[
https://issues.apache.org/jira/browse/FLINK-34266?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17811921#comment-17811921
]
Maximilian Michels commented on FLINK-34266:
--------------------------------------------
The way I understand the code is that for every observation, we will store the
total output rate of every vertex. During metric window evaluation, we will
average all of those. That is in line with how all the code works.
I agree 100% that all metrics should be observed over the entire metric window.
So rates should be computed by measuring the number of records produced at the
start and at the end up the window, then subtracting them from each other.
This request seems analogue to FLINK-34213 but for rates instead of busy time.
Is that fair to say?
> Output ratios should be computed over the whole metric window instead of
> averaged
> ---------------------------------------------------------------------------------
>
> Key: FLINK-34266
> URL: https://issues.apache.org/jira/browse/FLINK-34266
> Project: Flink
> Issue Type: Improvement
> Components: Autoscaler
> Reporter: Gyula Fora
> Priority: Critical
>
> Currently Output ratios are computed during metric collection based on the
> current in/out metrics an stored as part of the collected metrics.
> During evaluation the output ratios previously computed are then averaged
> together in the metric window. This however leads to incorrect computation
> due to the nature of the computation and averaging.
> Example:
> Let's look at a window operator that simply sorts and re-emits events in
> windows. During the window collection phase, output ratio will be computed
> and stored as 0. During the window computation the output ratio will be
> last_input_rate / window_size. Depending on the last input rate observation
> this can be off when averaged into any direction.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
