[SPARK-32199][SPARK-32198] Reduce job failures during decommissioning #29014
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
agrawaldevesh
force-pushed
the
decom_harden
branch
from
b6690d8 to
40cfea9
Compare
|
@holdenk ... can you take a quick look at the approach here please. Happy to chat further offline. Also happy to fold this PR (entirely or parts of it) into your PR. |
agrawaldevesh
force-pushed
the
decom_harden
branch
2 times, most recently
from
a4d7185 to
2254234
Compare
agrawaldevesh
changed the title
agrawaldevesh
force-pushed
the
decom_harden
branch
from
2254234 to
e7113a6
Compare
|
I'd rather avoid the hack configuration flag and just go straight to extending the current message. What do you think? |
|
If you need the changes I've made in the other (not yet merged PRs), feel free to base your PR on top of that and just make a note that it's the case for merge ordering and so folks don't duplicate review effort :) |
Sure !. I think obviously that is cleaner. If you aren't working on that extension, I would be happy to take a stab at that plumbing in a separate PR. In the meanwhile, I can also just back out that config since it is only needed when there are more than one executors in a node with external shuffle service. I will amend this PR to back out that config. |
agrawaldevesh
changed the title
agrawaldevesh
force-pushed
the
decom_harden
branch
2 times, most recently
from
bbee57d to
01ebb47
Compare
agrawaldevesh
changed the title
|
@holdenk, @jiangxb1987 @cloud-fan @Ngone51 -- This PR is ready for your review please. Thanks ! |
jiangxb1987
reviewed
Jul 10, 2020
core/src/main/scala/org/apache/spark/scheduler/DecommissionInfo.scala
Outdated
Show resolved
Hide resolved
core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala
Outdated
Show resolved
Hide resolved
core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala
Outdated
Show resolved
Hide resolved
core/src/main/scala/org/apache/spark/scheduler/ExecutorLossReason.scala
Outdated
Show resolved
Hide resolved
core/src/main/scala/org/apache/spark/scheduler/TaskSchedulerImpl.scala
Outdated
Show resolved
Hide resolved
Ngone51
reviewed
Jul 13, 2020
core/src/main/scala/org/apache/spark/scheduler/ExecutorLossReason.scala
Outdated
Show resolved
Hide resolved
Ngone51
reviewed
Jul 13, 2020
core/src/main/scala/org/apache/spark/scheduler/TaskSchedulerImpl.scala
Outdated
Show resolved
Hide resolved
core/src/main/scala/org/apache/spark/scheduler/TaskSchedulerImpl.scala
Outdated
Show resolved
Hide resolved
core/src/main/scala/org/apache/spark/scheduler/TaskSchedulerImpl.scala
Outdated
Show resolved
Hide resolved
core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala
Outdated
Show resolved
Hide resolved
core/src/main/scala/org/apache/spark/scheduler/ExecutorLossReason.scala
Outdated
Show resolved
Hide resolved
core/src/main/scala/org/apache/spark/executor/CoarseGrainedExecutorBackend.scala
Outdated
Show resolved
Hide resolved
core/src/main/scala/org/apache/spark/scheduler/DecommissionInfo.scala
Outdated
Show resolved
Hide resolved
agrawaldevesh
force-pushed
the
decom_harden
branch
from
d36097c to
5ce99b1
Compare
agrawaldevesh
force-pushed
the
decom_harden
branch
2 times, most recently
from
8d123c3 to
92d5e71
Compare
|
I cannot reproduce the failures in the |
agrawaldevesh
force-pushed
the
decom_harden
branch
2 times, most recently
from
c476ce1 to
30af8f4
Compare
|
ok to test |
|
Test build #125948 has finished for PR 29014 at commit
|
Ngone51
reviewed
Jul 16, 2020
core/src/main/scala/org/apache/spark/scheduler/ExecutorDecommissionInfo.scala
Outdated
Show resolved
Hide resolved
attilapiros
reviewed
Jul 28, 2020
core/src/test/scala/org/apache/spark/deploy/DecommissionWorkerSuite.scala
Outdated
Show resolved
Hide resolved
|
Test build #126713 has finished for PR 29014 at commit
|
attilapiros
reviewed
Jul 28, 2020
core/src/test/scala/org/apache/spark/deploy/DecommissionWorkerSuite.scala
Outdated
Show resolved
Hide resolved
attilapiros
reviewed
Jul 28, 2020
core/src/test/scala/org/apache/spark/deploy/DecommissionWorkerSuite.scala
Outdated
Show resolved
Hide resolved
attilapiros
reviewed
Jul 28, 2020
core/src/main/scala/org/apache/spark/scheduler/TaskSchedulerImpl.scala
Outdated
Show resolved
Hide resolved
|
@agrawaldevesh we are really close, sorry for being this picky (today) but I am sure this extra care will serve us well. |
attilapiros
reviewed
Jul 28, 2020
core/src/test/scala/org/apache/spark/deploy/DecommissionWorkerSuite.scala
Outdated
Show resolved
Hide resolved
agrawaldevesh
force-pushed
the
decom_harden
branch
from
806630a to
153e8e4
Compare
