xintongsong commented on a change in pull request #11323: [FLINK-16439][k8s]
Make KubernetesResourceManager starts workers using WorkerResourceSpec
requested by SlotManager
URL: https://github.com/apache/flink/pull/11323#discussion_r394865352
##########
File path:
flink-kubernetes/src/main/java/org/apache/flink/kubernetes/KubernetesResourceManager.java
##########
@@ -81,10 +80,8 @@
private final FlinkKubeClient kubeClient;
- private final ContaineredTaskManagerParameters taskManagerParameters;
-
- /** The number of pods requested, but not yet granted. */
- private int numPendingPodRequests = 0;
+ /** Map from pod name to worker resource. */
+ private final Map<String, WorkerResourceSpec> podWorkerResources;
Review comment:
1. True. I'll try to clean up when the pod with corresponding name is
removed from `workerNodes`. I believe that means we no longer need its resource
information.
2. I think it's by design. The original behavior is that RM asks SM for
pending slots, and will only request new pod only if there are more pending
slots than the slots of the pending TM/pods. In case of JM failover, there
should be no pending slots from previous attempt. Therefore, only slots
requested in the new attempt is considered. I think the current behavior aligns
with the original.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
