Are you certain that nimbus.task.timeout.secs is the correct config?
That config controls the length of time before nimbus thinks a worker has timed
out.
https://github.com/apache/incubator-storm/blob/master/storm-core/src/clj/backtype/storm/daemon/nimbus.clj#L369-L372
Its default is 30 seconds.
https://github.com/apache/incubator-storm/blob/master/conf/defaults.yaml#L45
storm.zookeeper.connection.timeout: 300000
storm.zookeeper.session.timeout: 300000
So these will make the situation worse while workers losing connections to ZK,
since it will cause the workers to wait longer before reconnecting. They could
wait until nimbus thinks the worker is dead before trying to reconnect.
supervisor: 2014-05-23 20:17:30 INFO supervisor:0 - Shutting down and clearing state for
id 94349373-74ec-484b-a9f8-a5076e17d474. Current supervisor time: 1400876250. State:
:disallowed, Heartbeat: #backtype.storm.daemon.common.WorkerHeartbeat{{:time-secs
1400876249, :storm-id "test-46-1400863199", :executors #{[-1 -1]}, :port 6700}
Here if the "State" is ":disallowed", then that means it is Nimbus that de-scheduled the
worker on that node--very probably in this case because it thought it was dead. When the supervisor sees
this, it will kill the worker. (A state of ":timed-out" means instead that the worker did not
heartbeat to its supervisor in time.)
If the CPU load on the worker was high enough to prevent heartbeats, then I
would expect to see :timed-out state above instead of :disallowed. The reason
is that the worker has only 5 seconds to do those heartbeats, while it has 30
seconds to heartbeat to nimbus (via ZK). (More often what happens to cause
this is memory has run out and garbage collection stops everything just long
enough.)
The real question is why connections from the worker to ZK are timing out in
the first place.
What about the ZK servers? Sometimes ZooKeeper servers cannot keep up, and
that causes pretty severe problems with timeouts.
--
Derek
On 5/30/14, 17:51, Michael Dev wrote:
Michael R,
We don't have GC logging enabled yet. I lean towards agreeing with Derek that I
don't think it's the issue but I will take a look at logging on Monday just to
verify.
Derek D,
Are you certain that nimbus.task.timeout.secs is the correct config? Tracing
through the github code it would seem to me that this is used as the timeout
value when making a Thrift connection to the Nimbus node. I thought the logs
indicated the timeout was occurring in the session connection to zookeeper as
evidenced by ClientCxn being a Zookeeper class.
I discovered that we were running with the default maxSessionTimeout zookeeper
config of 40 seconds. This would explain why our storm config of 5 minutes was
not being picked up (but obviously not the root problem nor why timeout
messages report 14-20 second timeout values). Typically we saw losses in
connection occur when our cluster becomes super busy with a burst of data
pushing workers to near 100% CPU. I'm testing the following configs over the
weekend to see if they at least allow us to prevent chronic worker restarting
during the brief high CPU periods.
Our current setup is as follows:
Storm 0.9.0.1
3 Storm node cluster
1 Supervisor running per Storm node
1-3 topologies deployed on the Storm cluster (depends on dev/prod/etc systems)
3 Workers per topology
Variable number of executors per component depending on how slow that component
is. Example file i/o has many executors (say 12) while in memory validation has
only 3 executors. Always maintaining a multiple of the number of workers for
even distribution.
