[ 
https://issues.apache.org/jira/browse/FLINK-26790?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17513247#comment-17513247
 ] 

Piotr Nowojski commented on FLINK-26790:
----------------------------------------

I think [~assassinj]'s confusion is about the picture. On the first figure 
captured "on first barrier", the checkpoint barrier for the lower connection is 
in the middle of the output buffers (between buffers "e" and "d"), which 
contradicts what is described below:
{quote}
by adding it to the end of the output buffers.
{quote}
I think indeed the graphic should be changed somehow.

I think I understand the [~arvid]'s intention behind it, that it was supposed 
to depict a scenario where two checkpoint barriers from two different upstream 
operators are arriving at different time, and the second channel is keep 
persisting the in-flight data until it receives it's checkpoint barrier, but 
indeed that's misleading as it is. Especially that it shows two operators 
connected to each other via two channels? Which apart of some esoteric scenario 
with for example union-ing output and side output of the same operator is 
impossible to achieve.

I would either simplify this graph, and focus on single input/output channel 
case (parallelism = 1), or keep it with upstream parallelism = 2, but then:
# Clearly adding one more upstream operator
# Adding one more extra step, where in the first figure only first upstream 
operator inject checkpoint barrier, and in the next figure second upstream 
operator is doing it

WDYT?

> Description image about Unaligned Checkpointing maybe incorrect.
> ----------------------------------------------------------------
>
>                 Key: FLINK-26790
>                 URL: https://issues.apache.org/jira/browse/FLINK-26790
>             Project: Flink
>          Issue Type: Improvement
>          Components: Documentation
>    Affects Versions: 1.13.6, 1.14.4
>            Reporter: jinghaihang
>            Priority: Minor
>         Attachments: image-2022-03-22-11-12-20-025.png
>
>
> Due to the second description :
>  * It immediately forwards the barrier to the downstream operator by adding 
> it to the end of the output buffers.
> so i think the picture maybe misleading readers that barrier will not  appear 
> in the middle of the data stream.
>  
> !image-2022-03-22-11-12-20-025.png|width=599,height=303!



--
This message was sent by Atlassian Jira
(v8.20.1#820001)

Reply via email to