hadoop-common-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Steve Loughran (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HADOOP-13783) Improve efficiency of WASB over page blobs
Date Wed, 02 Nov 2016 12:35:58 GMT

    [ https://issues.apache.org/jira/browse/HADOOP-13783?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15628817#comment-15628817
] 

Steve Loughran commented on HADOOP-13783:
-----------------------------------------

1. telemetry is always good; maybe make a separate JIRA. FWIW, S3a will print out its metrics
on a toString() call of an FS instance or output stream, so letting downstream projects log
the stats on a call without making changes to their code which only works on 2.8+ to use the
storage stats code. Worth copying.

There's a class in the Swift FS client. {{org.apache.hadoop.fs.swift.util.DurationStats}}
which is used to track stats on HTTP calls, including min/max/variance; I used this to track
down throttling of DELETE requests against one swift service. That class could be moved into
Hadoop-common and used in WASB as well as other blobstores, where it can isolate problems
at the HTTPS level

2. Interesting to hear that azure has blob placement issues on key name too; so does S3 —the
traditional layout of hive-friendly datasets appears to be suboptimal for object stores. It'd
be good to help work out a layout policy which Hive can use that lays out datasets better

> Improve efficiency of WASB over page blobs
> ------------------------------------------
>
>                 Key: HADOOP-13783
>                 URL: https://issues.apache.org/jira/browse/HADOOP-13783
>             Project: Hadoop Common
>          Issue Type: Bug
>          Components: azure
>            Reporter: NITIN VERMA
>            Assignee: NITIN VERMA
>
> 1)	Add telemetry to WASB driver. WASB driver is lack of any log or telemetry which makes
trouble shoot very difficult. For example, we don’t know where is high e2e latency between
HBase and Azure storage came from when Azure storage server latency was very low. Also we
don’t know why WASB can only do 166 IOPs which is way below azure storage 500 IOPs. And
we had several incidents before related to storage latency, because of lacking logs, we couldn’t
find the ownership of the incident quickly.
> 2)	Resolving the hot spotting issue when WASB driver partition the azure page blobs by
changing the key. Current key design is causing the hot spotting on azure storage. 



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

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


Mime
View raw message