[SPARK-35011][CORE][3.1] Avoid Block Manager registrations when StopExecutor msg is in-flight #33771
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…or msg is in-flight ### What changes were proposed in this pull request? This patch proposes a fix to prevent triggering BlockManager reregistration while `StopExecutor` msg is in-flight. Here on receiving `StopExecutor` msg, we do not remove the corresponding `BlockManagerInfo` from `blockManagerInfo` map, instead we mark it as dead by updating the corresponding `executorRemovalTs`. There's a separate cleanup thread running to periodically remove the stale `BlockManagerInfo` from `blockManangerInfo` map. Now if a recently removed `BlockManager` tries to register, the driver simply ignores it since the `blockManagerInfo` map already contains an entry for it. The same applies to `BlockManagerHeartbeat`, if the BlockManager belongs to a recently removed executor, the `blockManagerInfo` map would contain an entry and we shall not ask the corresponding `BlockManager` to re-register. ### Why are the changes needed? This changes are needed since BlockManager reregistration while executor is shutting down causes inconsistent bookkeeping of executors in Spark. Consider the following scenario: - `CoarseGrainedSchedulerBackend` issues async `StopExecutor` on executorEndpoint - `CoarseGrainedSchedulerBackend` removes that executor from Driver's internal data structures and publishes `SparkListenerExecutorRemoved` on the `listenerBus`. - Executor has still not processed `StopExecutor` from the Driver - Driver receives heartbeat from the Executor, since it cannot find the `executorId` in its data structures, it responds with `HeartbeatResponse(reregisterBlockManager = true)` - `BlockManager` on the Executor reregisters with the `BlockManagerMaster` and `SparkListenerBlockManagerAdded` is published on the `listenerBus` - Executor starts processing the `StopExecutor` and exits - `AppStatusListener` picks the `SparkListenerBlockManagerAdded` event and updates `AppStatusStore` - `statusTracker.getExecutorInfos` refers `AppStatusStore` to get the list of executors which returns the dead executor as alive. ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? - Modified the existing unittests. - Ran a simple test application on minikube that asserts on number of executors are zero once the executor idle timeout is reached. Closes apache#32114 from sumeetgajjar/SPARK-35011. Authored-by: Sumeet Gajjar <[email protected]> Signed-off-by: yi.wu <[email protected]>
|
Can one of the admins verify this patch? |
dongjoon-hyun
approved these changes
Aug 17, 2021
zhuqi-lucas
approved these changes
Aug 18, 2021
|
Merged to branch-3.1. Thank you, @sumeetgajjar and @zhuqi-lucas . |
dongjoon-hyun
pushed a commit
that referenced
this pull request
Aug 18, 2021
…xecutor msg is in-flight This PR backports #32114 to 3.1 <hr> ### What changes were proposed in this pull request? This patch proposes a fix to prevent triggering BlockManager reregistration while `StopExecutor` msg is in-flight. Here on receiving `StopExecutor` msg, we do not remove the corresponding `BlockManagerInfo` from `blockManagerInfo` map, instead we mark it as dead by updating the corresponding `executorRemovalTs`. There's a separate cleanup thread running to periodically remove the stale `BlockManagerInfo` from `blockManangerInfo` map. Now if a recently removed `BlockManager` tries to register, the driver simply ignores it since the `blockManagerInfo` map already contains an entry for it. The same applies to `BlockManagerHeartbeat`, if the BlockManager belongs to a recently removed executor, the `blockManagerInfo` map would contain an entry and we shall not ask the corresponding `BlockManager` to re-register. ### Why are the changes needed? This changes are needed since BlockManager reregistration while executor is shutting down causes inconsistent bookkeeping of executors in Spark. Consider the following scenario: - `CoarseGrainedSchedulerBackend` issues async `StopExecutor` on executorEndpoint - `CoarseGrainedSchedulerBackend` removes that executor from Driver's internal data structures and publishes `SparkListenerExecutorRemoved` on the `listenerBus`. - Executor has still not processed `StopExecutor` from the Driver - Driver receives heartbeat from the Executor, since it cannot find the `executorId` in its data structures, it responds with `HeartbeatResponse(reregisterBlockManager = true)` - `BlockManager` on the Executor reregisters with the `BlockManagerMaster` and `SparkListenerBlockManagerAdded` is published on the `listenerBus` - Executor starts processing the `StopExecutor` and exits - `AppStatusListener` picks the `SparkListenerBlockManagerAdded` event and updates `AppStatusStore` - `statusTracker.getExecutorInfos` refers `AppStatusStore` to get the list of executors which returns the dead executor as alive. ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? - Modified the existing unittests. - Ran a simple test application on minikube that asserts on number of executors are zero once the executor idle timeout is reached. Closes #33771 from sumeetgajjar/SPARK-35011-br-3.1. Authored-by: Sumeet Gajjar <[email protected]> Signed-off-by: Dongjoon Hyun <[email protected]>
