Luc, I'm not sure that I completely agree with your assessment of the WeakHashMap.
It does assume that the hash value is unique which is a correct usage. It is the responsibility of object.GetHashCode() to return a unique value. This is a problem in any Hashtable implementation and although I haven't seen the internal java code, I assume that it has the same limitation. The object would be the key only superficially as the hashcode would be used internally as well (based on the name of the object, WeakHashMap, anyway). If this is incorrect, I apologize for the noise. I did, however, notice that GetHashCode() is not overridden in the IndexReader and since the .NET framework does not guarantee a unique value (according to the documentation for oject.GetHashCode()), this could present the problem that you have identified. Is there a need to guarantee uniqueness by overriding the method or does anyone know if the value is always unique already? When the Cache object was implemented with Hashtable prior to this patch, the same problem would have been inherent with that implementation (of course with the additional bonus of the memory leak ;)) and I don't think I ever saw collisions although quite honestly I was never really looking for them. Also, you said that the call to WeakHashMap.Keys could return NULL values. This may be the reason why the Java implementation supports null keys. It does seem to be a possibility unless the call to Keys checks that WeakReference.Key.Target is not null at the time of iteration (this would eliminate this possibility since we would now have a strong reference to the object ie: object key = entries[i].Key.Target; if (object != null) keys.Add(object);). The same problem would exist with the Values property. Also, optimizing for few IndexReaders in the WeakHashMap is an assumption that specializes the class for IndexReaders. There is no guarantee that this will be the only use case in the future. The Java documentation says that the same initial parameters (loadfactor and capacity) as HashMap are used which suggests that it should be left in the .NET implementation with the same parameters as the default Hashtable which is its underlying container. Comments? -tim -----Original Message----- From: Vanlerberghe, Luc [mailto:luc.vanlerber...@bvdep.com] Sent: Wednesday, December 31, 2008 4:48 AM To: lucene-net-dev@incubator.apache.org Subject: Problems with patch for LUCENENET-106 Hi, I reviewed the fix for LUCENENET-106 and I see a couple of issues: 1. The implementation of WeakHashMap in SupportClass.cs has a bug in that it assumes all keys will have a unique HashCode. Although the chances of this occurring are *very* small, it cannot be guaranteed. A collision would mean that the cached value for one indexReader could be returned as cached value for another indexReader (using the way this class is used in Lucene as example) 2. This class attempts to be a complete port of the java.util.WeakHashMap, but: - A correct implementation is difficult to implement, and .Net does not have an equivalent of java.lang.ref.ReferenceQueue to accelerate the cleanup process. - Since the behaviour of the WeakHashMap is closely tied to the behaviour of the garbage collector, it is very difficult to test properly, if at all. - Lucene only uses the get(Object key) and put(K key, V value) methods - In Lucene, the keys are IndexReader instances. In normal use cases there will be only 1 IndexReader live, and perhaps another one if 'warming up' is used before switching IndexSearchers after an index update. - If this class is presented as a complete port of java.util.WeakHashMap, users might assume this is production quality code and copy/paste its code in their own projects. Any further bugs found might harm the credibility of the Lucene project, even if those sections are never used in Lucene. (e.g.: The collection returned by Keys might contain null values. There's no guarantee the GC won't intervene between the call to Cleanup and the calls to keys.Add(w.Key.Target). 3. Most support classes that are needed for the conversion of java to .Net are in the anonymous namespace, but this one is in Lucene.Net.Util. I would propose to keep the namespaces corresponding to the original java packages clean and either put all support classes in the anonymous namespace or in a Lucene.Net specific one (Lucene.Net.DotNetPort ?) (I see that George moved it already to SupportClass, thanks George!) I have been using the java version of Lucene in a project for a long time. For a new project that is currently under development, we will be using the .Net version. For now we are using the 2.1 version using a custom patch to avoid the memory leak problem. I'll post an up to date version of this patch as a replacement for the current implementation of Lucene.Net.Util.WeakHashMap sometime next week... - It doesn't try to be a complete implementation of WeakHashMap: only the methods strictly necessary are implemented, all other methods throw a NotImplementedException. (It should probably be renamed to SimplifiedWeakHashMap or something) - It's optimized for the 'low number of keys' case. - It's simple code, but I want to test it thoroughly first. The original patch was put directly in the FieldCacheImpl code, I want to make sure I didn't make some stupid mistake while converting it. I have an account on Jira, but I don't have rights to reopen the issue... Regards, Luc -----Original Message----- From: Digy (JIRA) [mailto:j...@apache.org] Sent: zaterdag 27 december 2008 19:47 To: lucene-net-dev@incubator.apache.org Subject: [jira] Closed: () Lucene.NET (Revision: 603121) is leaking memory [ https://issues.apache.org/jira/browse/LUCENENET-106?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Digy closed LUCENENET-106. -------------------------- Resolution: Fixed Assignee: Digy Patches are applied. DIGY. > Lucene.NET (Revision: 603121) is leaking memory > ----------------------------------------------- > > Key: LUCENENET-106 > URL: https://issues.apache.org/jira/browse/LUCENENET-106 > Project: Lucene.Net > Issue Type: Bug > Environment: .NET 2.0 > Reporter: Anton K. > Assignee: Digy > Priority: Critical > Attachments: DIGY-FieldCacheImpl.patch, Digy.rar, > luceneSrc_memUsage.patch, Paches for v2.3.1.rar, > WeakHashTable+FieldCacheImpl.rar, WeakReferences.rar > > > readerCache Hashtable field (see FieldCacheImpl.cs) never releases some hash > items that have closed IndexReader object as a key. So a lot of Term > instances are never released. > Java version of Lucene uses WeakHashMap and therefore doesn't have this > problem. > This bug can be reproduced only when Sort functionality used during search. > See following link for additional information. > http://www.gossamer-threads.com/lists/lucene/java-user/55681 > Steps to reproduce: > 1)Create index > 2) Modify index by IndexWiter; Close IndexWriter > 3) Use IndexSearcher for searching with Sort; Close InexSearcher > 4) Go to step 2 > You'll get OutOfMemoryException after some time of running this algorithm. -- This message is automatically generated by JIRA. - You can reply to this email to add a comment to the issue online.
