[zz]hadoop hbase
HBase suffers terribly from the inability of applications to flush file data to storage before the file is closed, and a crash of any portion of the HBase servers or a service interrupting crash of HDFS will result in data loss.In prior chapters there was a discussion of problems caused by applications or server processes attempting to exceed the system-imposed limit on the number of open files; HBase also has this problem. The problem is substantially aggravated because each Hadoop MapFile is actually two files and a directory in HDFS, and each HDFS file also has a hidden checksum file. Setting the per-process open file count very large is a necessity for the HBase servers. A storage file format, HFile, is under development and due for Hbase version 0.20.0, and is expected to solve many of the performance and reliability issues.
HBase relies utterly on a smoothly performing HDFS for its operation; any stalls or DataNode instability will show up as HBase errors. There are HDFS tuning parameters suggested in the troubleshooting section on the HBase wiki: /Troubleshooting. In particular, if the underlying HDFS cluster is experiencing a slow block report problem, HADOOP-4584 , HBase is not recommended.
HBase servers, particularly the version using memcached, are memory intensive and generally require at least a gigabyte of real memory per server; any paging will drastically affect performance. Java Virtual Machine (JVM) garbage collection thread stalls are also causing HBase failures.
HBase generally provides downloadable release bundles that track the Hadoop Core distributions. HBase is not part of the Hadoop Core distribution.
页:
[1]