|
Thank you @dongjoon-hyun and @zhuqi-lucas for approving this PR. |
|
Thanks for the work @sumeetgajjar |
dongjoon-hyun
pushed a commit
that referenced
this pull request
Sep 10, 2021
…en StopExecutor msg is in-flight" This reverts commit 79ea014. ### What changes were proposed in this pull request? Revert #33771 ### Why are the changes needed? It breaks the expected `BlockManager` re-registration (e.g., heartbeat loss of an active executor) due to deferred removal of `BlockManager`, see the check: https://github.com/apache/spark/blob/9cefde8db373a3433b7e3ce328e4a2ce83b1aca2/core/src/main/scala/org/apache/spark/storage/BlockManagerMasterEndpoint.scala#L551 ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? Pass existing tests. Closes #33960 from Ngone51/revert-35011-3.1. Authored-by: yi.wu <[email protected]> Signed-off-by: Dongjoon Hyun <[email protected]>
flyrain
pushed a commit
to flyrain/spark
that referenced
this pull request
Sep 21, 2021
…xecutor msg is in-flight This PR backports apache#32114 to 3.1 <hr> This patch proposes a fix to prevent triggering BlockManager reregistration while `StopExecutor` msg is in-flight. Here on receiving `StopExecutor` msg, we do not remove the corresponding `BlockManagerInfo` from `blockManagerInfo` map, instead we mark it as dead by updating the corresponding `executorRemovalTs`. There's a separate cleanup thread running to periodically remove the stale `BlockManagerInfo` from `blockManangerInfo` map. Now if a recently removed `BlockManager` tries to register, the driver simply ignores it since the `blockManagerInfo` map already contains an entry for it. The same applies to `BlockManagerHeartbeat`, if the BlockManager belongs to a recently removed executor, the `blockManagerInfo` map would contain an entry and we shall not ask the corresponding `BlockManager` to re-register. This changes are needed since BlockManager reregistration while executor is shutting down causes inconsistent bookkeeping of executors in Spark. Consider the following scenario: - `CoarseGrainedSchedulerBackend` issues async `StopExecutor` on executorEndpoint - `CoarseGrainedSchedulerBackend` removes that executor from Driver's internal data structures and publishes `SparkListenerExecutorRemoved` on the `listenerBus`. - Executor has still not processed `StopExecutor` from the Driver - Driver receives heartbeat from the Executor, since it cannot find the `executorId` in its data structures, it responds with `HeartbeatResponse(reregisterBlockManager = true)` - `BlockManager` on the Executor reregisters with the `BlockManagerMaster` and `SparkListenerBlockManagerAdded` is published on the `listenerBus` - Executor starts processing the `StopExecutor` and exits - `AppStatusListener` picks the `SparkListenerBlockManagerAdded` event and updates `AppStatusStore` - `statusTracker.getExecutorInfos` refers `AppStatusStore` to get the list of executors which returns the dead executor as alive. No - Modified the existing unittests. - Ran a simple test application on minikube that asserts on number of executors are zero once the executor idle timeout is reached. Closes apache#33771 from sumeetgajjar/SPARK-35011-br-3.1. Authored-by: Sumeet Gajjar <[email protected]> Signed-off-by: Dongjoon Hyun <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR backports #32114 to 3.1
What changes were proposed in this pull request?
This patch proposes a fix to prevent triggering BlockManager reregistration while
StopExecutormsg is in-flight.Here on receiving
StopExecutormsg, we do not remove the correspondingBlockManagerInfofromblockManagerInfomap, instead we mark it as dead by updating the correspondingexecutorRemovalTs. There's a separate cleanup thread running to periodically remove the staleBlockManagerInfofromblockManangerInfomap.Now if a recently removed
BlockManagertries to register, the driver simply ignores it since theblockManagerInfomap already contains an entry for it. The same applies toBlockManagerHeartbeat, if the BlockManager belongs to a recently removed executor, theblockManagerInfomap would contain an entry and we shall not ask the correspondingBlockManagerto re-register.Why are the changes needed?
This changes are needed since BlockManager reregistration while executor is shutting down causes inconsistent bookkeeping of executors in Spark.
Consider the following scenario:
CoarseGrainedSchedulerBackendissues asyncStopExecutoron executorEndpointCoarseGrainedSchedulerBackendremoves that executor from Driver's internal data structures and publishesSparkListenerExecutorRemovedon thelistenerBus.StopExecutorfrom the DriverexecutorIdin its data structures, it responds withHeartbeatResponse(reregisterBlockManager = true)BlockManageron the Executor reregisters with theBlockManagerMasterandSparkListenerBlockManagerAddedis published on thelistenerBusStopExecutorand exitsAppStatusListenerpicks theSparkListenerBlockManagerAddedevent and updatesAppStatusStorestatusTracker.getExecutorInfosrefersAppStatusStoreto get the list of executors which returns the dead executor as alive.Does this PR introduce any user-facing change?
No
How was this patch tested?