Avoid file system operations in BPOfferService thread while processing deletes
------------------------------------------------------------------------------
Key: HDFS-2500
URL: https://issues.apache.org/jira/browse/HDFS-2500
Project: Hadoop HDFS
Issue Type: Improvement
Components: data-node
Affects Versions: 0.23.0
Reporter: Todd Lipcon
Assignee: Todd Lipcon
Attachments: hdfs-2500.txt
While running a workload with concurrent writes and deletes, I saw a lot of
NotReplicatedYetExceptions being thrown due to late arrivals of blockReceived
reports from the DN. Looking at the DN logs, I found that the blockReceived
message was being delayed as much as 15 seconds because the OfferService thread
was blocked on file system operations processing deletes. We previously moved
the deletions to another thread, but it still accesses the file system to
determine the block length in the main thread. On a heavily loaded system this
can take a long time.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
