThreadLocal is used, every search(..) method should cleanup any ThreadLocal variables, or even deeper in the implementation. When the call leaves Lucene any used ThreadLocal should be cleaned up.

Michael McCandless wrote:

ThreadLocal, which we use in several places in Lucene, causes a leak in app servers because the classloader never fully deallocates Lucene's classes because the ThreadLocal is holding strong references.

Yet, ThreadLocal is very convenient for avoiding synchronization.

Does anyone have any ideas on how to solve this w/o falling back to "normal" synchronization?

Mike

Begin forwarded message:

From: "Yonik Seeley" <yonik@apache.org>
Date: July 7, 2008 3:30:28 PM EDT
Subject: Re: ThreadLocal in SegmentReader

On Mon, Jul 7, 2008 at 2:43 PM, Michael McCandless
So now I'm confused: the SegmentReader itself should no longer be reachable,
assuming you are not holding any references to your IndexReader.

Which means the ThreadLocal instance should no longer be reachable.

It will still be referenced from the Thread(s) ThreadLocalMap
The key (the ThreadLocal) will be weakly referenced, but the values
(now stale) are strongly referenced and won't be actually removed
until the table is resized (under the Java6 impl at least).
Nice huh?

-Yonik

---------------------------------------------------------------------
For additional commands, e-mail: java-user-help@lucene.apache.org



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



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



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



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



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