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

Re: [DISCUSS] Proposal of external shuffle service

Hi Andrey,

Thanks for further research on this topic and providing very helpful summaries.  

As we discussed before, I really like the idea of dividing two separate components on both JM and TM sides.

1. On JM side, the ShuffleMaster componenet created from ShuffleManager can manage and handle partition related issues properly.

1.1 The introduction of PartitionShuffleDescriptor and PartitiondDeploymentDescriptor is suitable for covering all the necessary infos related with partition during deployment process and other future extensions. The form of this new descriptor is also consistent with existing ResultPartitionDeploymentDescriptor and InputGateDeploymentDescriptor.

2. On TM side, the ShuffleService component created from ShuffleManager is a TM level service, which can be used for creating ResultPartitionWriter and InputGate during task deployment.

2.1 Concerning of updating UnknownShuffleDeploymentDescriptor,  I think it may bring an argument that whether the ShuffleService should provide a separate method for updating it or not. In other words, because the InputGate is created by ShuffleService, then whether all the possible operations for InputGate such as update or release should be handled via ShuffleService? I think it can be interpreted to operate InputGate directly if the update or release is general for all the ShuffleService implementations. But the InputGate interface should provide the explicit methods for releasing itself and updating input channels to make the whole process work.

2.2 In addition, some implementation details can be further confirmed in separate JIRAs,  such as whether we need task info related parameters during creating writer, and how to extract necessary components from current NetworkEnvrironment to wrap in specific ShuffleService implementation, etc.

3. For the points mentioned in future extensions, I agree with your analysis. We can focus on them separately step by step in different priorities. The above ShuffleMaster provides a basic precondition for decoupling the life cycles between partition state and task state. Then we can further extend the methods in ShuffleMaster to know whether the partition is still available for speeding up failover, and whether the partition is consumed by downstream to decide when to release TM or clean partition, etc. It is also a good idea to further refactor the interfaces on writer and reader sides to fine-grained handle raw record instead of Buffer. And it would be involved in more changes in current RecordWriter/StreamInputProcessor.

I think we can further confirm the above 2.1 issue, then I would adjust the google doc based on our conclusions which cover not only the first step, but also all the future extensions described and listed in priority. 
BTW, do you think it is necessary that we further co-author a FLIP for this feature? It is actually involved in many changes on both TM, JM sides.  :)


From:Andrey Zagrebin <andrey@xxxxxxxxxxxxxxxxx>
Send Time:2018年12月20日(星期四) 01:20
To:zhijiang <wangzhijiang999@xxxxxxxxxx>
Cc:dev <dev@xxxxxxxxxxxxxxxx>
Subject:Re: [DISCUSS] Proposal of external shuffle service

Hi Zhijiang,

Thanks for detailed answers! I am glad we are on the same page.

I spent some time on thinking more about our concerns and decided to make more suggestions for the discussion.
At the end, I also gathered some points related to possible extensions of shuffle API to verify that the final pluggable design can support them later with less changes.

