… GC or other pause occurs at inopportune point during test.
…r to long unit test. Cleaned up comments in ClientRequestExecutorPool to make intended semantics clearer.
- factored out a common base class in which all of the nested helper classes are defined - separated out the various types of tests into files: - simple/basic tests - contention tests that spawn threads to generate contention - specific race condition test - The specific race condition test for KeyedResourcePool shows that google issue 276 is resolved: http://code.google.com/p/project-voldemort/issues/detail?id=276
reset(dest) - changed behavior to match original KeyedResourcePool implementation. The original QueuedKeyedResourcePool.reset() was an unnecessary/bad behavior change that canceled enqueued requests. The original behavior was to destroy idle resources whenever pool is reset. test/integration/voldemort/socketpool/E2EClientRequestExecutorPoolAndFailureDetectorTest.java - stress test that has put and get threads contend for slow servers in such a manner as to trigger failure detection to mark nodes unavailable. This excercises connection tear down, reset(), and build up again. This also exercises the code paths in which callbacks do heavyweight work.
…to handle callback work. The ugly code is commented out.
- got rid of anonymous call back classes - factored out waitForResponses logic and processResponses logic for most of these classes. GetAll stands out as being fairly different from the others. - did not refactor to the point of sharing common code across classes, just refactored within each class. - added many TODOs to the code for further refactoring.
… enqueued requests upon setUnavailable. src/java/voldemort/cluster/failuredetector/ThresholdFailureDetector.java - Reduce amount of work done within synchronized section to reduce lock granularity and so ensure "side effects" of node being marked (un)available are not w/in sync section. src/java/voldemort/store/socket/clientrequest/ClientRequestExecutorPool.java - Added TODO/comment to decide whether we want to actively destroy all connections upon node being marked unavailable - Switched behavior to lazily destroying connections.
…ging in the face of a slop not being written. src/java/voldemort/cluster/failuredetector/AbstractFailureDetector.java - less verbose logging when a node is unavailable src/java/voldemort/store/routed/action/AbstractAction.java - less verbose logging in the face of expected "exceptional" responses. src/java/voldemort/store/slop/HintedHandoff.java - ensure that an error message is logged if a slop is not written (we can grep for "Slop write of key.*was not written" in logs) - added TODO because ObsoleteVersionExceptiosn are neither treated as failures or successes in the callback to sendHintParallel. test/integration/voldemort/nonblocking/E2ENonblockingCheckoutTest.java - minor fix test/unit/voldemort/server/socket/ClientRequestExecutorPoolTest.java - from junit3 to junit4
Protected all logger.(debug|info) statements I added with an is(Debug|Info)Enabled() check. Made AsyncRecoveryFailureDetector less verbose. When it polls a server to see if it is available, it now prints out a clean INFO level message. It had been printing out a WARN level message *with* a stacktrace that made this expected behavior look much scarier than it really is.
…-request instrumentation. Expect to remove most of this after debugging.
…entSocketStats. Expect to remove this after debugging.
…xpet to remove most of this after debugging. src/java/voldemort/store/socket/clientrequest/ClientRequestExecutorFactory.java - print out time to establish connection (if it takes longer than 1 ms) src/java/voldemort/utils/pool/KeyedResourcePool.java - print out info about object creation (connection establishment). In particular, how many outstanding creations (connection establishments) are in flight and how many idle resources are in the pool after the newly created resource is checked in.
…toreStats, SocketClientStats, Histogram) behavior. Expect to remove most of these messages after debugging. src/java/voldemort/store/stats/Histogram.java - one more check to harden the imnplementation of insert src/java/voldemort/store/stats/RequestCounter.java - durationMs from int to long - print out timing of histogram reset (q95, q99, reset)
…ropped unnecessary binary search. These changes preserve/correct behavior of the current Histogram. src/java/voldemort/store/stats/Histogram.java - Halved memory footprint by dropping unnecessary "bounds" array - Dropped unnecessary binary search, making insert O(1) rather than O(log(nBuckets)) - Improved documentation - Made interface consistent for type of values inserted/got from histogram (i.e., all are 'long') test/unit/voldemort/store/stats/HistogramTest.java - Added tests for boundary conditions: -ive values are dropped on insert, "too large" values are bucketed in the final bucket on insert. Minor fixes to RequestCounter calls to histogram to conform to 'long' interfaces.
…ronization methods. This should fix deadlock issue.
…reate resources (connections) when they are needed. Now, connections should only be created on demand. The initial code created connections until the max limit of the pool was Some minor tweaks to test that confirm the desired behavior at a unit test level.