Gurleen S Dhody created TEZ-4193:
------------------------------------
Summary: Async DIspatcher drained property
Key: TEZ-4193
URL: https://issues.apache.org/jira/browse/TEZ-4193
Project: Apache Tez
Issue Type: Bug
Affects Versions: 0.9.3
Reporter: Gurleen S Dhody
Fix For: 0.9.3[https://github.com/apache/tez/blob/cf111b1c896db94d40ea92fc6da2a6fcce356114/tez-common/src/main/java/org/apache/tez/common/AsyncDispatcher.java#L318] The drained property should be set after all dispatchers in the handle() [https://github.com/apache/tez/blob/cf111b1c896db94d40ea92fc6da2a6fcce356114/tez-common/src/main/java/org/apache/tez/common/AsyncDispatcher.java#L311] function are checked for. We should set {code:java} // drained = false; {code} [https://github.com/apache/tez/blob/cf111b1c896db94d40ea92fc6da2a6fcce356114/tez-common/src/main/java/org/apache/tez/common/AsyncDispatcher.java#L332 |https://github.com/apache/tez/blob/cf111b1c896db94d40ea92fc6da2a6fcce356114/tez-common/src/main/java/org/apache/tez/common/AsyncDispatcher.java#L332]here. Reason is that the thread on the event queue might already be empty and waiting on [https://github.com/apache/tez/blob/cf111b1c896db94d40ea92fc6da2a6fcce356114/tez-common/src/main/java/org/apache/tez/common/AsyncDispatcher.java#L106 |https://github.com/apache/tez/blob/cf111b1c896db94d40ea92fc6da2a6fcce356114/tez-common/src/main/java/org/apache/tez/common/AsyncDispatcher.java#L106]new event. Setting drained as false but passing the event to another dispatcher will cause issue when stopping this dispatcher service, when drainEventsOnStop is set true. The thread will be stuck in queue and drained property will be false causing a deadlock here [https://github.com/apache/tez/blob/cf111b1c896db94d40ea92fc6da2a6fcce356114/tez-common/src/main/java/org/apache/tez/common/AsyncDispatcher.java#L146] Needs to be addressed in [https://github.com/apache/tez/blob/cf111b1c896db94d40ea92fc6da2a6fcce356114/tez-common/src/main/java/org/apache/tez/common/AsyncDispatcherConcurrent.java] in AsyncConcurrecyDispatcher as well. -- This message was sent by Atlassian Jira (v8.3.4#803005)
