the blocking Java Socket API doesn't have a timeout on socketWrite invocations.
This means, if a TCP session is dropped or terminated without RST or FIN packets, the operating system it left to eventually time out the session. On the linux kernel this timeout usually takes 15 to 30minutes.
For this entire period, the AMQ server hangs, and producers and consumers are unable to use a topic.
I have created two patches for this at the page:
Let me show a bit more
"ActiveMQ Transport: tcp:///X.YYY.XXX.ZZZZ:2011" daemon prio=10 tid=0x0000000055d39000 nid=0xc78 runnable [0x00000000447c9000..0x00000000447cac10]
at java.net.SocketOutputStream.socketWrite0(Native Method)
This is a thread stuck in blocking IO, and can be stuck for 30 minutes during the kernel TCP retransmission attempts.
Unfortunately the thread dump is very misleading since the name of the thread, is not the destination or even remotely related to the socket it is operating on.
To mend this, a very simple (and configurable) ThreadNameFilter has been suggested to the patch, that appends the destination and helps the system administrator correctly identify the client that is about to receive data.
- locked <0x00002aaaec155818> (a org.apache.activemq.broker.region.Topic)
The lock being held at this issue unfortunately makes the entire Topic single threaded.
When this lock is being held, no other clients (producers and consumers) can publish to/receive from this topic.
And this lock can hold up to 30 minutes.
I consider solving this single threaded behavior a 'feature enhancement' that should be handled separately from this bug. Because even if it is solved, threads still risk being stuck in socketWrite0 for dropped connections that still appear to be established.
For this, I have implemented a socket timeout filter, based on a TransportFilter, this filter only times out connections that are actually writing data.
the binary 0000.jar applies to both 5.1 and trunk and can be used today in existing environments.