If worker to broker communication breaks down, the worker will sleep for refresh.leader.backoff.ms before throwing an error, at which point normal spark task retry (spark.task.maxFailures) comes into play.
If driver to broker communication breaks down, the driver will sleep for refresh.leader.backoff.ms before retrying the attempt to get offsets, up to spark.streaming.kafka.maxRetries number of times. The actual leader rebalancing process is entirely up to Kafka, which is why I'm saying if you're losing leaders, you should look at Kafka. On Fri, Apr 29, 2016 at 11:21 AM, swetha kasireddy <swethakasire...@gmail.com> wrote: > OK. So the Kafka side they use rebalance.backoff.ms of 2000 which is a > default for rebalancing and they say that refresh.leader.backoff.ms of 200 > to refresh leader is very aggressive and suggested us to increase it to > 2000. Even after increasing to 2500 I still get Leader Lost Errors. > > Is refresh.leader.backoff.ms the right setting in the app for it to wait > till the leader election and rebalance is done from the Kafka side assuming > that Kafka has rebalance.backoff.ms of 2000 ? > > Also, does Spark Kafka Direct try to restart the app when the leader is lost > or it will just wait till refresh.leader.backoff.ms and then retry again > depending on the number of retries? > > On Fri, Apr 29, 2016 at 8:14 AM, swetha kasireddy > <swethakasire...@gmail.com> wrote: >> >> OK. So the Kafka side they use rebalance.backoff.ms of 2000 which is a >> default for rebalancing and they say that refresh.leader.backoff.ms of 200 >> to refresh leader is very aggressive and suggested us to increase it to >> 2000. Even after increasing to 2500 I still get Leader Lost Errors. >> >> Is refresh.leader.backoff.ms the right setting in the app for it to wait >> till the leader election and rebalance is done from the Kafka side assuming >> that Kafka has rebalance.backoff.ms of 2000 ? >> >> On Wed, Apr 27, 2016 at 11:05 AM, Cody Koeninger <c...@koeninger.org> >> wrote: >>> >>> Seems like it'd be better to look into the Kafka side of things to >>> determine why you're losing leaders frequently, as opposed to trying >>> to put a bandaid on it. >>> >>> On Wed, Apr 27, 2016 at 11:49 AM, SRK <swethakasire...@gmail.com> wrote: >>> > Hi, >>> > >>> > We seem to be getting a lot of LeaderLostExceptions and our source >>> > Stream is >>> > working with a default value of rebalance.backoff.ms which is 2000. I >>> > was >>> > thinking to increase this value to 5000. Any suggestions on this? >>> > >>> > Thanks! >>> > >>> > >>> > >>> > -- >>> > View this message in context: >>> > http://apache-spark-user-list.1001560.n3.nabble.com/What-is-the-default-value-of-rebalance-backoff-ms-in-Spark-Kafka-Direct-tp26840.html >>> > Sent from the Apache Spark User List mailing list archive at >>> > Nabble.com. >>> > >>> > --------------------------------------------------------------------- >>> > To unsubscribe, e-mail: user-unsubscr...@spark.apache.org >>> > For additional commands, e-mail: user-h...@spark.apache.org >>> > >> >> > --------------------------------------------------------------------- To unsubscribe, e-mail: user-unsubscr...@spark.apache.org For additional commands, e-mail: user-h...@spark.apache.org