It might make sense for shuffle implementation to have component running on both JM and TM sides.
JM has a global view of what is happening and can interact with shuffling system independently on whether tasks are running or not. The component services could internally further communicate to each other outside of existing JM/TM APIs, depending on shuffle implementation.
It could help later with partition global life cycle management and cleanup.
Moreover, if we decide to use some ShuffleDeploymentDescriptor instead of ResultPartitionLocation or factories to instantiate Readers and Writers, they can be created in Task Executor. 
JM is probably not interested in this concern. ShuffleDeploymentDescriptor can be specific to shuffle implementation, like factories, and contain specific shuffle config for task side.
1. Configuration:
interface ShuffleManager {
  ShuffleMaster createMaster(Configuration flinkConfig);
  ShuffleService createService(Configuration flinkConfig);
ShuffleManager is a factory for ShuffleMaster (JM side) and ShuffleService (TM side).
Flink config could also contain specific shuffle configuration, like port etc.
Class which implements ShuffleManager in Flink cluster config, default is what we have now (can be the first step)
2. Job master side
class PartitionShuffleDescriptor {
  JobID, ExecutionAttemptID, ResultPartitionType, ResultPartitionLocation, TaskManagerLocation, etc
  later possibly ShuffleType/Descriptor to choose from available shuffle implementations
PartitionShuffleDescriptor contains all abstract information which JM can provide from the job/execution graph.
ResultPartitionType and ResultPartitionLocation are derived from graph and execution mode, 
so I think they are rather general parameters for any shuffle service and do not belong to particular shuffle implementation.
interface ShuffleMaster extends AutoClosable {
  ShuffleDeploymentDescriptor registerPartition(PartitionShuffleDescriptor);
  void deregisterPartition(PartionShuffleDescriptor);
JM process creates ShuffleMaster from configured per cluster ShuffleManager. JM is responsible for its life cycle.
ShuffleMaster is a global manager for partitions.
JM creates PartitionShuffleDescriptor and uses ShuffleMaster to register partition, e.g. when producer is deployed. 
ShuffleMaster transforms abstract PartitionShuffleDescriptor into a specific ShuffleDeploymentDescriptor.
ShuffleDeploymentDescriptor is put into ResultPartitionDeploymentDescriptor and InputGateDeploymentDescriptor.
It can contain specific partition config for ShuffleService on TM side to serve record readers and writers. 
If it is channel-based then further break down to channel configs.
Special UnknownShuffleDeploymentDescriptor could be used for eager deployment when task input is unknown yet.
Later, we could add an option to release partition globally by deregistering it with the ShuffleMaster. e.g. to clean it up.
3. Task executor side
interface ShuffleService extends AutoClosable {
  ResultPartitionWriter createResultPartitionWriter(ResultPartitionDeploymentDescriptor);
  InputGate createInputGate(InputGateDeploymentDescriptor);
TM process creates ShuffleService from configured per cluster ShuffleManager. TM is responsible for its life cycle.
ShuffleService could substitute NetworkEnvironment in TaskManagerServices.
4. Later extensions
4.1 Per job/job edge config
To keep jobs cluster independent, we could introduce abstract predefined ShuffleType’s or descriptors
for job developer to set it per job or job edge. The types are cluster-independent.
Cluster config could contain provided ShuffleManager implementation class for each supported ShuffleType or fallback to default for some types.
Instead of one ShuffleMaster/ShuffleService, JM/TM could have keep a registry of ShuffleMaster/ShuffleService’s per ShuffleType.
4.2 Delay TM shutdown until all local partitions have been consumed
JM could keep separately state of partition life cycle (e.g. in job state, HA). The task executor is to shutdown (e.g. after timeout in yarn) if all its tasks are done and all local partitions are consumed. If there are no local partitions then it can shutdown immediately. Whether JM should check that all produced by TM partitions are consumed is a feature of ShuffleManager. This could be done by calling some ShuffleManager.getFeatures() interface method.
4.3 Speed up failover
If partition is computed JM could reuse it as mention in fine-grained shuffle system design. Whether the partition is still available after task or task executor crash is also a feature of ShuffleManager.getFeatures().
4.4 Partition garbage collection
When the consumer task is done, the partition is to deregister and cleanup with the ShuffleMaster. 
In case of external storage, partitions are at risk to linger after job/cluster failures. The partition TTL is one option as mentioned in fine-grained shuffle system design. The TTL timer could be started when there is no partition access activity for certain period of time but there is always risk to lose partition too early. User could try to recover failed job any time later. So it might need more sophisticated approach, like manual cleanup triggering (ShuffleMaster.cleanup(PartitionsInUse)) which drops all currently unused partitions.
4.5 Shuffle Reader/Writers operation per record/byte[]/buffer
As discussed, ResultPartitionWriter/InputGate operates on buffers with serialised records data. Certain shuffle services might benefit from operating per serialised records or even java objects (e.g. local channel could hand over them or their copies from TypeSerializer.copy()). Record key could be treated as its meta info, additionally to bytes or to user java object.
ShuffleService could be refactored later to return RecordReader/RecordWriter. They could extend AbstractSerialisingRecordReader/Writer or AbstractBufferingRecordReader/Writer to import current behaviour and share code. This requires refactoring of StreamInputProcessor and RecordWriter to extract the interfaces.
It might be useful for ResultPartitionWriter/InputGate or RecordReader/RecordWriter also to extend AutoClosable in case the internal implementation needs a per task life cycle for them.

I hope it can help with the design. Feel free to give feedback.