Kyro serialization with Java Serialization failover disabled to ensure we're
using 100% kryo between bolts.
zoo.cfg
tickTime=2000
dataDir=/srv/zookeeper/data
clientPort=2182
initLimit=5
syncLimit=2
skipACL=true
maxClientCnxns=1000
maxSessionTimeout=300000
server.1=node1
server.2=node2
server.3=node3
storm yaml
storm.zookeeper.port: 2182
storm.local.dir=/srv/storm/data
nimbus.host: "node1"
storm.zookeeper.servers:
- "node1"
- "node2"
- "node3"
supervisor.slot.ports:
- 6700
- 6701
- 6702
- 6703
- 6704
java.library.path: "/usr/lib:/srv/storm/lib"
#Storm 0.9 netty support
storm.messaging.transport: "backtype.storm.messaging.netty.Context"
storm.messaging.netty.server_worker_threads: 1
storm.messaging.netty.client_worker_threads: 1
storm.messaging.netty.buffer_size: 5242880
storm.messaging.netty.max_retries: 100
storm.messaging.netty.max_wait_ms: 1000
storm.messaging.netty.min_wait_ms: 100
# Timeout band-aids in testing
topology.receiver.buffer.size: 2
storm.zookeeper.connection.timeout: 300000
storm.zookeeper.session.timeout: 300000
Date: Thu, 29 May 2014 12:56:19 -0500
From: der...@yahoo-inc.com
To: user@storm.incubator.apache.org
Subject: Re: Workers constantly restarted due to session timeout
OK, so GC is probably not the issue.
Specifically, this is a connection timeout to ZK from the worker, and it is
resulting in nimbus removing it from the assignments for that node. In turn,
the supervisor reads the schedule and shoots the worker because it is no longer
scheduled to be running.
The relevant config is nimbus.task.timeout.secs, and I think the default is
30s. What you could try is to make nimbus timeout longer than
storm.zookeeper.session.timeout. This would allow the ZK connections to
timeout and get a heartbeat in before nimbus decides they have timed out.
But the real question is why are the ZK sessions timing out at all?
Do you see this on several workers on that node? What about the supervisor?
What about other nodes? What do the ZK logs say?
--
Derek
On 5/29/14, 11:45, Michael Dev wrote:
Derek,
We are currently running with -Xmx60G and only about 20-30G of that has been
observed to be used. I'm still observing workers restarted every 2 minutes.
What timeout is relevant to increase for the heartbeats in question? Is it be a
config on the Zookeeper side we can increase to make our topology more
resilient to these restarts?
Michael
Date: Fri, 23 May 2014 15:50:50 -0500
From: der...@yahoo-inc.com
To: user@storm.incubator.apache.org
Subject: Re: Workers constantly restarted due to session timeout
2) Is this expected behavior for Storm to be unable to keep up with heartbeat
threads under high CPU or is our theory incorrect?
Check your JVM max heap size (-Xmx). If you use too much, the JVM will
garbage-collect, and that will stop everything--including the thread whose job
it is to do the heartbeating.
--
Derek
On 5/23/14, 15:38, Michael Dev wrote:
Hi all,
We are seeing our workers constantly being killed by Storm with to the
following logs:
worker: 2014-05-23 20:15:08 INFO ClientCxn:1157 - Client session timed out,
have not heard from the server in 28105ms for sessionid 0x14619bf2f4e0109,
closing socket and attempting reconnect
supervisor: 2014-05-23 20:17:30 INFO supervisor:0 - Shutting down and clearing state for
id 94349373-74ec-484b-a9f8-a5076e17d474. Current supervisor time: 1400876250. State:
:disallowed, Heartbeat: #backtype.storm.daemon.common.WorkerHeartbeat{{:time-secs
1400876249, :storm-id "test-46-1400863199", :executors #{[-1 -1]}, :port 6700}
Eventually Storm decides to just kill the worker and restart it as you see in
the supervisor log. We theorize this is the Zookeeper heartbeat thread and it
is being choked out due to very high CPU load on the machine (near 100%).
I have increased the connection timeouts in the storm.yaml config file yet
Storm seems to continue to use some unknown value for the above client session
timeout messages:
storm.zookeeper.connection.timeout: 300000
storm.zookeeper.session.timeout: 300000
1) What timeout config is appropriate for the above timeout message?
2) Is this expected behavior for Storm to be unable to keep up with heartbeat
threads under high CPU or is our theory incorrect?
Thanks,
Michael
