[
https://issues.apache.org/jira/browse/APEXCORE-379?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
David Yan updated APEXCORE-379:
-------------------------------
Description:
When the operator suddenly gets stuck, the latency value will also gets stuck,
until it is more than THROUGHPUT_CALCULATION_MAX_SAMPLES (default 1000) windows
behind, at which point STRAM estimates the latency by window width *
windowsBehind. But for 1000 windows, the latency for that operator is stuck
and hence the application latency is also based on stale value.
The way to fix this is for STRAM to estimate the latency when STRAM has not
received the end window stats for that operator for a certain amount of time.
was:
When the operator suddenly gets stuck, the latency value will also gets stuck,
until it is more than THROUGHPUT_CALCULATION_MAX_SAMPLES (default 1000) windows
behind, at which point STRAM estimates the latency by window width *
windowsBehind. But for 1000 windows, the latency for that operator is stuck
and hence the application latency is also based on stale value.
The way to fix this is for STRAM to estimate the latency when STRAM has not
received the end window stats for that operator for a certain amount of time.
Also, we should add to the REST API whether an operator latency is estimated.
> Latency is stuck when an operator is stuck (before the 1000 window threshold)
> -----------------------------------------------------------------------------
>
> Key: APEXCORE-379
> URL: https://issues.apache.org/jira/browse/APEXCORE-379
> Project: Apache Apex Core
> Issue Type: Improvement
> Reporter: David Yan
> Assignee: David Yan
>
> When the operator suddenly gets stuck, the latency value will also gets
> stuck, until it is more than THROUGHPUT_CALCULATION_MAX_SAMPLES (default
> 1000) windows behind, at which point STRAM estimates the latency by window
> width * windowsBehind. But for 1000 windows, the latency for that operator
> is stuck and hence the application latency is also based on stale value.
> The way to fix this is for STRAM to estimate the latency when STRAM has not
> received the end window stats for that operator for a certain amount of time.
>
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