On 10 Dec 2018, at 08:41, zhijiang <wangzhijiang999@xxxxxxxxxx> wrote:
Hi Andrey,

Thanks for providing so detailed concerns and enlightenments for this proposal. We exchanged our views of three main issues on google doc last week and it seems more appropriate to further contact here. :)

1. Configuration level for shuffle (cluster/job/operator)
- how do we share shuffle manager resources among different job tasks within one task executor process? It could be some static objects shared by all shuffle manager objects of some type but it might be not scalable approach. Example could be multiplexed netty connections (as I understand, current netty stack can become just custom shuffle service).
The creation of ShuffleManager instance on task level is just like the process of creating StateBackend in StateBackendLoader. The ShuffleService and ShuffleManager are two independent components, and the interaction between them is only registration mechanism. In detail, if some ShuffleManager instances want to rely ShuffleService to transport data, then it can register related infos to ShuffleService during creation of ResultPartitionWriter. So the ShuffleManager instance do not need  contain any objects like netty related stacks. The flink runtime can provide one unified netty-based ShuffleService which can be started in both internal TaskManager or external containers. The internal ShuffleService not only takes the role of tranporting data directly for some ShuffleManager instances but also takes the role of RPC server for communicating with external ShuffleService, such as register result partition to external service, otherwise the external service might need an additional RPC service to contact with TaskManager.  Here we have the implicit meaning to make intenral shuffle as a basic service started in TaskManager like the components of IOManager and MemoryManager, even thought useless for some type jobs.
- In case of having it per job, we might need to provide compatibility check between shuffle service and cluster mode (e.g. yarn ext shuffle service for standalone mode cluster) if it is an issue.
- Having it per job feels like the same complexity as having it per operator, at the first glance, just changes its granularity and where objects reside.
- what is the problem to use cluster per job mode? Then shuffle manager per cluster and per job is the same but might simplify other issues at the beginning. Streaming and batch jobs with different shuffle requirements could be started in different clusters per job.

I totally agree with the above concerns for per job configuration. As you mentioned, it is a option to run different type jobs in different clusters. But in some special scenarios like hybrid cluster to run online and offline jobs in differemt times, it is betterto support job level configuration for fexibility. Certainly it may not be a strong requirements for most cases, then we can reach an agreement to make the cluster level as the easiest way first and adjut the level if needed in future.

2. ShuffleManager interface

I think you mentioned three sub issues in this part:

2.1 Introduction of additional ResultPartitionWriterFactory && InputGateReaderFactory

I am not against the introduction of these two factories. The original introduction of pluggable ShuffleManager interface is for creating different writer and reader sides. If the ShuffleManager interface is used for creating factories, and then the factories are used for creating writer and reader. I still think the essence is same, and only the form is different.  That is the ShuffleManager concept is seen on JobManager side, and the task only sees the corresponding factories from ShuffleManager. In other words, we add another factory layer to distinguish between JobManager and task. The form might seem a bit better to introduce corresponding factories, so I am willing to take this way for implementation.

2.2 Whether to retain getResultPartitionLocation method in ShuffleManager interface

If I understand correctly, you mean to put this location as an argument in InputGateReaderFacotry constructor? If to do so, I think it makes sense and we can avoid have this explicit method in interface. But we also need to adjust the existing related process like updatePartitionInfo for downstream side. In this case, the partition location is unknown during deploying downstream tasks. Based on upstream's consumable notification, the location update is triggered by JobManager to downstream side.

2.3 ShuffleService interface

My initial thought is not making it as an interface. Because for internal or external shuffle cases, they can reuse the same unified netty-based shuffle service if we wrap the related componenets into current shuffle service well. If we want to furtherextend other implementations of shuffle service, like http-based shuffle service, then we can define an interface for it, the way as current RpcService interface to get ride of only akka implementations. So it also makes sense on my side to keep this interface. As for ShuffleServiceRegistry class, I agree with you to have this TaskManager level service for managing and sharing for all the internal tasks.

In summary, I think we do not have essential conflicts for above issues, almost for the implementation aspects. And I agree with the above points, especially for above 2.2 you might need double check if I understand correctly. 
Wish your further feedbacks then I can adjust the docs based on it.  Also welcome any other person's feedbacks!


