Coprocessors

From Google's Jeff Dean, in a keynote to LADIS 2009 (http://www.scribd.com/doc/21631448/Dean-Keynote-Ladis2009, slides 66 - 67):

BigTable Coprocessors (New Since OSDI'06)

• Arbitrary code that runs run next to each tablet in table
  • As tablets split and move, coprocessor code automatically splits/moves too

• High-level call interface for clients
  • Unlike RPC, calls addressed to rows or ranges of rows

• coprocessor client library resolves to actual locations
  • Calls across multiple rows automatically split into multiple parallelized RPCs

• Very flexible model for building distributed services
  • Automatic scaling, load balancing, request routing for apps

Example Coprocessor Uses

• Scalable metadata management for Colossus (next gen GFS-like file system)

• Distributed language model serving for machine translation system

• Distributed query processing for full-text indexing support

• Regular expression search support for code repository

For HBase, adding a coprocessor framework will allow for pluggable incremental addition of functionality. No more need to subclass the regionserver interface and implementation classes and set hbase.regionserver.class and hbase.regionserver.impl in hbase-site.xml. That mechanism allows for extension but at the exclusion of all others.

Also in HBASE-2001 currently there is a in-process map reduce framework for the regionservers. Coprocessors can optionally implement a 'MapReduce' interface which clients will be able to invoke concurrently on all regions of the table. Note this is not MapReduce on the table; this is MapReduce on each region, concurrently. One can implement MapReduce in a manner very similar to Hadoop's MR framework, or use shared variables to avoid the overhead of generating (and processing) a lot of intermediates. An initial application of this could be support for rapid calculation of aggregates over data stored in HBase.