spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Holden Karau <hol...@pigscanfly.ca>
Subject Re: Spark on Kubernetes scheduler variety
Date Thu, 24 Jun 2021 15:56:18 GMT
That's awesome, I'm just starting to get context around Volcano but maybe
we can schedule an initial meeting for all of us interested in pursuing
this to get on the same page.

On Wed, Jun 23, 2021 at 6:54 PM Klaus Ma <klaus1982.cn@gmail.com> wrote:

> Hi team,
>
> I'm kube-batch/Volcano founder, and I'm excited to hear that the spark
> community also has such requirements :)
>
> Volcano provides several features for batch workload, e.g. fair-share,
> queue, reservation, preemption/reclaim and so on.
> It has been used in several product environments with Spark; if necessary,
> I can give an overall introduction about Volcano's features and those use
> cases :)
>
> -- Klaus
>
> On Wed, Jun 23, 2021 at 11:26 PM Mich Talebzadeh <
> mich.talebzadeh@gmail.com> wrote:
>
>>
>>
>> Please allow me to be diverse and express a different point of view on
>> this roadmap.
>>
>>
>> I believe from a technical point of view spending time and effort plus
>> talent on batch scheduling on Kubernetes could be rewarding. However, if I
>> may say I doubt whether such an approach and the so-called democratization
>> of Spark on whatever platform is really should be of great focus.
>>
>> Having worked on Google Dataproc <https://cloud.google.com/dataproc> (A fully
>> managed and highly scalable service for running Apache Spark, Hadoop and
>> more recently other artefacts) for that past two years, and Spark on
>> Kubernetes on-premise, I have come to the conclusion that Spark is not a
>> beast that that one can fully commoditize it much like one can do with
>> Zookeeper, Kafka etc. There is always a struggle to make some niche areas
>> of Spark like Spark Structured Streaming (SSS) work seamlessly and
>> effortlessly on these commercial platforms with whatever as a Service.
>>
>>
>> Moreover, Spark (and I stand corrected) from the ground up has already a
>> lot of resiliency and redundancy built in. It is truly an enterprise class
>> product (requires enterprise class support) that will be difficult to
>> commoditize with Kubernetes and expect the same performance. After all,
>> Kubernetes is aimed at efficient resource sharing and potential cost saving
>> for the mass market. In short I can see commercial enterprises will work on
>> these platforms ,but may be the great talents on dev team should focus on
>> stuff like the perceived limitation of SSS in dealing with chain of
>> aggregation( if I am correct it is not yet supported on streaming datasets)
>>
>>
>> These are my opinions and they are not facts, just opinions so to speak :)
>>
>>
>>    view my Linkedin profile
>> <https://www.linkedin.com/in/mich-talebzadeh-ph-d-5205b2/>
>>
>>
>>
>> *Disclaimer:* Use it at your own risk. Any and all responsibility for
>> any loss, damage or destruction of data or any other property which may
>> arise from relying on this email's technical content is explicitly
>> disclaimed. The author will in no case be liable for any monetary damages
>> arising from such loss, damage or destruction.
>>
>>
>>
>>
>> On Fri, 18 Jun 2021 at 23:18, Holden Karau <holden@pigscanfly.ca> wrote:
>>
>>> I think these approaches are good, but there are limitations (eg dynamic
>>> scaling) without us making changes inside of the Spark Kube scheduler.
>>>
>>> Certainly whichever scheduler extensions we add support for we should
>>> collaborate with the people developing those extensions insofar as they are
>>> interested. My first place that I checked was #sig-scheduling which is
>>> fairly quite on the Kubernetes slack but if there are more places to look
>>> for folks interested in batch scheduling on Kubernetes we should definitely
>>> give it a shot :)
>>>
>>> On Fri, Jun 18, 2021 at 1:41 AM Mich Talebzadeh <
>>> mich.talebzadeh@gmail.com> wrote:
>>>
>>>> Hi,
>>>>
>>>> Regarding your point and I quote
>>>>
>>>> "..  I know that one of the Spark on Kube operators
>>>> supports volcano/kube-batch so I was thinking that might be a place I would
>>>> start exploring..."
>>>>
>>>> There seems to be ongoing work on say Volcano as part of  Cloud Native
>>>> Computing Foundation <https://cncf.io/> (CNCF). For example through
>>>> https://github.com/volcano-sh/volcano
>>>>
>>> <https://github.com/volcano-sh/volcano>
>>>>
>>>> There may be value-add in collaborating with such groups through CNCF
>>>> in order to have a collective approach to such work. There also seems to
be
>>>> some work on Integration of Spark with Volcano for Batch Scheduling.
>>>> <https://github.com/GoogleCloudPlatform/spark-on-k8s-operator/blob/master/docs/volcano-integration.md>
>>>>
>>>>
>>>>
>>>> What is not very clear is the degree of progress of these projects. You
>>>> may be kind enough to elaborate on KPI for each of these projects and where
>>>> you think your contributions is going to be.
>>>>
>>>>
>>>> HTH,
>>>>
>>>>
>>>> Mich
>>>>
>>>>
>>>>    view my Linkedin profile
>>>> <https://www.linkedin.com/in/mich-talebzadeh-ph-d-5205b2/>
>>>>
>>>>
>>>>
>>>> *Disclaimer:* Use it at your own risk. Any and all responsibility for
>>>> any loss, damage or destruction of data or any other property which may
>>>> arise from relying on this email's technical content is explicitly
>>>> disclaimed. The author will in no case be liable for any monetary damages
>>>> arising from such loss, damage or destruction.
>>>>
>>>>
>>>>
>>>>
>>>> On Fri, 18 Jun 2021 at 00:44, Holden Karau <holden@pigscanfly.ca>
>>>> wrote:
>>>>
>>>>> Hi Folks,
>>>>>
>>>>> I'm continuing my adventures to make Spark on containers party and I
>>>>> was wondering if folks have experience with the different batch
>>>>> scheduler options that they prefer? I was thinking so that we can
>>>>> better support dynamic allocation it might make sense for us to
>>>>> support using different schedulers and I wanted to see if there are
>>>>> any that the community is more interested in?
>>>>>
>>>>> I know that one of the Spark on Kube operators supports
>>>>> volcano/kube-batch so I was thinking that might be a place I start
>>>>> exploring but also want to be open to other schedulers that folks
>>>>> might be interested in.
>>>>>
>>>>> Cheers,
>>>>>
>>>>> Holden :)
>>>>>
>>>>> --
>>>>> Twitter: https://twitter.com/holdenkarau
>>>>> Books (Learning Spark, High Performance Spark, etc.):
>>>>> https://amzn.to/2MaRAG9
>>>>> YouTube Live Streams: https://www.youtube.com/user/holdenkarau
>>>>>
>>>>> ---------------------------------------------------------------------
>>>>> To unsubscribe e-mail: dev-unsubscribe@spark.apache.org
>>>>>
>>>>> --
>>> Twitter: https://twitter.com/holdenkarau
>>> Books (Learning Spark, High Performance Spark, etc.):
>>> https://amzn.to/2MaRAG9  <https://amzn.to/2MaRAG9>
>>> YouTube Live Streams: https://www.youtube.com/user/holdenkarau
>>>
>>

-- 
Twitter: https://twitter.com/holdenkarau
Books (Learning Spark, High Performance Spark, etc.):
https://amzn.to/2MaRAG9  <https://amzn.to/2MaRAG9>
YouTube Live Streams: https://www.youtube.com/user/holdenkarau

Mime
View raw message