STORM-946: We should remove Closed Client form cached-node+port->socket in worker

[tedxia]

Patch for STORM-946

The client may be Closed status after reconnect failed, and we will remove closed client from Context to escape memory leak.

But there is also reference for the closed Client in cached-node+port->socket in worker, for this reason we should also remove closed Client from cached-node+port->socket.
Meanwhile there is another reason for us to do so. Think about this situation: worker A connect to worker B1 B2, but for some reason worker B1 B2 died at the same, then nimbus reschedule worker B1 B1. And new B1 B2 may partly rescheduled at the some host:port as old B1 B2, that is (old B1: host1+port1, old B2: host2+port2, new B1: host2+port2, new B2: host3+port3). Worker A realized worker B1 B2 died and start reconnect to worker B1 B2, but before new worker B1 and old B2 have the same host+port, and by the current logic, we will remove old B1 Client and and create new Client for new worker B2, and do nothing to old B2 and new B1 because they have the same host+port. This will result the topology stop processing tuples. Once we remove closed Client from cached-node+port->socket before refresh-connections, this will not happen again.

[vesense]

+1

[caofangkun]

-1

We need take attention to unneeded connections but not closed connections ,
sometimes closed connections need to be reconnected by retry policy to make sure reconnect losted connections.
refresh-connections will try remove unneeded connections

by the way:

java.lang.IllegalArgumentException: No matching field found: isClosed for class backtype.storm.messaging.netty.Client 

May be you should use client.status() somewhere when you need .

[tedxia]

Yes we should use client.status(), and I am sorry for this error, I will add new patch to fix this.

@caofangkun Will you have a look at Client, once client's status change to Closed, it will never have change to change to other state, that is to say reconnect won't be called after client closed.

The problem happened in our product cluster when some net error happen, but after net recover, the topology can not recover. After I remove closed client first, this problem never come again.

[caofangkun]

@tedxia
I could only find two situations ( am I right?) will change client.status to closed by call (.close socket). and do not need reconnect at all
1: when reresh-connections remove unneeded connections and wil remove connecion from cached-node+port->socket
2: when shutdown worker

Does have any situation will call (.close socket) but not removed from cached-node+port->socket ?

[tedxia]

Will you have a look at the description of this pull request, it describe the situation.

[nicom]

@caofangkun

Sorry for commenting on such an old pull request, but we recently also encountered the problem this patch is trying to address.

Does have any situation will call (.close socket) but not removed from cached-node+port->socket ?

Indeed there is one other situation where the close() method is called:
https://github.com/apache/storm/blob/v0.9.6/storm-core/src/jvm/backtype/storm/messaging/netty/Client.java#L505

This only happens if the number of reconnects reaches the number set in the 'storm.messaging.netty.max_retries' config setting (I believe the default is 300). But the problem is aggravated by the fact, that the connectionAttempts counter is never reset. So if a topology runs for a long enough time this might become a problem.
Actually this issue has been fixed in 0.10.0

There is another more serious issue that lead to huge problems in one of out topologies whenever a worker crashed due to some exception.
If worker A sucessfully connects to worker B for the first time during startup but worker B closes the connection for some reason before the :worker-active-flag is set to true (here https://github.com/apache/storm/blob/v0.9.6/storm-core/src/clj/backtype/storm/daemon/worker.clj#L356), there will be no further reconnect attempts, since no messages will be processed and neither send() nor flushMessages() will ever be called.

In our case this frequenty happened because nimbus often rescheduled the whole toplogy right around the same time when the supervisor restarted the failed worker, leading to a race condition. This might have to do with the fact that nimbus.task.timeout.secs and supervisor.worker.timeout.secs have the same value by default.

Tedxia's patch won't fix this problem I belive, since the status returned by status() will be ConnectionWithStatus$Status/Connecting in this situation.

[revans2]

@nicom

Please file a JIRA at https://issues.apache.org/jira under the STORM project describing what you described here. We ran into a similar issue, but I would have to check on the progress of that fix/etc, and if it could be back ported to 0.9.x

[kevinconaway]

@revans2 was there ever an issue filed about this? We are seeing the same behavior in Storm 0.10.0

[kevinconaway]

Actually it looks like this issue

There is another more serious issue that lead to huge problems in one of out topologies whenever a worker crashed due to some exception.
If worker A sucessfully connects to worker B for the first time during startup but worker B closes the connection for some reason before the :worker-active-flag is set to true (here https://github.com/apache/storm/blob/v0.9.6/storm-core/src/clj/backtype/storm/daemon/worker.clj#L356), there will be no further reconnect attempts, since no messages will be processed and neither send() nor flushMessages() will ever be called.

may be fixed by STORM-1609 with the addition of the client keepalive TimerTask

[nicom]

may be fixed by STORM-1609 with the addition of the client keepalive TimerTask

I can confirm that it is