[
https://issues.apache.org/jira/browse/HDFS-11299?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15808360#comment-15808360
]
Xiaoyu Yao commented on HDFS-11299:
-----------------------------------
Thanks [~hanishakoneru] for working on this. The latest patch looks pretty good
to me. I just have few comments.
1. Metrics.md Line 328
`dfs.datanode.fileio.events.enabled.profiling` should be
`dfs.datanode.enable.fileio.profiling` to match with the code.
2. FaultInjectorFileIoEvents.java: checkstyle issues, looks like the
indentation is not setup properly in your editor.
3. FileIoProvider.java Line 819/821/823/898/900/902/959/961/963/1002/1004/1006
NIT: should we count read/write of a int as 4 bytes, i.e. using sizeof(int)
instead of 1 here to be consistent thought the len field is not being used by
the current provider implementation?
> Support multiple Datanode File IO hooks
> ---------------------------------------
>
> Key: HDFS-11299
> URL: https://issues.apache.org/jira/browse/HDFS-11299
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: hdfs
> Reporter: Hanisha Koneru
> Assignee: Hanisha Koneru
> Attachments: HDFS-11299.000.patch, HDFS-11299.001.patch
>
>
> HDFS-10958 introduces instrumentation hooks around DataNode disk IO and
> HDFS-10959 adds support for profiling hooks to expose latency statistics.
> Instead of choosing only one hook using Config parameters, we want to add two
> separate hooks - one for profiling and one for fault injection. The fault
> injection hook will be useful for testing purposes.
> This jira only introduces support for fault injection hook. The
> implementation for that will come later on.
> Also, now Default and Counting FileIOEvents would not be needed as we can
> control enabling the profiling and fault injection hooks using config
> parameters.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
