You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@flink.apache.org by Longda Feng <zh...@alibaba-inc.com> on 2016/06/29 07:18:22 UTC
[Discuss] Why different job's tasks can run in the single process.
hi ,
Sorry for asking the quest here? Any answer will be apprecated.
Why different job's tasks can run in the single process. (There are some different job's tasks in one TaskManager).It seems Flink-on-Yarn can let different job run on different process. But for standalone mode, this problem still exists.
Why design Flink like this?The advantage What I can thought is as following:(1) All task can share bigger memory pool.(2) The communication between the tasks in the same process will be fast.
But this design will impact to the stability. Flink provide User-Define-Function interface, if one of the User-Define-Function crash, It maybe crack the whole JVM, If the TaskManager crash, all other job's task in this TaskManager will be impacted. Even if the JVM don't crash, but maybe lead to some other unexpected problem, what's more this will make the code too sophisticated。Normal framework like Spark/Storm/Samza won't run different job's tasks in the same process。As one normal user, stability has the highest priority.
ThanksLongda
Re: [Discuss] Why different job's tasks can run in the single process.
Posted by Longda Feng <zh...@alibaba-inc.com>.
What I means is standalone mode is not suitable for production enironment. It is just for testing or prototype.
regardsLongda
------------------------------------------------------------------From:Kevin Jacobs <ke...@cern.ch>Send Time:2016年6月30日(星期四) 16:20To:dev <de...@flink.apache.org>; 封仲淹(纪君祥) <zh...@alibaba-inc.com>Subject:Re: [Discuss] Why different job's tasks can run in the single process.
In my opinion the streaming process can be perfectly simulated on a
single node. You can setup a message distribution system like Kafka on a
single node, you can run Spark on a single node and the only thing you
need to change when running it on a cluster is that you need to change
the environment. So there is no need to setup a cluster when testing the
streaming process.
Regards,
Kevin
On 30-06-16 09:54, Longda Feng wrote:
>
> This means Standalone mode is just for prototype.
> But I think we need a lightweight solution for streaming process, standalone is the best solution. Some times, we need setup a flink cluster on a small cluster. setup a yarn cluster isn't convenient.
> (1) in small company, the number of machine is small (2) When a data center is small, but we still need do some computing in this data center(3) some machines are in the while-list, they have been authorited to access some special data or machine, but the number of these machine is small.(4) some machine has critical data, they can't be shared with others, but the number of these machine is small.(5) when a team start to learn flink, he will setup a small cluster firstly, maybe he wo't want to setup a huge system, perfer to a small system
>
>
> regardsLongda
>
> ------------------------------------------------------------------From:Aljoscha Krettek <al...@apache.org>Send Time:2016年6月29日(星期三) 21:48To:封仲淹(纪君祥) <zh...@alibaba-inc.com>; dev <de...@flink.apache.org>Subject:Re: [Discuss] Why different job's tasks can run in the single process.
> Hi,
> yes, you are definitely right that allowing to run multiple user code tasks
> in the same TaskManager JVM is not good for stability. This mode is still
> there from the very early days of Flink where Yarn was not yet available.
> In a production environment I would now recommend to always run one
> Flink-Yarn cluster per job to get good isolation between different jobs.
>
> Cheers,
> Aljoscha
>
> On Wed, 29 Jun 2016 at 09:18 Longda Feng <zh...@alibaba-inc.com>
> wrote:
>
>> hi ,
>> Sorry for asking the quest here? Any answer will be apprecated.
>> Why different job's tasks can run in the single process. (There are some
>> different job's tasks in one TaskManager).It seems Flink-on-Yarn can let
>> different job run on different process. But for standalone mode, this
>> problem still exists.
>> Why design Flink like this?The advantage What I can thought is as
>> following:(1) All task can share bigger memory pool.(2) The communication
>> between the tasks in the same process will be fast.
>> But this design will impact to the stability. Flink provide
>> User-Define-Function interface, if one of the User-Define-Function crash,
>> It maybe crack the whole JVM, If the TaskManager crash, all other job's
>> task in this TaskManager will be impacted. Even if the JVM don't crash, but
>> maybe lead to some other unexpected problem, what's more this will make the
>> code too sophisticated。Normal framework like Spark/Storm/Samza won't run
>> different job's tasks in the same process。As one normal user, stability has
>> the highest priority.
