[ https://issues.apache.org/jira/browse/ZOOKEEPER-900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16623686#comment-16623686 ]
Yun Qin commented on ZOOKEEPER-900: ----------------------------------- I believe we have met an issue very similar to what's described in section b) of the original bug description. In our case, sid 1/2 weren't able to rejoin a quorum consisting of sid 3/4/5. The root cause was that QuorumCnxManager$Listener wasn't able to accept connection from sid 1/2 on sid 3/5. Restarting ZooKeeper on sid 3/5 solved the issue. I would really appreciate it if we can proceed with the fix, now that we have a preliminary patch and people willing to take over. Thanks. > FLE implementation should be improved to use non-blocking sockets > ----------------------------------------------------------------- > > Key: ZOOKEEPER-900 > URL: https://issues.apache.org/jira/browse/ZOOKEEPER-900 > Project: ZooKeeper > Issue Type: Bug > Reporter: Vishal Kher > Assignee: Martin Kuchta > Priority: Critical > Fix For: 3.6.0, 3.5.5 > > Attachments: ZOOKEEPER-900-part2.patch, ZOOKEEPER-900.patch, > ZOOKEEPER-900.patch1, ZOOKEEPER-900.patch2 > > > From earlier email exchanges: > 1. Blocking connects and accepts: > a) The first problem is in manager.toSend(). This invokes connectOne(), which > does a blocking connect. While testing, I changed the code so that > connectOne() starts a new thread called AsyncConnct(). AsyncConnect.run() > does a socketChannel.connect(). After starting AsyncConnect, connectOne > starts a timer. connectOne continues with normal operations if the connection > is established before the timer expires, otherwise, when the timer expires it > interrupts AsyncConnect() thread and returns. In this way, I can have an > upper bound on the amount of time we need to wait for connect to succeed. Of > course, this was a quick fix for my testing. Ideally, we should use Selector > to do non-blocking connects/accepts. I am planning to do that later once we > at least have a quick fix for the problem and consensus from others for the > real fix (this problem is big blocker for us). Note that it is OK to do > blocking IO in SenderWorker and RecvWorker threads since they block IO to the > respective peer. > b) The blocking IO problem is not just restricted to connectOne(), but also > in receiveConnection(). The Listener thread calls receiveConnection() for > each incoming connection request. receiveConnection does blocking IO to get > peer's info (s.read(msgBuffer)). Worse, it invokes connectOne() back to the > peer that had sent the connection request. All of this is happening from the > Listener. In short, if a peer fails after initiating a connection, the > Listener thread won't be able to accept connections from other peers, because > it would be stuck in read() or connetOne(). Also the code has an inherent > cycle. initiateConnection() and receiveConnection() will have to be very > carefully synchronized otherwise, we could run into deadlocks. This code is > going to be difficult to maintain/modify. > Also see: https://issues.apache.org/jira/browse/ZOOKEEPER-822 -- This message was sent by Atlassian JIRA (v7.6.3#76005)