[ http://issues.apache.org/jira/browse/POOL-86?page=comments#action_12445951 ] Mike Martin commented on POOL-86: ---------------------------------
Sandy wrote: > I think the changes you suggest to the evict method will fail to make forward > progress if getNumTests() is less than the number of idle objects for the > current key. This is why the _evictLastIndex was needed. _recentlyEvictedKeys can serve the same purpose. This change would do that: --- GenericKeyedObjectPool.java.prev Tue Oct 31 10:24:44 2006 +++ GenericKeyedObjectPool.java Tue Oct 31 10:25:06 2006 @@ -1207,6 +1207,7 @@ return; } key = keyIter.next(); + _recentlyEvictedKeys.add(key); LinkedList list = (LinkedList)_poolMap.get(key); objIter = list.listIterator(list.size()); } @@ -1259,7 +1260,6 @@ } } else { // else done evicting keyed pool - _recentlyEvictedKeys.add(key); key = null; } } > Back to the first part of your original submission: What I understand you > really want is a way to prune the pool size down after a peak load spike. > > Would a decorator that either occasionally discards returned objects or uses > some heuristics to determine when to discard returned objects meet your needs? > This could be implemented so that it doesn't need an eviction thread which > means we can guarantee thread-safety of the pool implementation but it would > require some pool activity to do it's work. I think what I want, and what most DB connection pools need, is the existing idle eviction facility as advertised. I've never encountered the thread-safety problems you mention. Is there an outstanding bug regarding that? If so, it's not occurring in my environment which as I said involves hundreds of keys (DB users), thousands of connections, and in fact is used by hundreds of threads. > GenericKeyedObjectPool retaining too many idle objects > ------------------------------------------------------ > > Key: POOL-86 > URL: http://issues.apache.org/jira/browse/POOL-86 > Project: Commons Pool > Issue Type: Bug > Affects Versions: 1.3 > Reporter: Mike Martin > Assigned To: Sandy McArthur > Attachments: pool-86.patch, pool-86.withtest.patch > > > There are two somewhat related problems in GenericKeyedObjectPool that cause > many more idle objects to be retained than should be, for much longer than > they > should be. > Firstly, borrowObject() is returning the LRU object rather than the MRU > object. > That minimizes rather than maximizes object reuse and tends to refresh all the > idle objects, preventing them from becoming evictable. > The idle LinkedList is being maintained with: > borrowObject: list.removeFirst() > returnObject: list.addLast() > These should either both be ...First() or both ...Last() so the list maintains > a newer-to-older, or vice-versa, ordering. The code in evict() works from the > end of the list which indicates newer-to-older might have been originally > intended. > Secondly, evict() itself has a couple of problems, both of which only show up > when many keys are in play: > 1. Once it processes a key it doesn't advance to the next key. > 2. _evictLastIndex is not working as documented ("Position in the _pool where > the _evictor last stopped"). Instead it's the position where the last > scan > started, and becomes the position at which it attempts to start scanning > *in the next pool*. That just causes objects eligible for eviction to > sometimes be skipped entirely. > Here's a patch fixing both problems: > GenericKeyedObjectPool.java > 990c990 > < pool.addLast(new ObjectTimestampPair(obj)); > --- > > pool.addFirst(new ObjectTimestampPair(obj)); > 1094,1102c1094,1095 > < } > < > < // if we don't have a keyed object pool iterator > < if (objIter == null) { > < final LinkedList list = (LinkedList)_poolMap.get(key); > < if (_evictLastIndex < 0 || _evictLastIndex > > list.size()) { > < _evictLastIndex = list.size(); > < } > < objIter = list.listIterator(_evictLastIndex); > --- > > LinkedList list = (LinkedList)_poolMap.get(key); > > objIter = list.listIterator(list.size()); > 1154,1155c1147 > < _evictLastIndex = -1; > < objIter = null; > --- > > key = null; > 1547,1551d1538 > < > < /** > < * Position in the _pool where the _evictor last stopped. > < */ > < private int _evictLastIndex = -1; > I have a local unit test for this but it depends on some other code I can't > donate. It works like this: > 1. Fill the pool with _maxTotal objects using many different keys. > 2. Select X as a small number, e.g. 2. > 3. Compute: > maxEvictionRunsNeeded = (maxTotal - X) / numTestsPerEvictionRun + 2; > maxEvictionTime = minEvictableIdleTimeMillis + maxEvictionRunsNeeded > * timeBetweenEvictionRunsMillis; > 4. Enter a loop: > a. Borrow X objects. > b. Exit if _totalIdle = 0 > c. Return the X objects. > Fail if loop doesn't exit within maxEvictionTime. > Mike -- This message is automatically generated by JIRA. - If you think it was sent incorrectly contact one of the administrators: http://issues.apache.org/jira/secure/Administrators.jspa - For more information on JIRA, see: http://www.atlassian.com/software/jira --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]