Connection loss causes infinite loop

[DavidVujic]

Connection loss causes infinite loop
When a connection loss occur, and the library reconnects, there is a scenario that causes the library to get stuck in an infinite loop between the zk_io_cb and yield functions in
the C++ wrapper src/node-zk.cpp.

To Reproduce
Steps to reproduce the behavior:

1. run the ZooKeeper server in a docker container
2. run a Node.js ZooKeeper client in a different docker container.
3. The Node.js client should have watchers on one or more nodes
4. run docker stop on the ZooKeeper server, and wait a couple of seconds.
5. Start the server again with docker start

This has only been reproduced in a docker container with a Linux OS.

Error Log example
A large number of log rows like this:
ZOO_ERROR@zk_io_cb@353: yield:zookeeper_process returned error: -117 - (not error) no server responses to process

Expected behavior
The library should emit a close event.

Desktop:

• OS: CentOS
• Node.js version: 10
• C/C++ compiler and version: gcc 7

[DavidVujic]

As I understand it, the C client (zookeeper.c) returns the status ZNOTHING when a socket read returns 0 (zero). The C++ wrapper does not handle that scenario, and will re-run the yield function.

A possible solution:

if (rc != ZOK) {
     LOG_ERROR(("yield:zookeeper_process returned error: %d - %s\n", rc, zerror(rc)));
            
     // Explicitly check for the return code, and close the session.
     // That will in turn emit a close event to the Node.js code
     if (rc == ZNOTHING) {
         zk->realClose(ZOO_EXPIRED_SESSION_STATE);
         return;
      }
}

[DavidVujic]

@mdlavin @kuebk Do you have any input on this? My C/C++ skills are very limited 😄

[kuebk]

Code seems to be ok, but I'm not sure should we close using ZOO_EXPIRED_SESSION_STATE, what about ZOO_CLOSED_STATE?

[DavidVujic]

I didn't find a ZOO_CLOSED_STATE in the wrapper code. Maybe it should be added in a separate PR @kuebk?