hadoop-common-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Varun Kapoor (Commented) (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HADOOP-8052) Hadoop Metrics2 should emit Float.MAX_VALUE (instead of Double.MAX_VALUE) to avoid making Ganglia's gmetad core
Date Thu, 09 Feb 2012 23:19:59 GMT

    [ https://issues.apache.org/jira/browse/HADOOP-8052?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13205025#comment-13205025
] 

Varun Kapoor commented on HADOOP-8052:
--------------------------------------

I chose to not change the data type housed in the MinMax class because that would, as is alluded
to by your helpful hint about clipping, be a much bigger change with a wider-reaching impact.


Instead, my justification for just changing the default values of 'max' and 'min' to Float.MIN_VALUE
and Float.MAX_VALUE, respectively, is:

a) It is the smallest required change that fixes both problems with gmetad I mentioned in
my original description.
b) It will likely be rare that we emit metrics outside of the range of [E-38:E+38], so the
regions on the number line outside of that range (that Double.MAX_VALUE and Double.MIN_VALUE
provide access to) are not really needed.

Also, I've left the types of the 2 new constants as 'double' in case we ever want to move
back to using Double's extremes in the future - my current fix signifies (or at least I hope
it does) that the *only* change we're making is in the default values of the 'min' and 'max'
fields.
                
> Hadoop Metrics2 should emit Float.MAX_VALUE (instead of Double.MAX_VALUE) to avoid making
Ganglia's gmetad core
> ---------------------------------------------------------------------------------------------------------------
>
>                 Key: HADOOP-8052
>                 URL: https://issues.apache.org/jira/browse/HADOOP-8052
>             Project: Hadoop Common
>          Issue Type: Bug
>          Components: metrics
>    Affects Versions: 0.23.0, 1.0.0
>            Reporter: Varun Kapoor
>            Assignee: Varun Kapoor
>              Labels: patch
>         Attachments: HADOOP-8052-branch-1.patch, HADOOP-8052-branch-1.patch, HADOOP-8052.patch,
HADOOP-8052.patch
>
>
> Ganglia's gmetad converts the doubles emitted by Hadoop's Metrics2 system to strings,
and the buffer it uses is 256 bytes wide.
> When the SampleStat.MinMax class (in org.apache.hadoop.metrics2.util) emits its default
min value (currently initialized to Double.MAX_VALUE), it ends up causing a buffer overflow
in gmetad, which causes it to core, effectively rendering Ganglia useless (for some, the core
is continuous; for others who are more fortunate, it's only a one-time Hadoop-startup-time
thing).
> The fix needed to Ganglia is simple - the buffer needs to be bumped up to be 512 bytes
wide, and all will be well - but instead of requiring a minimum version of Ganglia to work
with Hadoop's Metrics2 system, it might be more prudent to just use Float.MAX_VALUE.
> An additional problem caused in librrd (which Ganglia uses beneath-the-covers) by the
use of Double.MIN_VALUE (which functions as the default max value) is an underflow when librrd
runs the received strings through libc's strtod(), but the librrd code is good enough to check
for this, and only emits a warning - moving to Float.MIN_VALUE fixes that as well.

--
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

        

Mime
View raw message