发件人:Andrey Zagrebin <andrey@xxxxxxxxxxxxxxxxx>
发送时间:2018年12月10日(星期一) 05:18
收件人:dev <dev@xxxxxxxxxxxxxxxx>; zhijiang <wangzhijiang999@xxxxxxxxxx>
抄 送:Nico Kruber <nico@xxxxxxxxxxxxxxxxx>; Piotr Nowojski <piotr@xxxxxxxxxxxxxxxxx>; Stephan Ewen <sewen@xxxxxxxxxx>; Till Rohrmann <trohrmann@xxxxxxxxxx>
主 题:Re: [DISCUSS] Proposal of external shuffle service

Hi Zhijiang,

Thanks for sharing the document Zhijiang. 
I decided to compile my thoughts to consider here, not to overload document comments any more :)
I think I still have question about job level configuration for the shuffle service. You mentioned that we can keep several shuffle manager objects in one task executor for different jobs. This is valid. My concerns are:
- how do we share shuffle manager resources among different job tasks within one task executor process? It could be some static objects shared by all shuffle manager objects of some type but it might be not scalable approach. Example could be multiplexed netty connections (as I understand, current netty stack can become just custom shuffle service).
- In case of having it per job, we might need to provide compatibility check between shuffle service and cluster mode (e.g. yarn ext shuffle service for standalone mode cluster) if it is an issue.
- Having it per job feels like the same complexity as having it per operator, at the first glance, just changes its granularity and where objects reside.
- what is the problem to use cluster per job mode? Then shuffle manager per cluster and per job is the same but might simplify other issues at the beginning. Streaming and batch jobs with different shuffle requirements could be started in different clusters per job. 
As for ShuffleManager interface, I think I see your point with the ResultPartitionLocation. I agree that partition needs some addressing of underlying connection or resources in general. It can be thinked of as an argument of ShuffleManager factory methods.
My point is that task code might not need to be coupled to shuffle interface. This way we could keep task code more independent of records transfer layer. We can always change later how shuffle/network service is organised internally without any consequences for the general task code. If task code calls just factories provided by JM, it might not even matter for the task in future whether it is configured per cluster, job or operator. Internally, factory can hold location of concrete type if needed.
Code example could be:
Job Manager side:
interface ShuffleManager {
  ResultPartionWriterFactory createResultPartionWriterFactory(job/task/topology descriptors);
  // similar for input gate factory
class ShuffleManagerImpl implements ShuffleManager {
  private general config, services etc;
  ResultPartionWriterFactory createResultPartionWriterFactory(job/task/topology descriptors) {
    return new ResultPartionWriterFactoryImpl(location, job, oper id, other specific config etc);
  // similar for input gate factory
// somewhere in higher level code put ResultPartionWriterFactory into descriptor
Task executor side receives the factory inside the descriptor and calls factory.create(ShuffleServiceRegistry). Example of factory:
class ResultPartionWriterFactoryImpl implements ResultPartionWriterFactory {
  // all fields are lightweight and serialisable, received from JM
  private location, shuffle service id, other specific config etc;

 ResultPartionWriter create(ShuffleServiceRegistry registry, maybe more generic args) {
    // get or create task local specific ShuffleServiceImpl by id in registry
    // ShuffleServiceImpl object can be shared between jobs
    // register with the ShuffleServiceImpl by location, id, config etc
interface ShuffleService extends AutoClosable {
ShuffleServiceImpl manages resources and decides internally whether to do it per task executor, task, job or operator. It can contain network stack, e,g, netty connections etc. In case of external service, it can hold partition manager, transport client etc. It is not enforced to have it per job by this contract or even to have it at all. ShuffleServiceImpl also does not need to depend on all TaskManagerServices, only create relevant inside, e.g. network.
class ShuffleServiceRegistry {
  <T extends ShuffleService> T getShuffleService(id);
 registerShuffleService(ShuffleService, id);
  deregisterShuffleService(id); // remove and close ShuffleService
  close(); // close all
ShuffleServiceRegistry is just a generic container of all available ShuffleService’s. It could be part of TaskManagerServices instead of NetworkEnvironment which could go into specific ShuffleServiceImpl.

I might still miss some details, I would appreciate any feedback.


On 28 Nov 2018, at 08:59, zhijiang <wangzhijiang999@xxxxxxxxxx.INVALID> wrote:
Hi all,

I adjusted the umbrella jira [1] and corresponding google doc [2] to narrow down the scope of introducing pluggable shuffle manager architecture as the first step. 
Welcome further feedbacks and suggestions, then I would create specific subtasks for it to forward.


