HBase is using memory caches for the tables to reduce latency times. So dependent on the query, you will not see any disk I/O at all. How your data is distributed across the Hbase nodes depends on the 'sharding', which Hbase can do automatcally or you can define it during table creation. The operation you execute i.e. scan is executed on all RegionServers holding parts of the table in parallel. And to my understanding Yarn isn't balancing node resources between Hbase queries and MR jobs. In the cases I am aware of, HBase and yarn are configured to only use a share of the available hardware resources on the node, i.e. the RAM, to avoid issues.
... View more