Zhijiang Wang created FLINK-4021:
------------------------------------
Summary: Problem of setting autoread for netty channel when more
tasks sharing the same Tcp connection
Key: FLINK-4021
URL: https://issues.apache.org/jira/browse/FLINK-4021
Project: Flink
Issue Type: Bug
Components: Distributed Runtime
Affects Versions: 1.0.2
Reporter: Zhijiang Wang
Assignee: Zhijiang Wang
Fix For: 1.1.0
More than one task sharing the same Tcp connection for shuffling data.
If the downstream task said as "A" has no available memory segment to read
netty buffer from network, it will set autoread as false for the channel.
When the task A is failed or has available segments again, the netty handler
will be notified to process the staging buffers first, then reset autoread as
true. But in some scenarios, the autoread will not be set as true any more.
That is when processing staging buffers, first find the corresponding input
channel for the buffer, if the task for that input channel is failed, the
decodeMsg method in PartitionRequestClientHandler will return false, that means
setting autoread as true will not be done anymore.
In summary, if one task "A" sets the autoread as false because of no available
segments, and resulting in some staging buffers. If another task "B" is failed
by accident corresponding to one staging buffer. When task A trys to reset
autoread as true, the process can not work because of task B failed.
I have fixed this problem in our application by adding one boolean parameter in
decodeBufferOrEvent method to distinguish whether this method is invoke by
netty IO thread channel read or staged message handler task in
PartitionRequestClientHandler.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
