Hi all,
Till started a discussion about FLIP-160: Declarative scheduler [1] earlier
today, the first major feature based on that effort will be FLIP-159:
Reactive Mode. It allows users to operate Flink in a way that it reactively
scales the job up or down depending on the provided resources: addin
Thanks a lot for the proposal, Robert and Till.
> No fixed parallelism for any of the operators
Regarding this limitation, can the scheduler only adjust the default
parallelism? if some operators set parallelism explicitly (like always 1),
just leave them unchanged.
On Fri, Jan 22, 2021 at 8:42
Thanks for preparing the FLIP and starting the discussion, Robert.
## Cluster vs. Job configuration
As I have commented on the FLIP-160 discussion thread [1], I'm a bit unsure
about activating the reactive execution mode via a cluster level
configuration option. I'm aware that in the first step th
Thanks Robert for creating this FLIP and starting the discussion.
This is a great start point to make Flink work with auto scaling service.
The reactive mode
is very useful in containerized environment(e.g. docker, Kubernetes). For
example, combined
with Kubernetes "Horizontal Pod Autoscaler"[1],
Thank you very much for the comments so far.
@Steven:
No fixed parallelism for any of the operators
>
> Regarding this limitation, can the scheduler only adjust the default
> parallelism? if some operators set parallelism explicitly (like always 1),
> just leave them unchanged.
We will respect
## configuration option
I see your point that autoscaling mode might be more suitable for session
clusters. It doesn't change that `execution-mode` could be a job-level
configuration. But I'm good with keeping it cluster-level and marking it
experimental at the moment, so we can change it later if
Thanks for your thoughts Xintong! What you are writing is very valuable
feedback for me, as I have limited experience with real-world deployments.
It seems that autoscaling support is a really important follow up.
## active resource managers
I guess you can consider reactive mode a special case o
Thanks for the explanation, Robert.
Now I see how these things are expected to be supported in steps.
I think you are right. Demanding a fixed finite amount of resources can be
considered as a special case of `ScalingPolicy`. I'm now good with the
current scope of reactive mode as a first step, a
Thanks a lot for all the feedback Steven, Yang Wang and Xintong. I have a
few more comments to add.
# Keep it simple and stupid
As Robert said we would like to keep the new feature initially as simple as
possible in order to quickly implement it. Once we have a basic
implementation, we want to re
Thanks Robert and Till for the thorough explanation.
Now I understand key difference between reactive mode and auto scaling mode.
For the latter, we could dynamically adjust the desired value based on
monitoring
the metrics(e.g. cpu, memory, latency, delay, etc.). Since the reactive is
simpler
and
The discussion has been open for 6 days, and it seems that all questions
and concerns raised so far have been addressed.
I will start a VOTE thread for FLIP-159 now.
On Tue, Jan 26, 2021 at 3:45 PM Yang Wang wrote:
> Thanks Robert and Till for the thorough explanation.
>
> Now I understand key
11 matches
Mail list logo