In every minor compaction job of HBase,
org.apache.phoenix.schema.stats.DefaultStatisticsCollector.initGuidePostDepth() is called,
and SYSTEM.CATALOG table is open to get guidepost width via
htable = env.getTable(
This function call creates one zookeeper connection to get cluster id.
DefaultStatisticsCollector doesn't close this zookeeper connection immediately after get guidepost width, and the zookeeper connection remains alive until HRegion is closed.
This is not a problem with small number of Regions, but when number of Region is large and upsert operation is frequent, the number of zookeeper connection gradually increases to hundreds, and the zookeeper server nodes experience short of available TCP/IP ports.
This zookeeper connection should be closed immediately after get guidepost width.