So should we enforce having 2 instances now or defer this to a later date?

I'd rather do this early since it changes 2 assumptions that ShuffleMaster can currently make:
- every partition release is preceded by a registration of said partition
- the release of partitions may rely on local data

On 04/10/2019 17:10, Till Rohrmann wrote:
Thanks for updating the FLIP.

I think the RM does not need to have access to a full fledged ShuffleMaster
implementation. Instead it should enough to give it a leaner interface
which only supports to delete result partitions and list available global
partitions. This might entail that one will have a ShuffleMaster
implementation running on the Dispatcher and a
GlobalResultPartitionsShuffleMaster implementation running on the RM. Long
story short, if we separate the RM from the Dispatcher, then this might
entail that we will have two ShuffleMaster incarnations running in each
process.

Cheers,
Till

On Fri, Oct 4, 2019 at 3:34 PM Chesnay Schepler <ches...@apache.org> wrote:

I have updated the FLIP.

- consistently use "local"/"global" terminology; this incidentally should
make it easier to update the terminology if we decide on other names
- inform RM via heartbeats from TE about available global partitions
- add dedicated method for releasing global partitions
- add dedicated section for required changes to the ShuffleMaster (mostly
clarification)
- added some items to the "Rejected Alternatives" section
- updated discussion link


While writing the ShuffleMaster section I noticed the following:

If, at any point, the JM/RM are moved into dedicated processes we either
a) have multiple ShuffleMaster instances for the same shuffle service
active
b) require a single ShuffleMaster on the RM, to which JM calls are being
forwarded.

Neither of these are without pain-points;
a) introduces additional constraints on ShuffleMaster implementations in
that no local state must be kept
b) again forces the JM to regularly be in touch with the RM, and limits
the ShuffleMaster interface to being RPC-friendly.

I'm wondering whether this issue was already an anyone's radar.


On 04/10/2019 14:12, Till Rohrmann wrote:



On Fri, Oct 4, 2019 at 12:37 PM Chesnay Schepler <ches...@apache.org>
wrote:

*Till: In the FLIP you wrote "The set of partitions to release may contain local
and/or global partitions; the promotion set must only refer to local
partitions." to describe the `releasePartitions`. I think the JM should
never be in the situation to release a global partition. Moreover, I
believe we should have a separate RPC to release global result partitions
which might come from the RM.*

We can certainly add a separate RPC method for explicitly releasing global 
partitions.
You are correct that the JM should not be able to release those, just like the 
RM should not be able to release non-global partitions.
*Till: Once the JM has obtained the required slots to run a job, it no longer
needs to communicate with the RM. Hence, a lost RM connection won't
interfere with the job. I would like to keep it like this by letting the TE
announce global result partitions to the RM and not to introduce another
communication roundtrip.

*Agreed, this is a nice property to retain.
*Till: How big do you expect the payload to become?

*I don't know, which is precisely why I want to be cautious about it.
The last time I made a similar assumption I didn't expect anyone to have 
hundreds of thousands of metrics on a single TM, which turned out to be wrong.
I wouldn't exclude the possibility of a similar number of partitions being 
hosted on a single TE.


One problem we have to solve with the heartbeat-based approach is that 
partitions may be lost without the TE noticing, due to disk-failures or 
external delete operations.
Currently, for scheduling purposes we rely on information stored in the JM, and 
update said information if a job fails due to a missing partition. However, 
IIRC the JM is informed about with an exception that is thrown by the consumer 
of said partition, not the producer. As far as the producing TM is concerned, 
it is still hosting that partition.
This means we have to forward errors for missing partitions from the network 
stack on the producers side to the TE, so that it can inform the RM about it.


Yes, I think you are right Chesnay. This would also be a good addition for
the local result partitions.

Cheers,
Till

On 02/10/2019 16:21, Till Rohrmann wrote:

Thanks for addressing our comments Chesnay. See some comments inline.

On Wed, Oct 2, 2019 at 4:07 PM Chesnay Schepler <ches...@apache.org> 
<ches...@apache.org> wrote:


Thank you for your comments; I've aggregated them a bit and added
comments to each of them.

1) Concept name (proposal: persistent)

I agree that "global" is rather undescriptive, particularly so since we
never had a notion of "local" partitions.
I'm not a fan of "persistent"; as to me this always implies reliable
long-term storage which as I understand we aren't shooting for here.

I was thinking of "cached" partitions.

To Zhijiangs point, we should of course make the naming consistent
everywhere.

