[HADOOP-9898] Set SO_KEEPALIVE on all our sockets

• Type: Bug
• Status: Closed
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 3.0.0-alpha1
• Fix Version/s: 2.3.0
• Component/s: ipc, net
• Target Version/s: 2.1.1-beta
• Assignee: Todd Lipcon
• Reporter: Todd Lipcon
• Created: 22/Aug/13 21:16
• Updated: 12/May/16 18:22
• Resolved: 31/Oct/13 20:57

Description

We recently saw an issue where network issues between slaves and the NN caused ESTABLISHED TCP connections to pile up and leak on the NN side. It looks like the RST packets were getting dropped, which meant that the client thought the connections were closed, while they hung open forever on the server.

Setting the SO_KEEPALIVE option on our sockets would prevent this kind of leak from going unchecked.

Issue Links

• relates to: HDFS-8894 Set SO_KEEPALIVE on DN server sockets