Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Plain View
Zookeeper, mail # user - trying to understand the keep alive code


Copy link to this message
-
trying to understand the keep alive code
Brian Tarbox 2013-02-27, 19:03
I've been trying to debug a problem related to getting session timeouts
when running an app that does tons of non-zookeeper related I/O.  That I/O
pressure has led to lots of zookeeper timeouts.  This has led me to try to
understand the timeout code.

I am puzzled because the timeout code seems far more complicated than it
might need to be...I'm not trying to critique the code so much as wondering
if I'm missing something.

Rather than simply spinning a thread that sends a ping every so often the
code seems to go to great lengths to avoid sending pings and even to avoid
calling System.currentTimeMillis (which it uses to determine if its really
time to send a ping).

While I'm a big fan of efficiency the resulting code seems
convoluted....help me see what I'm missing.

The main keep alive loop is basically the following code:

*int to = readTimeout - clientCnxnSocket.getIdleRecv();
int timeToNextPing = (readTimeout / 2)  -
clientCnxnSocket.getIdleSend();
if (timeToNextPing <= 0) {
            sendPing();
            clientCnxnSocket.updateLastSend();
 } else {
           if (timeToNextPing < to)
                     to = timeToNextPing;
}
clientCnxnSocket .doTransport(to); // blocks for "to" ms in a select, then
processes I/O and can take tens of seconds*

*
*
There is added complexity in that getIdleSend and getIdleRecv maintain
separate counters of how long its been since a send or receive...but only
relative to "now" which represents the current time...which is in turn
updated by various methods when various things happen.

I guess my question boils down to: is the system really that sensitive to
an extra call to check the time or an extra ping packet every few seconds
that the code in ClientCnxn.java is justified?   I ask because I've seen

a) "now" get significantly skewed from the actual time
b) under heavy I/O pressure doTransport takes 20, 30 or 40 seconds, causing
session timeouts
c) increasing the client timeout mostly doesn't help because the code
increases "to" based on the timeout

Again, sorry to be critical but I'm really asking in the spirit of trying
to understand.

Thanks.

http://about.me/BrianTarbox