[
https://issues.apache.org/jira/browse/KAFKA-3148?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15116072#comment-15116072
]
Robert Joseph Evans commented on KAFKA-3148:
--------------------------------------------
I created some code that can show the issue on
https://github.com/revans2/kafka/tree/KAFKA-3148
https://github.com/revans2/kafka/blob/KAFKA-3148/clients/src/test/java/org/apache/kafka/common/network/EchoServer.java#L85-L89
is the sleep to cause the hang to happen all the time. It may need to be
longer, up to 4 seconds, to guarantee that it always happens on slower boxes.
https://github.com/revans2/kafka/blob/KAFKA-3148/clients/src/test/java/org/apache/kafka/common/network/SslSelectorTest.java#L118-L119
adds in a 60 second timeout on the test so it does not hang, but fails instead.
https://github.com/revans2/kafka/blob/KAFKA-3148/clients/src/test/java/org/apache/kafka/common/network/SslSelectorTest.java#L112-L117
is an ugly hack that shows once you are in this situation (in this case the
test has run for 15 seconds) send another message to get things going again.
Typically what happens is that the client server intermix sends and receives
until the renegotiate happens and the client finishes sending the 500 messages
before the echo server wakes up from the sleep. It renegotiates and then
finishes echoing back the rest of the 500 messages without the client waking up
to read anything. A few seconds later the 15 second timeout happens and the
extra send goes out and all of the buffered messages are read.
> SslSelectorTest.testRenegotiation can hang on slow boxes
> --------------------------------------------------------
>
> Key: KAFKA-3148
> URL: https://issues.apache.org/jira/browse/KAFKA-3148
> Project: Kafka
> Issue Type: Bug
> Components: clients
> Affects Versions: 0.9.0.0
> Reporter: Robert Joseph Evans
>
> The SslSelectorTest hangs very frequently (about 75% of the time) for me when
> running on a very slow Linux Virtual Machine, but I can artificially
> reproduce the issue on a fast box by inserting a sleep right before the
> renegotiate happens in the EchoServer.
> It appears that after the renegotiate happens the client will not be woken up
> to receive anything unless a send is first done.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
