Zhu Zhu created FLINK-17019:
-------------------------------
Summary: Implement FIFO Physical Slot Assignment in SlotPoolImpl
Key: FLINK-17019
URL: https://issues.apache.org/jira/browse/FLINK-17019
Project: Flink
Issue Type: Sub-task
Components: Runtime / Coordination
Affects Versions: 1.11.0
Reporter: Zhu Zhu
Fix For: 1.11.0
The SlotPool should try to fulfill the oldest pending slot request once it
receives an available slot, no matter if the slot is returned by another
terminated task or is just offered from a task manager. This naturally ensures
that slot requests of an earlier scheduled region will be fulfilled earlier
than requests of a later scheduled region.
We only need to change the slot assignment logic on slot offers. This is
because the fields {{pendingRequests}} and {{waitingForResourceManager}} store
the pending requests in LinkedHashMaps . Therefore,
{{tryFulfillSlotRequestOrMakeAvailable(...)}} will naturally fulfill the
pending requests in inserted order.
When a new slot is offered via {{SlotPoolImpl#offerSlot(...)}} , we should use
it to fulfill the oldest fulfillable slot request directly by invoking
{{tryFulfillSlotRequestOrMakeAvailable(...)}}.
If a pending request (say R1) exists with the allocationId of the offered slot,
and it is different from the request to fulfill (say R2), we should update the
pendingRequest to replace AllocationID of R1 to be the AllocationID of R2. This
ensures failAllocation(...) can fail slot allocation requests to trigger
restarting tasks and re-allocating slots.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