@attilapiros Always always appreciated !. Better to be nitpicky once than confuse a thousand future readers :-P I have reworked the whole config confusion in the test. Now there are only the following configs:
This has really cleaned up the config mess so thanks for spotting this. |
agrawaldevesh
force-pushed
the
decom_harden
branch
3 times, most recently
from
1fd2cca to
a906041
Compare
|
Test build #126729 has finished for PR 29014 at commit
|
|
LGTM |
|
Locally lint runs successfully: |
attilapiros
approved these changes
Jul 28, 2020
|
Test build #126730 has finished for PR 29014 at commit
|
|
Test build #126731 has finished for PR 29014 at commit
|
This PR reduces the prospect of a job loss during decommissioning. It fixes two holes in the current decommissioning framework: - (a) Loss of decommissioned executors is not treated as a job failure: We know that the decom'd executor would be dying soon, so its death is clearly not caused by the application. - (b) Shuffle files on the decommissioned host are cleared when the first fetch failure is detected from a decommissioned host: This is a bit tricky in terms of when to clear the shuffle state ? Ideally you want to clear it the millisecond before the shuffle service on the node dies (or the executor dies when there is no external shuffle service) -- too soon and it could lead to some wasteage and too late would lead to fetch failures. (These two things are a bit intertwined so it is easier to do them in a single commit) The approach here is to do this clearing when the very first fetch failure is observed on the decom'd block manager, without waiting for other blocks to also signal a failure. Added a new unit test `DecommissionWorkerSuite` to test the new behavior.
agrawaldevesh
force-pushed
the
decom_harden
branch
from
a906041 to
dbc478b
Compare
|
Thanks a lot for a very very solid review and co-authoring of this PR @attilapiros. I learnt a lot. @holdenk all yours now :-) |
|
Test build #126745 has finished for PR 29014 at commit
|
holdenk
approved these changes
Jul 30, 2020
|
FYI, I saw the flakiness of the test I will keep monitoring and see if it's frequent. |
asfgit
pushed a commit
that referenced
this pull request
Aug 26, 2020
…kSchedulerImpl realm ### What changes were proposed in this pull request? The decommissioning state is a bit fragment across two places in the TaskSchedulerImpl: #29014 stored the incoming decommission info messages in TaskSchedulerImpl.executorsPendingDecommission. While #28619 was storing just the executor end time in the map TaskSetManager.tidToExecutorKillTimeMapping (which in turn is contained in TaskSchedulerImpl). While the two states are not really overlapping, it's a bit of a code hygiene concern to save this state in two places. With #29422, TaskSchedulerImpl is emerging as the place where all decommissioning book keeping is kept within the driver. So consolidate the information in _tidToExecutorKillTimeMapping_ into _executorsPendingDecommission_. However, in order to do so, we need to walk away from keeping the raw ExecutorDecommissionInfo messages and instead keep another class ExecutorDecommissionState. This decoupling will allow the RPC message class ExecutorDecommissionInfo to evolve independently from the book keeping ExecutorDecommissionState. ### Why are the changes needed? This is just a code cleanup. These two features were added independently and its time to consolidate their state for good hygiene. ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? Existing tests. Closes #29452 from agrawaldevesh/consolidate_decom_state. Authored-by: Devesh Agrawal <devesh.agrawal@gmail.com> Signed-off-by: Holden Karau <hkarau@apple.com>
holdenk
added a commit
to holdenk/spark
that referenced
this pull request
Oct 27, 2020
[SPARK-21040][CORE] Speculate tasks which are running on decommission executors This PR adds functionality to consider the running tasks on decommission executors based on some config. In spark-on-cloud , we sometimes already know that an executor won't be alive for more than fix amount of time. Ex- In AWS Spot nodes, once we get the notification, we know that a node will be gone in 120 seconds. So if the running tasks on the decommissioning executors may run beyond currentTime+120 seconds, then they are candidate for speculation. Currently when an executor is decommission, we stop scheduling new tasks on those executors but the already running tasks keeps on running on them. Based on the cloud, we might know beforehand that an executor won't be alive for more than a preconfigured time. Different cloud providers gives different timeouts before they