...

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Motivation

As described in FLINK-15959, currently ResourceManager will only start a a new worker when there's resource requirements from job. But in certain scenarios, it is necessary to maintain a certain amount of slots as soon as ResourceManager is started. And these slots should not release even if all slots are free. Here are some scenarios which can benefit from this proposal:

Currently, the Flink cluster only requests Task Managers (TMs) when there is a resource requirement, and idle TMs are released after a certain period of time. However, in certain scenarios, we need to improve the efficiency of job execution by maintaining a certain number of available resources in the system. For example:

1. For session clusters running short lived-jobs like OLAP, we should treat session clusters as a long-running service. Having running workers all the time can greatly improve the service stability and reduce the job's cold-start latency;

2. For application mode, batch job might be scheduled stage by stage. If the next region requires more resources, it might take more time to pull up required resources. If users

Users

3. actually know how many resources are needed when running a single job, initialize all workers when cluster starts can speed up

startup process

4. Job schedule in topology order, the next operator won't schedule until prior execution slot allocated. The TaskExecutors will start in several batches . In some cases, it might slow down the startup speed

5. Flink supports FLINK-12122 [Spread out tasks evenly across all available registered TaskManagers], but it will only effect if all TMs are registered. Starting all TMs at the beginning can solve this problem

For session clusters running short lived-jobs, clusters should have workers all the time to reduce job's E2E latency

6. the resource allocation process;

Public Interfaces

Notice: Like the maximum resource options, the minimum resource options will also be invalidated in standalone clusters.

Proposed Changes

Minimum required resource check in DefaultResourceAllocationStrategy

The minimum resource allocation limitation will be implemented in the DefaultResourceAllocationStrategy of FineGrainedSlotManager.

Each time when SlotManager needs to reconcile the cluster resources or fulfill job resource requirements, the DefaultResourceAllocationStrategy will check if the minimum resource requirement has been fulfilled. If it is not, DefaultResourceAllocationStrategy will add PendingTaskManagers to the ResourceAllocationResult/ResourceReconcileResult request new PendingTaskManagers and FineGrainedSlotManager will allocate new worker resources accordingly.

Notice: The minimum resource limitation will be fulfilled in the best effort manner. The unfulfilled request will be retried in the next resource check but will not block job scheduling.

Resource stability check in SlotManagerConfiguration

After minimum required resource options are introduced, users can configure both minimum and maximum required resources in SlotManager, which might lead to worker resource oscillation. Here is an example:

...

To avoid this oscillation, we need to check the worker number derived from minimum and maximum resource configuration is consistent before starting SlotManager.

Notice: The resource stability check will be disabled when the minimum resource option is 0 which is set as default.

Compatibility, Deprecation, and Migration Plan

• What impact (if any) will there be on existing users?  No

Test Plan

UT & Manually verify that the min resource allocation limitation is working as expected.

Rejected Alternatives

If there are alternative ways of accomplishing the same thing, what were they? The purpose of this section is to motivate why the design is the way it is and not some other way.