>>
>> ThanksLongda
>>
>>
>>
>>
Re: [Discuss] Why different job's tasks can run in the single process.
Posted by Aljoscha Krettek <al...@apache.org>.
I'm not aware of any plans to change this. It would be possible though, to
add a mode where a standalone cluster only accepts a single job.
On Fri, 1 Jul 2016 at 04:50 Jark Wu <wu...@alibaba-inc.com> wrote:
> As standalone mode has the disadvantage that the TaskManager JVM can’t
> isolate different jobs. Does we have any plan to improve this ?
>
> - Jark Wu
>
> > 在 2016年6月30日，下午4:19，Kevin Jacobs <ke...@cern.ch> 写道：
> >
> > In my opinion the streaming process can be perfectly simulated on a
> single node. You can setup a message distribution system like Kafka on a
> single node, you can run Spark on a single node and the only thing you need
> to change when running it on a cluster is that you need to change the
> environment. So there is no need to setup a cluster when testing the
> streaming process.
> >
> > Regards,
> > Kevin
> >
> > On 30-06-16 09:54, Longda Feng wrote:
> >>
> >> This means Standalone mode is just for prototype.
> >> But I think we need a lightweight solution for streaming process,
> standalone is the best solution. Some times, we need setup a flink cluster
> on a small cluster. setup a yarn cluster isn't convenient.
> >> (1) in small company, the number of machine is small (2) When a data
> center is small, but we still need do some computing in this data center(3)
> some machines are in the while-list, they have been authorited to access
> some special data or machine, but the number of these machine is small.(4)
> some machine has critical data, they can't be shared with others, but the
> number of these machine is small.(5) when a team start to learn flink, he
> will setup a small cluster firstly, maybe he wo't want to setup a huge
> system, perfer to a small system
> >>
> >>
> >> regardsLongda
> >>
> >>
> ------------------------------------------------------------------From:Aljoscha
> Krettek <al...@apache.org>Send Time:2016年6月29日(星期三) 21:48To:封仲淹(纪君祥) <
> zhongyan.feng@alibaba-inc.com>; dev <de...@flink.apache.org>Subject:Re:
> [Discuss] Why different job's tasks can run in the single process.
> >> Hi,
> >> yes, you are definitely right that allowing to run multiple user code
> tasks
> >> in the same TaskManager JVM is not good for stability. This mode is
> still
> >> there from the very early days of Flink where Yarn was not yet
> available.
> >> In a production environment I would now recommend to always run one
> >> Flink-Yarn cluster per job to get good isolation between different jobs.
> >>
> >> Cheers,
> >> Aljoscha
> >>
> >> On Wed, 29 Jun 2016 at 09:18 Longda Feng <zhongyan.feng@alibaba-inc.com
> >
> >> wrote:
> >>
> >>> hi ,
> >>> Sorry for asking the quest here? Any answer will be apprecated.
> >>> Why different job's tasks can run in the single process. (There are
> some
> >>> different job's tasks in one TaskManager).It seems Flink-on-Yarn can
> let
> >>> different job run on different process. But for standalone mode, this
> >>> problem still exists.
> >>> Why design Flink like this?The advantage What I can thought is as
> >>> following:(1) All task can share bigger memory pool.(2) The
> communication
> >>> between the tasks in the same process will be fast.
> >>> But this design will impact to the stability. Flink provide
> >>> User-Define-Function interface, if one of the User-Define-Function
> crash,
> >>> It maybe crack the whole JVM, If the TaskManager crash, all other
> job's
> >>> task in this TaskManager will be impacted. Even if the JVM don't
> crash, but
> >>> maybe lead to some other unexpected problem, what's more this will
> make the
> >>> code too sophisticated。Normal framework like Spark/Storm/Samza won't
> run
> >>> different job's tasks in the same process。As one normal user,
> stability has
> >>> the highest priority.
> >>>
> >>> ThanksLongda
> >>>
> >>>
> >>>
> >>>
>
>
Re: [Discuss] Why different job's tasks can run in the single process.
Posted by Jark Wu <wu...@alibaba-inc.com>.
As standalone mode has the disadvantage that the TaskManager JVM can’t isolate different jobs. Does we have any plan to improve this ?
