What are some tips for configuring HBase?

   

Jeff Hammerbacher 8 endorsements

2 votes by Oleksiy Kovyrin and Alex Kamil

Much of this content is taken from the  HBase Overview [1] and the HBase Default Configuration [2] .

System

• Increase the default per-process file handle limit [3] in 
  /etc/security/limits.conf

HDFS

• Set dfs.datanode.max.xceivers to 2047 [4] in 
  $HADOOP_HOME/conf/hdfs-site.xml

• Set dfs.datanode.socket.write.timeout to 0 [5]

HBase

• First, note that the default configuration values are stored atsrc/main/resources/hbase-default.xml [6] in the source tree
• For your site-specific configuration values, edit conf/hbase-site.xml
• Set hbase.rootdir to point to the directory in HDFS where HBase will put its data; e.g. 
  hdfs://localhost:9000/hbase

Per-Cluster

• hfile.block.cache.size controls the amount of region server heap space to devote to the block cache. Currently defaults to 20%.

Per-Table

• Max File Size: for clusters with lots of data, can be tuned up to 1 GB to result in less regions on the cluster.
• MemStore Flush Size

Per-Family

• Compression
• Bloom filters

Per-Region Server

• hbase.regionserver.global.memstore.upperLimit is used to cap the amount of heap room in each region server to reserve for all MemStores [7] served by that region. It defaults to 40% of the heap. 
• hbase.hregion.memstore.flush.size is the threshold for deciding when to flush a single MemStore to disk. It defaults to 64 MB.
• hbase.hregion.memstore.block.multiplier controls when to start blocking writes to keep the MemStore size sane. It defaults to 2 (multiplied by the memstore.flush.size). For production clusters with lots of RAM that you monitor closely, you can up to something like 8.
• hbase.hregion.max.filesize determines how big a StoreFile is allowed to grow before splitting a region. Defaults to 256 MB.

Per-Store

• hbase.hstore.blockingStoreFiles determines the maximum number of StoreFiles per Store to allow before blocking writes and forcing a compaction. The default is 7, but in production clusters monitored closely, it may make sense to up to 15.

[1]  http://hbase.apache.org/docs/cur...
[2]  http://hbase.apache.org/docs/cur...
[3]  http://wiki.apache.org/hadoop/Hb...
[4]  http://wiki.apache.org/hadoop/Hb...
[5]  http://hbase.markmail.org/thread...
[6]  http://svn.apache.org/viewvc/hba...
[7]  http://hbase.apache.org/docs/cur...

Wed Jun 09 2010 04:37:14 GMT+0800 (China Standard Time)