2) Naming of last parameter of TE#releasePartitions (proposal:
partitionsToRetain / partitionsToPersistent)

I can see where you're coming from ("promote" is somewhat abstract), but
I think both suggestions have downsides.

"partitionsToPersistent" to me implies an additional write operation to
somewhere, but we aren't doing that.
"partitionsToRetain" kind of results in a redundancy with the other
argument since retaining is the opposite to releasing a partition; if I
want to retain a partition, why am I not just excluding it from the set
to release?

I quite like "promote" personally; we fundamentally change how the
lifecycle for these partitions work, and introducing new keywords isn't
a inherently a bad choice.

3) Naming of TE#releasePartitions (proposal: releaseOrPromotePartitions;
Note: addition of "OrPromote" is dependent on 2) )

Good point.

4) /Till: I'm not sure whether partitionsToRelease should contain a//
//global/persistent result partition id. I always thought that the user
will//
//be responsible for managing the lifecycle of a global/persistent//
//result partition./

@Till Please elaborate; which method/argument are you referring to?


In the FLIP you wrote "The set of partitions to release may contain local
and/or global partitions; the promotion set must only refer to local
partitions." to describe the `releasePartitions`. I think the JM should
never be in the situation to release a global partition. Moreover, I
believe we should have a separate RPC to release global result partitions
which might come from the RM.


4)/Dedicated PartitionTable for global partitions/

Since there is only one RM for each TE a PartitionTable is unnecessary;
a simple set will suffice.
Alternatively, we could introduce such a dedicated set into the
PartitionTable to keep these data-structures close.

5) /Zhijiang: Nit: TM->TE in the section of Proposed Changes: "TMs
retain global partitions for successful jobs"/

Will fix it.

6) /Zhijiang: Considering ShuffleMaster, it was built inside JM and
expected to interactive with JM before. Now the RM also needs to
interactive with ShuffleMaster to release global partitions. Then it
might be better to move ShuffleMaster outside of JM, and the lifecycle
of ShuffleMaster should be consistent with RM./

Yes, I alluded to this in the FLIP but should've been more explicit; the
shuffle master must outlive the JM. This is somewhat tricky when
considering the future a bit; if we assume that different jobs or even a
single one can use different shuffle services, then we need a way to
associate the partitions with the corresponding shuffle master. This
will likely require the introduction of a ShuffleMasterID that is
included in the ShuffleDescriptor.

7) Handover

/Till: The handover logic between the JM and the RM for the
global/persistent//
//result partitions seems a bit brittle to me. What will happen if the JM//
//cannot reach the RM? I think it would be better if the TM announces the//
//global/persistent result partitions to the RM via its heartbeats. That
way//
//we don't rely on an established connection between the JM and RM and we//
//keep the TM as the ground of truth. Moreover, the RM should simply
forward//
//the release calls to the TM without much internal logic./

As for your question, if the JM cannot reach the RM the handover will
fail, the JM will likely shutdown without promoting any partition and
the TE will release all partitions.
What is the defined behavior for the JM in case of the RM disconnect
after a job has finished? Does it always/sometimes/never shutdown
with/-out communicating the result to the client / updating HA data;
or simply put, does the JM behave to the user as if nothing has happened
in all cases?


Once the JM has obtained the required slots to run a job, it no longer
needs to communicate with the RM. Hence, a lost RM connection won't
interfere with the job. I would like to keep it like this by letting the TE
announce global result partitions to the RM and not to introduce another
communication roundtrip.


A heartbeat-based approach is useful and can alleviate some failure
cases (see below); but we need to make sure we don't exceed the akka
framesize or otherwise interfere with the heartbeat mechanism (like we
did with metrics in the past). Ideally we would only submit updates to
the partition set (added/removed partitions), but I'm not sure if the
heartbeats are reliable enough for this to work.


How big do you expect the payload to become?


8. Failure cases:
/Becket:/
/a) The TEs may remove the result partition while the RM does not//
//know. In this case, the client will receive a runtime error and submit
the//
//full DAG to recompute the missing result partition. In this case, RM
should//
//release the incomplete global partition. How would RM be notified to do//
//that?//
//b) Is it possible the RM looses global partition metadata while//
//the TE still host the data? For example, RM deletes the global
partition//
//entry while the release partition call to TE failed.//
//c) What would happen if the JM fails before the global partitions//
//are registered to RM? Are users exposed to resource leak if JM does not//
//have HA?//
//d) What would happen if the RM fails? Will TE release the//
//partitions by themselves?/

1.a) This is a good question that I haven't considered. This will likely
require a heartbeat-like report of available partitions.


The hearbeat based synchronization approach seems to crystalize as the way
to go forward with this FLIP.