- Jark Wu
> 在 2016年6月30日，下午4:19，Kevin Jacobs <ke...@cern.ch> 写道：
>
> In my opinion the streaming process can be perfectly simulated on a single node. You can setup a message distribution system like Kafka on a single node, you can run Spark on a single node and the only thing you need to change when running it on a cluster is that you need to change the environment. So there is no need to setup a cluster when testing the streaming process.
>
> Regards,
> Kevin
>
> On 30-06-16 09:54, Longda Feng wrote:
>>
>> This means Standalone mode is just for prototype.
>> But I think we need a lightweight solution for streaming process, standalone is the best solution. Some times, we need setup a flink cluster on a small cluster. setup a yarn cluster isn't convenient.
>> (1) in small company, the number of machine is small (2) When a data center is small, but we still need do some computing in this data center(3) some machines are in the while-list, they have been authorited to access some special data or machine, but the number of these machine is small.(4) some machine has critical data, they can't be shared with others, but the number of these machine is small.(5) when a team start to learn flink, he will setup a small cluster firstly, maybe he wo't want to setup a huge system, perfer to a small system
>>
>>
>> regardsLongda
>>
>> ------------------------------------------------------------------From:Aljoscha Krettek <al...@apache.org>Send Time:2016年6月29日(星期三) 21:48To:封仲淹(纪君祥) <zh...@alibaba-inc.com>; dev <de...@flink.apache.org>Subject:Re: [Discuss] Why different job's tasks can run in the single process.
>> Hi,
>> yes, you are definitely right that allowing to run multiple user code tasks
>> in the same TaskManager JVM is not good for stability. This mode is still
>> there from the very early days of Flink where Yarn was not yet available.
>> In a production environment I would now recommend to always run one
>> Flink-Yarn cluster per job to get good isolation between different jobs.
>>
>> Cheers,
>> Aljoscha
>>
>> On Wed, 29 Jun 2016 at 09:18 Longda Feng <zh...@alibaba-inc.com>
>> wrote:
>>
>>> hi ,
>>> Sorry for asking the quest here? Any answer will be apprecated.
>>> Why different job's tasks can run in the single process. (There are some
>>> different job's tasks in one TaskManager).It seems Flink-on-Yarn can let
>>> different job run on different process. But for standalone mode, this
>>> problem still exists.
>>> Why design Flink like this?The advantage What I can thought is as
>>> following:(1) All task can share bigger memory pool.(2) The communication
>>> between the tasks in the same process will be fast.
>>> But this design will impact to the stability. Flink provide
>>> User-Define-Function interface, if one of the User-Define-Function crash,
>>> It maybe crack the whole JVM, If the TaskManager crash, all other job's
>>> task in this TaskManager will be impacted. Even if the JVM don't crash, but
>>> maybe lead to some other unexpected problem, what's more this will make the
>>> code too sophisticated。Normal framework like Spark/Storm/Samza won't run
>>> different job's tasks in the same process。As one normal user, stability has
>>> the highest priority.
>>>
>>> ThanksLongda
>>>
>>>
>>>
>>>
Re: [Discuss] Why different job's tasks can run in the single
process.
Posted by Kevin Jacobs <ke...@cern.ch>.
In my opinion the streaming process can be perfectly simulated on a
single node. You can setup a message distribution system like Kafka on a
single node, you can run Spark on a single node and the only thing you
need to change when running it on a cluster is that you need to change
the environment. So there is no need to setup a cluster when testing the
streaming process.
Regards,
Kevin
On 30-06-16 09:54, Longda Feng wrote:
>
> This means Standalone mode is just for prototype.
> But I think we need a lightweight solution for streaming process, standalone is the best solution. Some times, we need setup a flink cluster on a small cluster. setup a yarn cluster isn't convenient.
> (1) in small company, the number of machine is small (2) When a data center is small, but we still need do some computing in this data center(3) some machines are in the while-list, they have been authorited to access some special data or machine, but the number of these machine is small.(4) some machine has critical data, they can't be shared with others, but the number of these machine is small.(5) when a team start to learn flink, he will setup a small cluster firstly, maybe he wo't want to setup a huge system, perfer to a small system
>
>
> regardsLongda
>
> ------------------------------------------------------------------From:Aljoscha Krettek <al...@apache.org>Send Time:2016\u5e746\u670829\u65e5(\u661f\u671f\u4e09) 21:48To:\u5c01\u4ef2\u6df9(\u7eaa\u541b\u7965) <zh...@alibaba-inc.com>; dev <de...@flink.apache.org>Subject:Re: [Discuss] Why different job's tasks can run in the single process.
