[jira] [Comment Edited] (AURORA-1827) Fix SLA percentile calculation

["Reza Motamedi (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/AURORA-1827?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15700339#comment-15700339 ] 

Reza Motamedi edited comment on AURORA-1827 at 11/27/16 10:17 PM:
------------------------------------------------------------------

based on the comment description in the SlaUtils::percentile, it reports the percentile value from the given list ordered in a non-descending order. My concern is with the provided example. 
   * Example: [30, 60, 70, 90], the 75 percentile is 30 (i.e. 75% of elements are greater).
The current implementation works correctly according to this example. However, if we assume that assume that these sampled values represent a continuous distribution the 75 percentile 52.5.

I believe the second number captures the underlying distribution better. For instance, having 99 percentile as the input, the current implementation again reports 30, and the second calculation reports 30.9.



was (Author: rezam):
based on the comment description in slaUtils the SlaUtils::percentile reports the percentile value from the given list ordered in a non-descending order. My concern is with the provided example. 
   * Example: [30, 60, 70, 90], the 75 percentile is 30 (i.e. 75% of elements are greater).
The current implementation works correctly according to this example. However, if we assume that assume that these sampled values represent a continuous distribution the 75 percentile 52.5.

I believe the second number captures the underlying distribution better. For instance, having 99 percentile as the input, the current implementation again reports 30, and the second calculation reports 30.9.


> Fix SLA percentile calculation 
> -------------------------------
>
>                 Key: AURORA-1827
>                 URL: https://issues.apache.org/jira/browse/AURORA-1827
>             Project: Aurora
>          Issue Type: Story
>            Reporter: Reza Motamedi
>            Assignee: Reza Motamedi
>            Priority: Trivial
>              Labels: newbie, sla
>
> The calculation of mttX (median-time-to-X) depends on the computation of percentile values. The current implementation does not behave nicely with a small sample size. For instance, for a given sample set of  {50, 150}, 50-percentile is reported to be 50. Although, 100 seems a more appropriate return value.
> One solution is to modify `SlaUtil` to perform an extrapolation when the sample size is small or when the corresponding index to a percentile value is not an integer. 



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