1. ZooKeeper
  2. ZOOKEEPER-1057

zookeeper c-client, connection to offline server fails to successfully fallback to second zk host


    • Type: Bug Bug
    • Status: Closed
    • Priority: Blocker Blocker
    • Resolution: Fixed
    • Affects Version/s: 3.3.1, 3.3.2, 3.3.3
    • Fix Version/s: 3.4.6, 3.5.0
    • Component/s: c client
    • Labels:
    • Environment:


      Hello, I'm a contributor for the node.js zookeeper module: https://github.com/yfinkelstein/node-zookeeper
      i'm using zk 3.3.3 for the purposes of this issue, but i have validated it fails on 3.3.1 and 3.3.2

      i'm having an issue when trying to connect when one of my zookeeper servers is offline.
      if the first server attempted is online, all is good.

      if the offline server is attempted first, then the client is never able to connect to any server.
      inside zookeeper.c a connection loss (-4) is received, the socket is closed and buffers are cleaned up, it then attempts the next server in the list, creates a new socket (which gets the same fd as the previously closed socket) and connecting fails, and it continues to fail seemingly forever.
      The nature of this "fail" is not that it gets -4 connection loss errors, but that zookeeper_interest doesn't find anything going on on the socket before the user provided timeout kicks things out. I don't want to have to wait 5 minutes, even if i could make myself.

      this is the message that follows the connection loss:
      2011-04-27 23:18:28,355:13485:ZOO_ERROR@handle_socket_error_msg@1530: Socket [] zk retcode=-7, errno=60(Operation timed out): connection timed out (exceeded timeout by 3ms)
      2011-04-27 23:18:28,355:13485:ZOO_ERROR@yield@213: yield:zookeeper_interest returned error: -7 - operation timeout

      While investigating, i decided to comment out close(zh->fd) in handle_error (zookeeper.c#1153)
      now everything works (obviously i'm leaking an fd). Connection the the second host works immediately.
      this is the behavior i'm looking for, though i clearly don't want to leak the fd, so i'm wondering why the fd re-use is causing this issue.
      close() is not returning an error (i checked even though current code assumes success).

      i'm on osx 10.6.7
      i tried adding a setsockopt so_linger (though i didn't want that to be a solution), it didn't work.

      full debug traces are included in issue here: https://github.com/yfinkelstein/node-zookeeper/issues/6

      1. ZOOKEEPER-1057.patch
        1 kB
        Germán Blanco
      2. ZOOKEEPER-1057.patch
        1 kB
        Germán Blanco
      3. ZOOKEEPER-1057.patch
        1 kB
        Germán Blanco
      4. ZOOKEEPER-1057.patch
        4 kB
        Germán Blanco
      5. ZOOKEEPER-1057-b3.4.patch
        4 kB
        Germán Blanco
      6. ZOOKEEPER-1057.patch
        4 kB
        Michi Mutsuzaki
      7. ZOOKEEPER-1057.patch
        4 kB
        Michi Mutsuzaki


        No work has yet been logged on this issue.


          • Assignee:
            Michi Mutsuzaki
            Woody Anderson
          • Votes:
            2 Vote for this issue
            8 Start watching this issue


            • Created: