Details
-
Improvement
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
-
None
-
Reviewed
Description
We have did some usability improvements for scheduler activities based on YARN3.1 in our cluster as follows:
1. Not available for multi-thread asynchronous scheduling. App and node activities maybe confused when multiple scheduling threads record activities of different allocation processes in the same variables like appsAllocation and recordingNodesAllocation in ActivitiesManager. I think these variables should be thread-local to make activities clear among multiple threads.
2. Incomplete activities for multi-node lookup mechanism, since ActivitiesLogger will skip recording through {{if (node == null || activitiesManager == null) }} when node is null which represents this allocation is for multi-nodes. We need support recording activities for multi-node lookup mechanism.
3. Current app activities can not meet requirements of diagnostics, for example, we can know that node doesn't match request but hard to know why, especially when using placement constraints, it's difficult to make a detailed diagnosis manually. So I propose to improve the diagnoses of activities, add diagnosis for placement constraints check, update insufficient resource diagnosis with detailed info (like 'insufficient resource names:[memory-mb]') and so on.
4. Add more useful fields for app activities, in some scenarios we need to distinguish different requests but can't locate requests based on app activities info, there are some other fields can help to filter what we want such as allocation tags. We have added containerPriority, allocationRequestId and allocationTags fields in AppAllocation.
5. Filter app activities by key fields, sometimes the results of app activities is massive, it's hard to find what we want. We have support filter by allocation-tags to meet requirements from some apps, more over, we can take container-priority and allocation-request-id as candidates if necessary.
6. Aggregate app activities by diagnoses. For a single allocation process, activities still can be massive in a large cluster, we frequently want to know why request can't be allocated in cluster, it's hard to check every node manually in a large cluster, so that aggregation for app activities by diagnoses is necessary to solve this trouble. We have added groupingType parameter for app-activities REST API for this, supports grouping by diagnostics.
I think we can have a discuss about these points, useful improvements which can be accepted will be added into the patch. Thanks.
Running design doc is attached here.
