[HBASE-11267] Dynamic metrics2 metrics may consume large amount of heap memory

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 0.99.0, 0.98.3
• Hadoop Flags: Reviewed
• Assignee: Ted Yu
• Reporter: Ted Yu
• Created: 28/May/14 21:48
• Updated: 21/Feb/15 23:30
• Resolved: 31/May/14 01:27

Description

User sunweiei provided the following jmap output in 0.96 deployment:

 num     #instances         #bytes  class name
----------------------------------------------
   1:      14917882     3396492464  [C
   2:       1996994     2118021808  [B
   3:      43341650     1733666000  java.util.LinkedHashMap$Entry
   4:      14453983     1156550896  [Ljava.util.HashMap$Entry;
   5:      14446577      924580928  org.apache.hadoop.metrics2.lib.Interns$CacheWith2Keys$2

Heap consumption by Interns$CacheWith2Keys$2 could be due to calls to Interns.info() in DynamicMetricsRegistry.
This scenario would arise when large number of regions are tracked through metrics2.
Interns class doesn't provide API to remove entries in its internal Map.
Meaning, DynamicMetricsRegistry#clearMetrics() is unable to reclaim that part of heap.

Issue Links

• is related to: HBASE-11165 Scaling so cluster can host 1M regions and beyond (50M regions?)
• relates to: HBASE-10089 Metrics intern table names cause eventual permgen OOM in 0.94