[
https://issues.apache.org/jira/browse/KAFKA-3568?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15248554#comment-15248554
]
Greg Zoller commented on KAFKA-3568:
------------------------------------
Ok... I don't believe my test results, but I have repeated them.
I found that build 9cfb99 (Mar 22) worked for me -- KafkaProducer didn't
hang/time-out. I get meaningful commit metadata on send() callback and null as
the exception, and it returned instantly as expected. The next-newer build,
73470b0, KafkaProducer send() calls would hang for 60 seconds then
timeout--null for callback metadata and timeout exceptions populated.
The only differences between the content of these two builds were documentation
changes and a tiny config file change. So I started making the changes
manually one at a time until I found the issue.... and I can't believe it.
Build 9cfb99 removed two lines in config/server.properties:
#advertised.host.name=<hostname routable by clients>
#advertised.port=<port accessible by clients>
(It also commented out a listeners config variable but turning this on/off made
no difference to my issue.) Note: THE TWO LINES ARE COMMENTED OUT! I flipped
this a few times in disbelief but the results were the same.
My process:
1) ./gradlew clean
2) ./gradlew -PscalaVersion=2.11 releaseTarGz -x signArchives
3) (build a Docker image with the tgz file just like in spotify/kafka)
4) ./gradlew -PscalaVersion=2.11 install_2_11
5) clean, update, then recompile my test code in sbt, built against the release
in .m2 local maven repo
6) run
This process is repeatable and I applied it consistently through all tests.
Not convinced yet? I went back to trunk and pasted in those 2 commented out
lines in config/server.properties, leaving everything else as-is and rebuilt
using the process above. Worked! Here's the git diff on my 2-line change to
trunk:
$ git diff
diff --git a/config/server.properties b/config/server.properties
index aebcb87..d9d17e8 100644
--- a/config/server.properties
+++ b/config/server.properties
@@ -29,6 +29,9 @@ broker.id=0
# listeners = PLAINTEXT://your.host.name:9092
#listeners=PLAINTEXT://:9092
+#advertised.host.name=<hostname routable by clients>
+#advertised.port=<port accessible by clients>
+
# Hostname and port the broker will advertise to producers and consumers. If
not set,
# it uses the value for "listeners" if configured. Otherwise, it will use the
value
# returned from java.net.InetAddress.getCanonicalHostName().
I dunno what to say. I'll keep flipping this and have a friend confirm the
results. Is some code or script textually looking at this file (not parsing
it) for either of these 2 properties?
> KafkaProducer fails with timeout on message send()
> --------------------------------------------------
>
> Key: KAFKA-3568
> URL: https://issues.apache.org/jira/browse/KAFKA-3568
> Project: Kafka
> Issue Type: Bug
> Components: clients
> Affects Versions: 0.10.1.0
> Environment: MacOS Docker
> Reporter: Greg Zoller
> Labels: producer
>
> I had a KafkaProducer working fine in 0.9.0.1. I was having unrelated
> problems in that version so thought to try 0.10.1.0. I built it as I did
> 0.9.0.1:
> Fresh build against Scala 2.11.7. Built the tgz build plus local maven
> install. From the tgz I created a Docker image similar to spotify/kafka. I
> linked my producer code to the maven jars. This process worked in 0.9.
> Code is here:
> https://gist.github.com/gzoller/145faef1fefc8acea212e87e06fc86e8
> At the bottom you can see a clip from the output... there's a warning about
> metadata (not sure if its important or not) and then its trying to send()
> messages and timing out. I clipped the output, but it does fail the same way
> for each message sent in 0.10.1.0. Same code compiled against 0.9.0.1
> populates the topic's partitions w/o problem.
> Was there a breaking change between 0.9 and 0.10, or is this a bug?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
