This is part of a broader effort to enable dynamic scaling of executors ([SPARK-3174](https://issues.apache.org/jira/browse/SPARK-3174)). This is intended to work alongside SPARK-3795 (#2746), SPARK-3796 and SPARK-3797, but is functionally independently of these other issues.
The logic is built on top of PraveenSeluka's changes at #2798. This is different from the changes there in a few major ways: (1) the mechanism is implemented within the existing scheduler backend framework rather than in new `Actor` classes. This also introduces a parent abstract class `YarnSchedulerBackend` to encapsulate common logic to communicate with the Yarn `ApplicationMaster`. (2) The interface of requesting executors exposed to the `SparkContext` is the same, but the communication between the scheduler backend and the AM uses total number executors desired instead of an incremental number. This is discussed in #2746 and explained in the comments in the code.
I have tested this significantly on a stable Yarn cluster.
------------
A remaining task for this issue is to tone down the error messages emitted when an executor is removed.
Currently, `SparkContext` and its components react as if the executor has failed, resulting in many scary error messages and eventual timeouts. While it's not strictly necessary to fix this as of the first-cut implementation of this mechanism, it would be good to add logic to distinguish this case. I prefer to address this in a separate PR. I have filed a separate JIRA for this task at SPARK-4134.
Author: Andrew Or <andrew@databricks.com>
Author: Andrew Or <andrewor14@gmail.com>
Closes#2840 from andrewor14/yarn-scaling-mechanism and squashes the following commits:
485863e [Andrew Or] Minor log message changes
4920be8 [Andrew Or] Clarify that public API is only for Yarn mode for now
1c57804 [Andrew Or] Reword a few comments + other review comments
6321140 [Andrew Or] Merge branch 'master' of github.com:apache/spark into yarn-scaling-mechanism
02836c0 [Andrew Or] Limit scope of synchronization
4e2ed7f [Andrew Or] Fix bug: keep track of removed executors properly
73ade46 [Andrew Or] Wording changes (minor)
2a7a6da [Andrew Or] Add `sc.killExecutor` as a shorthand (minor)
665f229 [Andrew Or] Mima excludes
79aa2df [Andrew Or] Simplify the request interface by asking for a total
04f625b [Andrew Or] Fix race condition that causes over-allocation of executors
f4783f8 [Andrew Or] Change the semantics of requesting executors
005a124 [Andrew Or] Fix tests
4628b16 [Andrew Or] Merge branch 'master' of github.com:apache/spark into yarn-scaling-mechanism
db4a679 [Andrew Or] Merge branch 'master' of github.com:apache/spark into yarn-scaling-mechanism
572f5c5 [Andrew Or] Unused import (minor)
f30261c [Andrew Or] Kill multiple executors rather than one at a time
de260d9 [Andrew Or] Simplify by skipping useless null check
9c52542 [Andrew Or] Simplify by skipping the TaskSchedulerImpl
97dd1a8 [Andrew Or] Merge branch 'master' of github.com:apache/spark into yarn-scaling-mechanism
d987b3e [Andrew Or] Move addWebUIFilters to Yarn scheduler backend
7b76d0a [Andrew Or] Expose mechanism in SparkContext as developer API
47466cd [Andrew Or] Refactor common Yarn scheduler backend logic
c4dfaac [Andrew Or] Avoid thrashing when removing executors
53e8145 [Andrew Or] Start yarn actor early to listen for AM registration message
bbee669 [Andrew Or] Add mechanism in yarn client mode
1df05a40eb