Github user Ngone51 commented on a diff in the pull request: https://github.com/apache/spark/pull/20604#discussion_r182086337 --- Diff: core/src/main/scala/org/apache/spark/SparkContext.scala --- @@ -1643,7 +1646,10 @@ class SparkContext(config: SparkConf) extends Logging { def killExecutors(executorIds: Seq[String]): Boolean = { schedulerBackend match { case b: ExecutorAllocationClient => - b.killExecutors(executorIds, replace = false, force = true).nonEmpty + require(executorAllocationManager.isEmpty, --- End diff -- Hi, @squito , I'm quite questioned about the cases: > If you've got just one executor, and then you kill it, should your app sit with 0 executors? if app sit with 0 executors, then pending tasks increase, which lead to `ExecutorAllocationManager` increases target number of executors. So, app will not always sit with 0 executors. > Or even if you've got 10 executors, and you kill one -- when is dynamic allocation allowed to bump the total back up? for this case, to be honest, I really do not get your point. But, it must blame my poor English. And, what will happens if we use this method without `ExecutorAllocationManager `? Or do we really need adjust TargetNumExecutors (set `adjustTargetNumExecutors = true` below) if we are not using `ExecutorAllocationManager `? see these several lines in `killExecutors()`: ``` if (adjustTargetNumExecutors) { requestedTotalExecutors = math.max(requestedTotalExecutors - executorsToKill.size, 0) ... doRequestTotalExecutors(requestedTotalExecutors) } ``` Set `adjustTargetNumExecutors = true` will change `requestedTotalExecutors` . And IIUC, `requestedTotalExecutors ` is only used in dynamic allocation mode. So, if we are not using `ExecutorAllocationManager `, allocation client will request `requestedTotalExecutors = 0` number of executors to cluster manager (this is really terrible). But, actually, app without `ExecutorAllocationManager ` do not have a limit requesting executors (in default). Actually, I think this series methods, including `killAndReplaceExecutor `, `requestExecutors`, etc, are designed with dynamic allocation mode. And if we still want use these methods while app do not use `ExecutorAllocationManager`, we should not change `requestedTotalExecutors `, or even not request cluster manager with a specific number. WDYT?
--- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org