I did a small test, and I a setting mapred.tasktracker.expiry.interval=60000 worked for me (TT became considered as lost after around 66 seconds).
Can the formula be: mapred.tasktracker.expiry.interval + 2 * some-heartbeat-interval-that-is-3-sec-by-default? Otherwise, is the 6 sec some kind of time needed to make a decision to consider TT as lost? 2013/12/3 Hansi Klose <[email protected]> > I forget to say that we use Cloudera 2.0.0-mr1-cdh4.2.0 > > > Gesendet: Dienstag, 03. Dezember 2013 um 17:38 Uhr > > Von: "Hansi Klose" <[email protected]> > > An: [email protected] > > Betreff: mapreduce.jobtracker.expire.trackers.interval no effect > > > > Hi, > > > > we want to set the heartbeat timout for a tasktracker. > > > > If the tasktracker does not send heartbeats for 60 seconds he should > > be marked as lost. > > > > I found the parameter mapreduce.jobtracker.expire.trackers.interval > > which sounds right to me. > > > > I set > > > > <property> > > <name>mapreduce.jobtracker.expire.trackers.interval</name> > > <value>60000</value> > > </property> > > > > in the mapred-site.xml on all servers and restarted the jobtracker and > all tasktrackers. > > > > I started a benchmark "hadoop jar hadoop-examples.jar randomwriter rand" > and every tasktracker gets 2 jobs. > > It is a small test environment. > > > > On one tasktracker i stopped the network. On the jobtracker i could see > the "Seconds since heartbeat" > > increasing. But after 60 seconds the tasktracker was still in the > overview. > > Even in the log of the jobtracker I found nothing. > > > > After over 600 seconds i found the message > > org.apache.hadoop.mapred.JobTracker: Lost tracker ..... > > And the tasktracker wasn't shown any more on the jobtracker. > > > > Isn't this the right setting? > > > > Regards Hansi > > >
