Created on 12-22-201603:28 PM - edited 09-16-202201:37 AM
Consider increasing network capacity to overcome
the challenge caused due to non locality of blocks.
Create configuration groups of datanodes exclusively for HBASE,
disabling HDFS balancer on this group and allow only hbase balancer. Follow
this url Host_Config_Groups to create host config groups.
Few temporary workarounds can also be applied if problem is
severe and need immediate attention :-
Disable HDFS balancer permanently on the cluster and run it
manually on need basis. (Please spin a support case and have the situation
discussed before implementing this workaround.)
In case the performance issue needs to be fixed post running of
HDFS Balancer, a major compaction could be manually initiated. For performance
gains, major compaction is run on off peak hours such as weekends. This article Compaction_Best_Practices is a recommended read here.
Scheduling major compaction after scheduled balancer rather than
vice versa.
HDFS although has introduced concept of "favored
nodes" feature but HBase APIs are not yet equipped to choose specific
nodes during data writing.
Please note that these are expert level configurations and
procedures, if unsure of their implications, its always recommended to open a
support case with us.
Refer following Apache URLs to track the progress of region
blocks pinning implementation.