He Xiaoqiao created HDFS-14511:
----------------------------------
Summary: FSEditlog write both Quorum Journal and Local disk by
default in HA using QJM scenario
Key: HDFS-14511
URL: https://issues.apache.org/jira/browse/HDFS-14511
Project: Hadoop HDFS
Issue Type: Bug
Components: namenode, qjm
Reporter: He Xiaoqiao
Assignee: He Xiaoqiao
Recently, I meet case about FSEditLog in HA using QJM scenario. NameNode enter
suspended state and can not process other RPC requests any more.
The root cause is load of local disk is very high, it will block edit log
recored flush local, then following RPC request will occupy all RPC handlers
since #FSEditLog write edit log record to both FileJournal (which is local
directory located at the same as FsImage) and QuorumJournal in proper order by
default and no configuration to switch off FileJournal. However local edit log
is not used any time soon.
More detailed information, the location where edit log write to is decided by
configuration items 'dfs.namenode.shared.edits.dir' and 'dfs.namenode.name.dir'
(since 'dfs.namenode.edits.dir' is deprecated item, if not set it will be
overrided/replaced by 'dfs.namenode.name.dir' where fsimage located.) by
default. So JournalSet = QuorumJournal (SharedEditsDirs, set by
'dfs.namenode.shared.edits.dir') + FileJournal (LocalStorageEditsDirs, set by
'dfs.namenode.name.dir' by default). Another side, these two config items have
to set in HA with QJM.
In one word, edit log is double write to both QJM and local disk by default and
no way to turn off local write with current implementation. I propose we should
offer some choice or turn off local edit log write by default in HA using QJM
for users.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-dev-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-dev-h...@hadoop.apache.org
