Hi all, we have been experiencing the same problem with 2 of our clusters. We are currently using HDP 2.1 that comes with HBase 0.98.
The problem manifested by showing a huge differences (hundreds of GB) between the output of "df" and "du" of the hdfs data directories. Eventually, other systems complained for the lack of space before shutting down. We identified the problem and discovered that all the RegionServers were holding lots of open file descriptors to deleted files, which prevented the OS to free the disk space occupied (hence the difference between "df" and "du"). The deleted files were pointing to the local HDFS blocks of old HFiles deleted from HDFS during the compaction and/or split operations. Apparently those file descriptors were stored by the HDFS ShortCircuit cache. My question is, isn't the shortcircuit feautre supposed to get "notified" somehow of file deletion on a file on HDFS so it can remove the open fds from the cache? This creates huge leaks whenever HBase is heavily loaded and we had to restart the RegionServer periodically until before identifying the problem. We solved the problem first by disabling shortcircuit from HDFS and then enabling it and reducing the cache size so to trigger often the caching policies (this leads to some performance loss). p.s. I am aware of the "dfs.client.read.shortcircuit.streams.cache.expiry.ms " directoparameter, but for some reason the default value (5 mins) does not work out-of-the-box on HDP 2.1, moreover the problem persists for high timeouts and big cache sizes. Kind Regards
