[IGNITE-11454] Race in ClientImpl may lead to client node segmentation on fast reconnect

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 2.8
• Assignee: Alexey Goncharuk
• Reporter: Alexey Goncharuk
• Created: 28/Feb/19 15:14
• Updated: 01/Mar/19 13:06
• Resolved: 01/Mar/19 13:04

Description

We have the following code in ClientImpl#tryJoin:

            if (spi.joinTimeout > 0) {
                final int joinCnt0 = joinCnt;

                timer.schedule(new TimerTask() {
                    @Override public void run() {
                        if (joinCnt == joinCnt0 && joining())
                            queue.add(JOIN_TIMEOUT);
                    }
                }, spi.joinTimeout);
            }

We have a window when the timeout object is still scheduled, but the node is already connected to the cluster. The following sequence is possible: a node disconnects, clears it's queue, then timeout object is fired, adds a message to the queue, then tryJoin is called. In this case, the node will be immediately segmented.
ClientReconnectAfterClusterRestartTest demonstrates this if join timeout is set to 10s.

Issue Links

• links to: GitHub Pull Request #6204