[
https://issues.apache.org/jira/browse/FLINK-15603?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Piotr Nowojski updated FLINK-15603:
-----------------------------------
Description:
One of the most important metrics is missing in the checkpoint stats: "start
delay" (aka "barrier lag"), meaning the time it between when the checkpoint was
triggered and when the barriers arrive at a task.
That time is critical to identify if a checkpoint takes too long because of
backpressure or other contention.
You can implicitly calculate this by "end_to_end_time - sync_time -
async_time", but it is much more obvious for users that something is up when
this number is explicitly shown.
was:
One of the most important metrics is missing in the checkpoint stats: "barrier
lag", meaning the time it between when the checkpoint was triggered and when
the barriers arrive at a task.
That time is critical to identify if a checkpoint takes too long because of
backpressure or other contention.
You can implicitly calculate this by "end_to_end_time - sync_time -
async_time", but it is much more obvious for users that something is up when
this number is explicitly shown.
> Show checkpoint start delay in checkpoint statistics
> ----------------------------------------------------
>
> Key: FLINK-15603
> URL: https://issues.apache.org/jira/browse/FLINK-15603
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Web Frontend
> Reporter: Stephan Ewen
> Assignee: Piotr Nowojski
> Priority: Critical
> Labels: pull-request-available, usability
> Fix For: 1.11.0
>
> Attachments: Screenshot 2020-01-23 at 15.43.26.png
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> One of the most important metrics is missing in the checkpoint stats: "start
> delay" (aka "barrier lag"), meaning the time it between when the checkpoint
> was triggered and when the barriers arrive at a task.
> That time is critical to identify if a checkpoint takes too long because of
> backpressure or other contention.
> You can implicitly calculate this by "end_to_end_time - sync_time -
> async_time", but it is much more obvious for users that something is up when
> this number is explicitly shown.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
