[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [DISCUSS] Proposal of external shuffle service

Thanks for starting this design discussion Zhijiang!

I really like the idea to introduce a ShuffleService abstraction which
allows to have different implementations depending on the actual use case.
Especially for batch jobs I can clearly see the benefits of persisting the
results somewhere else.

Do you already know which interfaces we need to extend and where to
introduce new abstractions?


On Mon, Aug 27, 2018 at 1:57 PM Zhijiang(wangzhijiang999)
<wangzhijiang999@xxxxxxxxxx.invalid> wrote:

> Hi all!
> The shuffle service is responsible for transporting upstream produced data
> to the downstream side. In flink, the NettyServer is used for network
> transport service and this component is started in the TaskManager process.
> That means the TaskManager can support internal shuffle service which
> exists some concerns:
> 1. If a task finishes, the ResultPartition of this task still retains
> registered in TaskManager, because the output buffers have to be
> transported by internal shuffle service in TaskManager. That means the
> TaskManager can not be released by ResourceManager until ResultPartition
> released. It may waste container resources and can not support well for
> dynamic resource scenarios.
> 2. If we want to expand another shuffle service implementation, the
> current mechanism is not easy to handle, because the output level (result
> partition) and transport level (shuffle service) are not divided clearly
> and loss of abstraction to be extended.
> For above considerations, we propose the external shuffle service which
> can be deployed on any other external contaienrs, e.g. NodeManager
> container in yarn. Then the TaskManager can be released ASAP ifneeded when
> all the internal tasks finished. The persistent output files of these
> finished tasks can be served to transport by external shuffle service in
> the same machine.
> Further we can abstract both of the output level and transport level to
> support different implementations. e.g. We realized merging the data of all
> the subpartitions into limited persistent local files for disk improvements
> in some scenarios instead of one-subpartiton-one-file.
> I know it may be a big work for doing this, and I just point out some
> ideas, and wish getting any feedbacks from you!
> Best,
> Zhijiang