[
https://issues.apache.org/jira/browse/FLINK-4348?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15535527#comment-15535527
]
ASF GitHub Bot commented on FLINK-4348:
---------------------------------------
Github user mxm commented on the issue:
https://github.com/apache/flink/pull/2571
Thanks for the feedback, @beyond1920 and @KurtYoung.
You're right, the changes don't allow slots to be released by a
TaskExecutor. We can change that by explicitly reporting to the RM if a slot
becomes free. This may also decrease latency in case a tasks finishes and new
ones are waiting to be deployed.
>b. When we handleSlotRequestFailedAtTaskManager, we will make this slot
free again. If the slot is occupied by some other task now, we will
continuously failed for all allocation on this slot. ( this can be fixed by 3)
How can that happen? The slot will not appear free while it is allocated at
the TaskExecutor. When allocation fails, it is marked as free and then the
request is retried immediately. It must succeed eventually if the initial
decision to allocate the slot was correct. However, we need to explicitly check
if a TaskExecutor has deregistered, to make sure old TaskExecutors don't send
failures which triggers slot allocation of already removed slots (due to
TaskExecutor deregistration). That should be fix with this PR.
> 1. We can remove the update status part entirely, since it can only do
new slot registration now, we can just move it to the task executor first
registration.
Very good suggestion. Let's move the initial registration and
reconciliation of slots to the registration message.
To wrap up, let's change the following:
1. Move the slot registration and allocation report to the registration of
the TaskExecutor
2. Let the TaskExecutor immediately notify the ResourceManager once a slot
becomes free
3. Change the fencing in handleSlotRequestFailedAtTaskManager to protect
against TaskExecutors which are not registered anymore.
Let me know if that would work for you.
> Implement slot allocation protocol with TaskExecutor
> ----------------------------------------------------
>
> Key: FLINK-4348
> URL: https://issues.apache.org/jira/browse/FLINK-4348
> Project: Flink
> Issue Type: Sub-task
> Components: Cluster Management
> Reporter: Kurt Young
> Assignee: Maximilian Michels
>
> When slotManager finds a proper slot in the free pool for a slot request,
> slotManager marks the slot as occupied, then tells the taskExecutor to give
> the slot to the specified JobMaster.
> when a slot request is sent to taskExecutor, it should contain following
> parameters: AllocationID, JobID, slotID, resourceManagerLeaderSessionID.
> There exists 3 following possibilities of the response from taskExecutor, we
> will discuss when each possibility happens and how to handle.
> 1. Ack request which means the taskExecutor gives the slot to the specified
> jobMaster as expected.
> 2. Decline request if the slot is already occupied by other AllocationID.
> 3. Timeout which could caused by lost of request message or response message
> or slow network transfer.
> On the first occasion, ResourceManager need to do nothing. However, under the
> second and third occasion, ResourceManager need to notify slotManager,
> slotManager will verify and clear all the previous allocate information for
> this slot request firstly, then try to find a proper slot for the slot
> request again. This may cause some duplicate allocation, e.g. the slot
> request to TaskManager is successful but the response is lost somehow, so we
> may request a slot in another TaskManager, this causes two slots assigned to
> one request, but it can be taken care of by rejecting registration at
> JobMaster.
> There are still some question need to discuss in a step further.
> 1. Who send slotRequest to taskExecutor, SlotManager or ResourceManager? I
> think it's better that SlotManager delegates the rpc call to ResourceManager
> when SlotManager need to communicate with outside world. ResourceManager
> know which taskExecutor to send the request based on ResourceID. Besides this
> RPC call which used to request slot to taskExecutor should not be a
> RpcMethod, because we hope only SlotManager has permission to call the
> method, but the other component, for example JobMaster and TaskExecutor,
> cannot call this method directly.
> 2. If JobMaster reject the slot offer from a TaskExecutor, the TaskExecutor
> should notify the free slot to ResourceManager immediately, or wait for next
> heartbeat sync. The advantage of first way is the resourceManager’s view
> could be updated faster. The advantage of second way is save a RPC method in
> ResourceManager.
> 3. There are two communication type. First, the slot request could be sent as
> an ask operation where the response is returned as a future. Second,
> resourceManager send the slot request in fire and forget way, the response
> could be returned by an RPC call. I prefer the first one because it is more
> simple and could save a RPC method in ResourceManager (for callback in the
> second way).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
