[jira] [Commented] (HIVE-18034) Improving logging with HoS executors spend lots of time in GC

["Sahil Takiar (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HIVE-18034?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16356393#comment-16356393 ] 

Sahil Takiar commented on HIVE-18034:
-------------------------------------

Looks like the Spark Web UI uses the following logic to detect if excessive GC time has been spent:

{code}
// totalDuration range from 0 to 50% GC time, alpha max = 1
function totalDurationAlpha(totalGCTime, totalDuration) {
    return totalDuration > 0 ?
        (Math.min(totalGCTime / totalDuration + 0.5, 1)) : 1;
}

// When GCTimePercent is edited change ToolTips.TASK_TIME to match
var GCTimePercent = 0.1;

function totalDurationStyle(totalGCTime, totalDuration) {
    // Red if GC time over GCTimePercent of total time
    return (totalGCTime > GCTimePercent * totalDuration) ?
        ("hsla(0, 100%, 50%, " + totalDurationAlpha(totalGCTime, totalDuration) + ")") : "";
{code}

We should be able to use the info in {{o.a.hive.spark.client.metrics}} to do the same calculation. {{executorRunTime}} + {{ executorDeserializeTime}} should give us the total executor duration, and {{jvmGCTime}} should give the GC time. I think a {{LOG#warn}} and a warning on the console should be printed.

> Improving logging with HoS executors spend lots of time in GC
> -------------------------------------------------------------
>
>                 Key: HIVE-18034
>                 URL: https://issues.apache.org/jira/browse/HIVE-18034
>             Project: Hive
>          Issue Type: Sub-task
>          Components: Spark
>            Reporter: Sahil Takiar
>            Priority: Major
>
> There are times when Spark will spend lots of time doing GC. The Spark History UI shows a bunch of red flags when too much time is spent in GC. It would be nice if those warnings are propagated to Hive.



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