hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Wei-Chiu Chuang (JIRA)" <j...@apache.org>
Subject [jira] [Comment Edited] (HDFS-14084) Need for more stats in DFSClient
Date Thu, 06 Dec 2018 21:35:00 GMT

    [ https://issues.apache.org/jira/browse/HDFS-14084?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16712046#comment-16712046
] 

Wei-Chiu Chuang edited comment on HDFS-14084 at 12/6/18 9:34 PM:
-----------------------------------------------------------------

bq. I'm not sure on the best way to get the information out... logging (trace leve?) might
be too verbose but I don't see many more ways.

IMO logging latency number at trace level (for each op?) doesn't work too well -- based on
my experience trying to measure NameNode RPC latency.

For most part I am interested in the distribution of latency number. For example, 50%-tile,90%-tile,99%-tile,
of OP_DELETE, over some period of time, say the past 30 seconds, 5 minutes.

We already have something similar at RPC server level (via config key dfs.metrics.percentiles.intervals),
just that we don't have that in the client side.


was (Author: jojochuang):
bq. I'm not sure on the best way to get the information out... logging (trace leve?) might
be too verbose but I don't see many more ways.

IMO logging latency number at trace level (for each op?) doesn't work too well.

For most part I am interested in the distribution of latency number. For example, 50%-tile,90%-tile,99%-tile,
of OP_DELETE, over some period of time, say the past 30 seconds, 5 minutes.

We already have something similar at RPC server level (via config key dfs.metrics.percentiles.intervals),
just that we don't have that in the client side.

> Need for more stats in DFSClient
> --------------------------------
>
>                 Key: HDFS-14084
>                 URL: https://issues.apache.org/jira/browse/HDFS-14084
>             Project: Hadoop HDFS
>          Issue Type: Improvement
>    Affects Versions: 3.0.0
>            Reporter: Pranay Singh
>            Assignee: Pranay Singh
>            Priority: Minor
>         Attachments: HDFS-14084.001.patch
>
>
> The usage of HDFS has changed from being used as a map-reduce filesystem, now it's becoming
more of like a general purpose filesystem. In most of the cases there are issues with the
Namenode so we have metrics to know the workload or stress on Namenode.
> However, there is a need to have more statistics collected for different operations/RPCs
in DFSClient to know which RPC operations are taking longer time or to know what is the frequency
of the operation.These statistics can be exposed to the users of DFS Client and they can periodically
log or do some sort of flow control if the response is slow. This will also help to isolate
HDFS issue in a mixed environment where on a node say we have Spark, HBase and Impala running
together. We can check the throughput of different operation across client and isolate the
problem caused because of noisy neighbor or network congestion or shared JVM.
> We have dealt with several problems from the field for which there is no conclusive evidence
as to what caused the problem. If we had metrics or stats in DFSClient we would be better
equipped to solve such complex problems.
> List of jiras for reference:
> -------------------------
>  HADOOP-15538 HADOOP-15530 ( client side deadlock)



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-help@hadoop.apache.org


Mime
View raw message