> Hi,
> yes, you are definitely right that allowing to run multiple user code tasks
> in the same TaskManager JVM is not good for stability. This mode is still
> there from the very early days of Flink where Yarn was not yet available.
> In a production environment I would now recommend to always run one
> Flink-Yarn cluster per job to get good isolation between different jobs.
>
> Cheers,
> Aljoscha
>
> On Wed, 29 Jun 2016 at 09:18 Longda Feng <zh...@alibaba-inc.com>
> wrote:
>
>> hi ,
>> Sorry for asking the quest here? Any answer will be apprecated.
>> Why different job's tasks can run in the single process. (There are some
>> different job's tasks in one TaskManager).It seems Flink-on-Yarn can let
>> different job run on different process. But for standalone mode, this
>> problem still exists.
>> Why design Flink like this?The advantage What I can thought is as
>> following:(1) All task can share bigger memory pool.(2) The communication
>> between the tasks in the same process will be fast.
>> But this design will impact to the stability. Flink provide
>> User-Define-Function interface, if one of the User-Define-Function crash,
>> It maybe crack the whole JVM, If the TaskManager crash, all other job's
>> task in this TaskManager will be impacted. Even if the JVM don't crash, but
>> maybe lead to some other unexpected problem, what's more this will make the
>> code too sophisticated\u3002Normal framework like Spark/Storm/Samza won't run
>> different job's tasks in the same process\u3002As one normal user, stability has
>> the highest priority.
>>
>> ThanksLongda
>>
>>
>>
>>
Re: [Discuss] Why different job's tasks can run in the single process.
Posted by Aljoscha Krettek <al...@apache.org>.
Hi,
no, it is not just for prototyping. Since it is actually the oldest mode of
execution it is also the most stable. You should have no problem using it.
Cheers,
Aljoscha
On Thu, 30 Jun 2016 at 09:54 Longda Feng <zh...@alibaba-inc.com>
wrote:
>
>
> This means Standalone mode is just for prototype.
> But I think we need a lightweight solution for streaming process,
> standalone is the best solution. Some times, we need setup a flink cluster
> on a small cluster. setup a yarn cluster isn't convenient.
> (1) in small company, the number of machine is small (2) When a data
> center is small, but we still need do some computing in this data center(3)
> some machines are in the while-list, they have been authorited to access
> some special data or machine, but the number of these machine is small.(4)
> some machine has critical data, they can't be shared with others, but the
> number of these machine is small.(5) when a team start to learn flink, he
> will setup a small cluster firstly, maybe he wo't want to setup a huge
> system, perfer to a small system
>
>
> regardsLongda
>
> ------------------------------------------------------------------From:Aljoscha
> Krettek <al...@apache.org>Send Time:2016年6月29日(星期三) 21:48To:封仲淹(纪君祥) <
> zhongyan.feng@alibaba-inc.com>; dev <de...@flink.apache.org>Subject:Re:
> [Discuss] Why different job's tasks can run in the single process.
> Hi,
> yes, you are definitely right that allowing to run multiple user code tasks
> in the same TaskManager JVM is not good for stability. This mode is still
> there from the very early days of Flink where Yarn was not yet available.
> In a production environment I would now recommend to always run one
> Flink-Yarn cluster per job to get good isolation between different jobs.
>
> Cheers,
> Aljoscha
>
> On Wed, 29 Jun 2016 at 09:18 Longda Feng <zh...@alibaba-inc.com>
> wrote:
>
> > hi ,
> > Sorry for asking the quest here? Any answer will be apprecated.
> > Why different job's tasks can run in the single process. (There are some
> > different job's tasks in one TaskManager).It seems Flink-on-Yarn can let
> > different job run on different process. But for standalone mode, this
> > problem still exists.
> > Why design Flink like this?The advantage What I can thought is as
> > following:(1) All task can share bigger memory pool.(2) The communication
> > between the tasks in the same process will be fast.