发件人:zhijiang <wangzhijiang999@xxxxxxxxxx.INVALID>
发送时间:2018年11月1日(星期四) 17:19
收件人:dev <dev@xxxxxxxxxxxxxxxx>; Jin Sun <isunjin@xxxxxxxxx>
抄 送:Nico Kruber <nico@xxxxxxxxxxxxxxxxx>; Piotr Nowojski <piotr@xxxxxxxxxxxxxxxxx>; Stephan Ewen <sewen@xxxxxxxxxx>
主 题:回复:[DISCUSS] Proposal of external shuffle service

Thanks for the efficient response till!

Thanks sunjin for the good feedbacks, we will further confirm with the comments then! :)
发件人:Jin Sun <isunjin@xxxxxxxxx>
发送时间:2018年11月1日(星期四) 06:42
收件人:dev <dev@xxxxxxxxxxxxxxxx>
抄 送:Zhijiang(wangzhijiang999) <wangzhijiang999@xxxxxxxxxx>; Nico Kruber <nico@xxxxxxxxxxxxxxxxx>; Piotr Nowojski <piotr@xxxxxxxxxxxxxxxxx>; Stephan Ewen <sewen@xxxxxxxxxx>
主 题:Re: [DISCUSS] Proposal of external shuffle service

Thanks Zhijiang for the proposal. I like the idea of external shuffle service, have left some comments on the document. 

On Oct 31, 2018, at 2:26 AM, Till Rohrmann <trohrmann@xxxxxxxxxx> wrote:

Thanks for the update Zhijiang! The community is currently quite busy with
the next Flink release. I hope that we can finish the release in two weeks.
After that people will become more responsive again.


On Wed, Oct 31, 2018 at 7:49 AM zhijiang <wangzhijiang999@xxxxxxxxxx> wrote:

I already created the umbrella jira [1] for this improvement, and attched
the design doc [2] in this jira.

Welcome for further discussion about the details.



发件人:Zhijiang(wangzhijiang999) <wangzhijiang999@xxxxxxxxxx.INVALID>
发送时间:2018年9月11日(星期二) 15:21
收件人:dev <dev@xxxxxxxxxxxxxxxx>
抄 送:dev <dev@xxxxxxxxxxxxxxxx>
主 题:回复:[DISCUSS] Proposal of external shuffle service

Many thanks Till!

I would create a JIRA for this feature and design a document attched with it.
I will let you know after ready! :)


发件人:Till Rohrmann <trohrmann@xxxxxxxxxx>
发送时间:2018年9月7日(星期五) 22:01
收件人:Zhijiang(wangzhijiang999) <wangzhijiang999@xxxxxxxxxx>
抄 送:dev <dev@xxxxxxxxxxxxxxxx>
主 题:Re: [DISCUSS] Proposal of external shuffle service

The rough plan sounds good Zhijiang. I think we should continue with what
you've proposed: Open a JIRA issue and creating a design document which
outlines the required changes a little bit more in detail. Once this is
done, we should link the design document in the JIRA issue and post it here
for further discussion.


On Wed, Aug 29, 2018 at 6:04 PM Zhijiang(wangzhijiang999) <
wangzhijiang999@xxxxxxxxxx> wrote:

Glad to receive your positive feedbacks Till!

Actually our motivation is to support batch job well as you mentioned.

For output level, flink already has the Subpartition abstraction(writer),
and currently there are PipelinedSubpartition(memory output) and
SpillableSubpartition(one-sp-one-file output) implementations. We can
extend this abstraction to realize other persistent outputs (e.g.

For transport level(shuffle service), the current SubpartitionView
abstraction(reader) seems as the brige linked with the output level, then

the view can understand and read the different output formats. The current
NetworkEnvironment seems take the role of internal shuffle service in
TaskManager and the transport server is realized by netty inside. This

component can also be started in other external containers like NodeManager
of yarn to take the role of external shuffle service. Further we can

abstract to extend the shuffle service for transporting outputs by http or

rdma instead of current netty.  This abstraction should provide the way for
output registration in order to read the results correctly, similar with
current SubpartitionView.

The above is still a rough idea. Next I plan to create a feature jira to
cover the related changes if possible. It would be better if getting help
from related committers to review the detail designs together.


发件人:Till Rohrmann <trohrmann@xxxxxxxxxx>
发送时间:2018年8月29日(星期三) 17:36
收件人:dev <dev@xxxxxxxxxxxxxxxx>; Zhijiang(wangzhijiang999) <
主 题: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!