1.b) RM should only delete entries if it received an ack from the TE;
otherwise we could easily end up leaking partitions. I believe I forgot
writing this down.
1.c) As described in the FLIP the handoff to the RM must occur before
partitions are promoted.
      If the JM fails during the handoff then the TE will cleanup all
partitions since it lost the connection to the JM, and partitions
weren't promoted yet.
      If the JM fails after the handoff but before the promotion, same as
above. The RM would contain invalid entries in this case; see 1.a) .
      If the JM fails after the handoff and promotion partitions we don't
leak anything since the RM is now fully responsible.
1.d) yes; if the connection to the RM is disrupted the TE will cleanup
all global partitions, similar to how it cleans up all partitions
associated with a given job if the connection to the corresponding JM is
disrupted.

9. /Becket: It looks that TE should be the source of truth of the result
partition//
//existence. Does it have to distinguish between global and local result//
//partitions? If TE does not need to distinguish them, it seems the the//
//releasePartition() method in TE could just provide the list of
partitions//
//to release, without the partitions to promote./

The promotion is a hard requirement, as this is the signal to the TE
that this partition is no longer bound to the life-cycle of a job.
Without the promotion the TE would delete the partition once the JM has
shutdown; this is a safety net to ensure cleanup of partitions in case
of a disconnect.

10. /In the current design, RM should be able to release result//
//partitions using ShuffleService. Will RM do this by sending RPC to the
TEs?//
//Or will the RM do it by itself?/

The RM will send a release call to each TM and issue a release call to
the ShuffleMaster, just like the JobMaster handles partition releases.

11. /Becket: How do we plan to handle the case when there are different
shuffle//
//services in the same Flink cluster? For example, a shared standalone//
//cluster./

This case is not considered; there are so many changes necessary in
other parts of the runtime that we would jump the gun in addressing it
here.
Ultimately though, I would think that that the addition of a shuffle
master instance ID and shuffle service identifier should suffice.

The identifier is used in subsequent jobs to load the appropriate
shuffle service for a given partition (think of it like a class name),
while the shuffle master instance ID is used to differentiate between
the different shuffle master instances running in the cluster (which
partitions have to be associated with so we can issue the correct
release calls).

12. /Becket: Minor: usually REST API uses `?` to pass the parameters. Is
there a//
//reason we use `:` instead?/

That's netty syntax for path parameters.

On 30/09/2019 08:34, Becket Qin wrote:

Forgot to say that I agree with Till that it seems a good idea to let TEs
register the global partitions to the RM instead of letting JM do it.

This

simplifies quite a few things.

Thanks,

Jiangjie (Becket) Qin

On Sun, Sep 29, 2019 at 11:25 PM Becket Qin <becket....@gmail.com> 
<becket....@gmail.com>

wrote:

Hi Chesnay,

Thanks for the proposal. My understanding of the entire workflow step by
step is following:

     - JM maintains the local and global partition metadata when the task
runs to create result partitions. The tasks themselves does not

distinguish

between local / global partitions. Only the JM knows that.
     - JM releases the local partitions as the job executes. When a job
finishes successfully, JM registers the global partitions to the RM. The
global partition IDs are set on the client instead of randomly

generated,

so the client can release global partitions using them. (It would be

good

to have some human readable string associated with the global result
partitions).
     - Client issues REST call to list / release global partitions.

A few thoughts / questions below:
1. Failure cases:
            * The TEs may remove the result partition while the RM does

not

know. In this case, the client will receive a runtime error and submit

the

full DAG to recompute the missing result partition. In this case, RM

should

release the incomplete global partition. How would RM be notified to do
that?
            * Is it possible the RM looses global partition metadata

while

the TE still host the data? For example, RM deletes the global partition
entry while the release partition call to TE failed.
            * What would happen if the JM fails before the global

partitions

are registered to RM? Are users exposed to resource leak if JM does not
have HA?
            * What would happen if the RM fails? Will TE release the
partitions by themselves?

2. It looks that TE should be the source of truth of the result

partition

existence. Does it have to distinguish between global and local result
partitions? If TE does not need to distinguish them, it seems the the
releasePartition() method in TE could just provide the list of

partitions

to release, without the partitions to promote.

3. In the current design, RM should be able to release result
partitions using ShuffleService. Will RM do this by sending RPC to the

TEs?

Or will the RM do it by itself?

4. How do we plan to handle the case when there are different shuffle
services in the same Flink cluster? For example, a shared standalone
cluster.

