Guys, during debugging of failed test (GridSessionCheckpointSelfTest.testSharedFsCheckpoint) I've noticed that GridTaskWorker can invoke reduce() method in caller thread.
If task isn't annotated by @ComputeTaskMapAsync then mapping job will be run in caller thread. Since job mapping will be finished processDelayedResponses() method will be invoked and if delayed responses queue isn't empty then caller thread can invoke reduce() method eventually and perform reducing synchronously. It can be usefull in case of synchronous execution but, IMHO, it is very strange behavior for asynchronous case because user expects that method will return after creation of task. Similar behavior is possible for all places where code invokes GridTaskProcessor.execute() method (IgniteCompute.broadcast(), IgniteCAche.size(), REST handlers, etc.) I see three options in order to fix the problem: 1. Remove GridTaskWorker.processDelayedResponses() method and all calls. Perhaps, performance can sufer a little bit (but I'm not sure). 2. Add special flag to execute method (e.g. usePool) that will give guarantees that task will not be executed in caller thread. Of course this flag should be add for all methods in call chain. 3. Use task process thread context (GridTaskProcessor.thCtx) and special key that will represent requirement about execution task in pool similar to usePool flag. In case of 2nd and 3rd options we should analyze every usage of GridTaskProcessor.execute() method and solve should caller thread execute task or not. Maybe I missed something and there is better way to solve this problem. I will be grateful for any advice or idea. -- Andrey Gura GridGain Systems, Inc. www.gridgain.com