> > But this design will impact to the stability. Flink provide
> > User-Define-Function interface, if one of the User-Define-Function crash,
> > It maybe crack the whole JVM, If the TaskManager crash, all other job's
>
> > task in this TaskManager will be impacted. Even if the JVM don't crash, but
>
> > maybe lead to some other unexpected problem, what's more this will make the
> > code too sophisticated。Normal framework like Spark/Storm/Samza won't run
>
> > different job's tasks in the same process。As one normal user, stability has
> > the highest priority.
> >
> > ThanksLongda
> >
> >
> >
> >
>
>
Re: [Discuss] Why different job's tasks can run in the single process.
Posted by Longda Feng <zh...@alibaba-inc.com>.
This means Standalone mode is just for prototype.
But I think we need a lightweight solution for streaming process, standalone is the best solution. Some times, we need setup a flink cluster on a small cluster. setup a yarn cluster isn't convenient.
(1) in small company, the number of machine is small (2) When a data center is small, but we still need do some computing in this data center(3) some machines are in the while-list, they have been authorited to access some special data or machine, but the number of these machine is small.(4) some machine has critical data, they can't be shared with others, but the number of these machine is small.(5) when a team start to learn flink, he will setup a small cluster firstly, maybe he wo't want to setup a huge system, perfer to a small system
regardsLongda
------------------------------------------------------------------From:Aljoscha Krettek <al...@apache.org>Send Time:2016年6月29日(星期三) 21:48To:封仲淹(纪君祥) <zh...@alibaba-inc.com>; dev <de...@flink.apache.org>Subject:Re: [Discuss] Why different job's tasks can run in the single process.
Hi,
yes, you are definitely right that allowing to run multiple user code tasks
in the same TaskManager JVM is not good for stability. This mode is still
there from the very early days of Flink where Yarn was not yet available.
In a production environment I would now recommend to always run one
Flink-Yarn cluster per job to get good isolation between different jobs.
Cheers,
Aljoscha
On Wed, 29 Jun 2016 at 09:18 Longda Feng <zh...@alibaba-inc.com>
wrote:
> hi ,
> Sorry for asking the quest here? Any answer will be apprecated.
> Why different job's tasks can run in the single process. (There are some
> different job's tasks in one TaskManager).It seems Flink-on-Yarn can let
> different job run on different process. But for standalone mode, this
> problem still exists.
> Why design Flink like this?The advantage What I can thought is as
> following:(1) All task can share bigger memory pool.(2) The communication
> between the tasks in the same process will be fast.
> But this design will impact to the stability. Flink provide
> User-Define-Function interface, if one of the User-Define-Function crash,
> It maybe crack the whole JVM, If the TaskManager crash, all other job's
> task in this TaskManager will be impacted. Even if the JVM don't crash, but
> maybe lead to some other unexpected problem, what's more this will make the
> code too sophisticated。Normal framework like Spark/Storm/Samza won't run
> different job's tasks in the same process。As one normal user, stability has
> the highest priority.
>
> ThanksLongda
>
>
>
>
Re: [Discuss] Why different job's tasks can run in the single process.
Posted by Aljoscha Krettek <al...@apache.org>.
Hi,
yes, you are definitely right that allowing to run multiple user code tasks
in the same TaskManager JVM is not good for stability. This mode is still
there from the very early days of Flink where Yarn was not yet available.
In a production environment I would now recommend to always run one
Flink-Yarn cluster per job to get good isolation between different jobs.
Cheers,
Aljoscha
On Wed, 29 Jun 2016 at 09:18 Longda Feng <zh...@alibaba-inc.com>
wrote:
> hi ,
> Sorry for asking the quest here? Any answer will be apprecated.
> Why different job's tasks can run in the single process. (There are some
> different job's tasks in one TaskManager).It seems Flink-on-Yarn can let
> different job run on different process. But for standalone mode, this
> problem still exists.
> Why design Flink like this?The advantage What I can thought is as
> following:(1) All task can share bigger memory pool.(2) The communication
> between the tasks in the same process will be fast.
> But this design will impact to the stability. Flink provide
> User-Define-Function interface, if one of the User-Define-Function crash,
> It maybe crack the whole JVM, If the TaskManager crash, all other job's
> task in this TaskManager will be impacted. Even if the JVM don't crash, but
> maybe lead to some other unexpected problem, what's more this will make the
> code too sophisticated。Normal framework like Spark/Storm/Samza won't run
> different job's tasks in the same process。As one normal user, stability has
> the highest priority.
>
> ThanksLongda
>
>
>
>