5. Minor: usually REST API uses `?` to pass the parameters. Is there a
reason we use `:` instead?

Thanks,

Jiangjie (Becket) Qin

On Tue, Sep 17, 2019 at 3:22 AM zhijiang<wangzhijiang...@aliyun.com.invalid> 
<wangzhijiang...@aliyun.com.invalid> wrote:


Thanks Chesnay for this FLIP and sorry for touching it a bit delay on

my

side.

I also have some similar concerns which Till already proposed before.

1. The consistent terminology in different components. On JM side,
PartitionTracker#getPersistedBlockingPartitions is defined for getting
global partitions. And on RM side, we define the method of
#registerGlobalPartitions correspondingly for handover the partitions

from

JM. I think it is better to unify the term in different components for

for

better understanding the semantic. Concering whether to use global or
persistent, I prefer the "global" term personally. Because it

describes the

scope of partition clearly, and the "persistent" is more like the

partition

storing way or implementation detail. In other words, the global

partition

might also be cached in memory of TE, not must persist into files from
semantic requirements. Whether memory or persistent file is just the
implementation choice.

2. On TE side, we might rename the method #releasePartitions to
#releaseOrPromotePartitions which describes the function precisely and
keeps consistent with
PartitionTracker#stopTrackingAndReleaseOrPromotePartitionsFor().

3. Very agree with Till's suggestions of global PartitionTable on TE

side

and sticking to TE's heartbeat report to RM for global partitions.

4. Considering ShuffleMaster, it was built inside JM and expected to
interactive with JM before. Now the RM also needs to interactive with
ShuffleMaster to release global partitions. Then it might be better to

move

ShuffleMaster outside of JM, and the lifecycle of ShuffleMaster should

be

consistent with RM.

5. Nit: TM->TE in the section of Proposed Changes: "TMs retain global
partitions for successful jobs"

Best,
Zhijiang


------------------------------------------------------------------
From:Till Rohrmann <trohrm...@apache.org> <trohrm...@apache.org>
Send Time:2019年9月10日(星期二) 10:10
To:dev <dev@flink.apache.org> <dev@flink.apache.org>
Subject:Re: [DISCUSS] FLIP-67: Global partitions lifecycle

Thanks Chesnay for drafting the FLIP and starting this discussion.

I have a couple of comments:

* I know that I've also coined the terms global/local result partition

but

maybe it is not the perfect name. Maybe we could rethink the

terminology

and call them persistent result partitions?
* Nit: I would call the last parameter of void releasePartitions(JobID
jobId, Collection<ResultPartitionID> partitionsToRelease,
Collection<ResultPartitionID> partitionsToPromote) either
partitionsToRetain or partitionsToPersistent.
* I'm not sure whether partitionsToRelease should contain a
global/persistent result partition id. I always thought that the user

will

be responsible for managing the lifecycle of a global/persistent
result partition.
* Instead of extending the PartitionTable to be able to store
global/persistent and local/transient result partitions, I would rather
introduce a global PartitionTable to store the global/persistent result
partitions explicitly. I think there is a benefit in making things as
explicit as possible.
* The handover logic between the JM and the RM for the

global/persistent

result partitions seems a bit brittle to me. What will happen if the JM
cannot reach the RM? I think it would be better if the TM announces the
global/persistent result partitions to the RM via its heartbeats. That

way

we don't rely on an established connection between the JM and RM and we
keep the TM as the ground of truth. Moreover, the RM should simply

forward

the release calls to the TM without much internal logic.

Cheers,
Till

On Fri, Sep 6, 2019 at 3:16 PM Chesnay Schepler <ches...@apache.org> 
<ches...@apache.org>
wrote:


Hello,

FLIP-36 (interactive programming)
<


https://cwiki.apache.org/confluence/display/FLINK/FLIP-36%3A+Support+Interactive+Programming+in+Flink

proposes a new programming paradigm where jobs are built incrementally
by the user.

To support this in an efficient manner I propose to extend partition
life-cycle to support the notion of /global partitions/, which are
partitions that can exist beyond the life-time of a job.

These partitions could then be re-used by subsequent jobs in a fairly
efficient manner, as they don't have to persisted to an external

storage

first and consuming tasks could be scheduled to exploit data-locality.

The FLIP outlines the required changes on the JobMaster, TaskExecutor
and ResourceManager to support this from a life-cycle perspective.

This FLIP does /not/ concern itself with the /usage/ of global
partitions, including client-side APIs, job-submission, scheduling and
reading said partitions; these are all follow-ups that will either be
part of FLIP-36 or spliced out into separate FLIPs.






Reply via email to