Hi Bo, The additional replica could be an intermediate state. For example, if a replica will be moved to somewhere else, Helix needs to boot up the new replication before drop the old one. During this process, there could be more replicas than the config. Moreover, the error state transition might block normal load balance that cleans up additional replicas. Could you please confirm if it is transitional or a stable status? Then please also verify if there is any error partition in the resource.
About load balance issue, please try our new CRUSH based strategy "CRUSH-ed", which is designed for even distribution. It is available after version 0.8.0. Hope this helps. Thanks. Best Regards, Jiajun On Fri, May 18, 2018 at 3:53 PM, Bo Liu <[email protected]> wrote: > The cluster level configuration is as below. > > allowParticipantAutoJoin true > > DELAY_REBALANCE_DISABLE false > > DELAY_REBALANCE_ENABLE true > > DELAY_REBALANCE_TIME 600000 > > FAULT_ZONE_TYPE pg > > MAX_OFFLINE_INSTANCES_ALLOWED 10 > > TOPOLOGY /az/pg/instance > > TOPOLOGY_AWARE_ENABLED true > > > On Fri, May 18, 2018 at 3:51 PM, Bo Liu <[email protected]> wrote: > >> Hi folks, >> >> We are running Helix in FULL_AUTO mode with the following configurations. >> The resource is configured to have 4 replicas. However, we noticed that a >> few partitions actually get 5 replicas (shown in helix UI, and we checked >> them on the hosts). >> And we have a few hosts which don't host any partitions. >> >> We tried to rebalance the resource through Helix restful API and got no >> luck. >> Could you please provide some inputs? >> >> Thanks, >> >> >> >> IDEAL_STATE_MODE AUTO_REBALANCE >> >> MIN_ACTIVE_REPLICAS 2 >> >> NUM_PARTITIONS 1500 >> >> REBALANCE_MODE FULL_AUTO >> >> REBALANCE_STRATEGY org.apache.helix.controller.re >> balancer.strategy.MultiRoundCrushRebalanceStrategy >> >> REBALANCER_CLASS_NAME org.apache.helix.controller.re >> balancer.DelayedAutoRebalancer >> >> REPLICAS 4 >> >> STATE_MODEL_DEF_REF MasterSlave >> >> STATE_MODEL_FACTORY_NAME DEFAULT >> >> >> >> >> -- >> Best regards, >> Bo >> >> > > > -- > Best regards, > Bo > >
