[CASSANDRA-11038] Is node being restarted treated as node joining?

• Type: Bug
• Status: Resolved
• Priority: Low
• Resolution: Fixed
• Fix Version/s: 2.2.7, 3.0.8, 3.8
• Component/s: Legacy/Distributed Metadata
• Severity: Low
• Assignee: Sam Tunnicliffe
• Reporter: cheng ren

Authors

Sam Tunnicliffe

Reviewers

Joel Knighton

• Created: 20/Jan/16 01:05
• Updated: 08/Jul/20 10:03
• Resolved: 20/Jun/16 13:32

Description

Hi,
What we found recently is that every time we restart a node, all other nodes in the cluster treat the restarted node as a new node joining and issue node joining notification to clients. We have traced the code path being hit when a peer node detected a restarted node:

src/java/org/apache/cassandra/gms/Gossiper.java

    private void handleMajorStateChange(InetAddress ep, EndpointState epState)
    {
        if (!isDeadState(epState))
        {
            if (endpointStateMap.get(ep) != null)
                logger.info("Node {} has restarted, now UP", ep);
            else
                logger.info("Node {} is now part of the cluster", ep);
        }
        if (logger.isTraceEnabled())
            logger.trace("Adding endpoint state for " + ep);
        endpointStateMap.put(ep, epState);

        // the node restarted: it is up to the subscriber to take whatever action is necessary
        for (IEndpointStateChangeSubscriber subscriber : subscribers)
            subscriber.onRestart(ep, epState);

        if (!isDeadState(epState))
            markAlive(ep, epState);
        else
        {
            logger.debug("Not marking " + ep + " alive due to dead state");
            markDead(ep, epState);
        }
        for (IEndpointStateChangeSubscriber subscriber : subscribers)
            subscriber.onJoin(ep, epState);
    }

subscriber.onJoin(ep, epState) ends up with calling onJoinCluster in Server.java

src/java/org/apache/cassandra/transport/Server.java
        public void onJoinCluster(InetAddress endpoint)
        {
server.connectionTracker.send(Event.TopologyChange.newNode(getRpcAddress(endpoint), server.socket.getPort()));
        }

We have a full trace of code path and skip some intermedia function calls here for being brief.

Upon receiving the node joining notification, clients would go and scan system peer table to fetch the latest topology information. Since we have tens of thousands of client connections, scans from all of them put an enormous load to our cluster.

Although in the newer version of driver, client skips fetching peer table if the new node has already existed in local metadata, we are still curious why node being restarted is handled as node joining on server side? Did we hit a bug or this is the way supposed to be? Our old java driver version is 1.0.4 and cassandra version is 2.0.12.

Thanks!

Issue Links

• is duplicated by: CASSANDRA-11731 dtest failure in pushed_notifications_test.TestPushedNotifications.move_single_node_test
• is related to: CASSANDRA-11360 dtest failure in pushed_notifications_test.TestPushedNotifications.restart_node_test
• is related to: CASSANDRA-15912 Occasional NEW_NODE event sent instead of NODE_UP in dTest