[FLUME-926] Memory leak in 0.9.4

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Won't Fix
• Affects Version/s: 0.9.4
• Fix Version/s: 0.9.5
• Component/s: Node
• Labels: memory_leak

Environment

Debian Squeeze
Sun Java 1.6.0_23 (32 or 64 depending on the collector)

• Assignee: Unassigned
• Reporter: Thomas Vachon
• Created: 11/Jan/12 14:15
• Updated: 05/Nov/14 11:21
• Resolved: 05/Nov/14 11:21

Description

Without setting Xmx to a sane value for EC2 (between 250-500M depending on the server). A Flume collector in autoCollector mode with 8 logical streams consumes all available server ram until it is gone. I have not been able to see any GC activity on the node/collector which leaves me to believe that flume is leaving objects in an unreapable state. I restarted one flume instance yesterday and within 12 hours I have seen it take 75M of ram for 6534 counts of the int[] object.

We have it configured with 8 logical collectors per a collector/node separated via flows with all source nodes in autoE2E mode. 4 flows write to HDFS, 4 flows write to S3.

I have been experiencing this every 2-3 days on the 32 bit machine (1.7 GB RAM) and every 10 or so days on the 64 bit machine (7 GB RAM).

I have forced Xmx on both machines, but without the GC process working, I have a feeling the will OOM and crash (maybe the watchdog will be able to restart them, but it is unknown at this time).