take away the nodes. For Ex- In case of AWS spot nodes, an executor won't be alive for more than 120 seconds. We can utilize this information in cloud environments and take better decisions about speculating the already running tasks on decommission executors. Yes. This PR adds a new config "spark.executor.decommission.killInterval" which they can explicitly set based on the cloud environment where they are running. Added UT. Closes apache#28619 from prakharjain09/SPARK-21040-speculate-decommission-exec-tasks. Authored-by: Prakhar Jain <prakharjain09@gmail.com> Signed-off-by: Holden Karau <hkarau@apple.com> [SPARK-20629][CORE][K8S] Copy shuffle data when nodes are being shutdown This pull request adds the ability to migrate shuffle files during Spark's decommissioning. The design document associated with this change is at https://docs.google.com/document/d/1xVO1b6KAwdUhjEJBolVPl9C6sLj7oOveErwDSYdT-pE . To allow this change the `MapOutputTracker` has been extended to allow the location of shuffle files to be updated with `updateMapOutput`. When a shuffle block is put, a block update message will be sent which triggers the `updateMapOutput`. Instead of rejecting remote puts of shuffle blocks `BlockManager` delegates the storage of shuffle blocks to it's shufflemanager's resolver (if supported). A new, experimental, trait is added for shuffle resolvers to indicate they handle remote putting of blocks. The existing block migration code is moved out into a separate file, and a producer/consumer model is introduced for migrating shuffle files from the host as quickly as possible while not overwhelming other executors. Recomputting shuffle blocks can be expensive, we should take advantage of our decommissioning time to migrate these blocks. This PR introduces two new configs parameters, `spark.storage.decommission.shuffleBlocks.enabled` & `spark.storage.decommission.rddBlocks.enabled` that control which blocks should be migrated during storage decommissioning. New unit test & expansion of the Spark on K8s decom test to assert that decommisioning with shuffle block migration means that the results are not recomputed even when the original executor is terminated. This PR is a cleaned-up version of the previous WIP PR I made apache#28331 (thanks to attilapiros for his very helpful reviewing on it :)). Closes apache#28708 from holdenk/SPARK-20629-copy-shuffle-data-when-nodes-are-being-shutdown-cleaned-up. Lead-authored-by: Holden Karau <hkarau@apple.com> Co-authored-by: Holden Karau <holden@pigscanfly.ca> Co-authored-by: “attilapiros” <piros.attila.zsolt@gmail.com> Co-authored-by: Attila Zsolt Piros <attilazsoltpiros@apiros-mbp16.lan> Signed-off-by: Holden Karau <hkarau@apple.com> [SPARK-24266][K8S] Restart the watcher when we receive a version changed from k8s Restart the watcher when it failed with a HTTP_GONE code from the kubernetes api. Which means a resource version has changed. For more relevant information see here: fabric8io/kubernetes-client#1075 No Running spark-submit to a k8s cluster. Not sure how to make an automated test for this. If someone can help me out that would be great. Closes apache#28423 from stijndehaes/bugfix/k8s-submit-resource-version-change. Authored-by: Stijn De Haes <stijndehaes@gmail.com> Signed-off-by: Holden Karau <hkarau@apple.com> [SPARK-32217] Plumb whether a worker would also be decommissioned along with executor This PR is a giant plumbing PR that plumbs an `ExecutorDecommissionInfo` along with the DecommissionExecutor message. The primary motivation is to know whether a decommissioned executor would also be loosing shuffle files -- and thus it is important to know whether the host would also be decommissioned. In the absence of this PR, the existing code assumes that decommissioning an executor does not loose the whole host with it, and thus does not clear the shuffle state if external shuffle service is enabled. While this may hold in some cases (like K8s decommissioning an executor pod, or YARN container preemption), it does not hold in others like when the cluster is managed by a Standalone Scheduler (Master). This is similar to the existing `workerLost` field in the `ExecutorProcessLost` message. In the future, this `ExecutorDecommissionInfo` can be embellished for knowing how long the executor has to live for scenarios like Cloud spot kills (or Yarn preemption) and the like. No Tweaked an existing unit test in `AppClientSuite` Closes apache#29032 from agrawaldevesh/plumb_decom_info. Authored-by: Devesh Agrawal <devesh.agrawal@gmail.com> Signed-off-by: Holden Karau <hkarau@apple.com> [SPARK-32199][SPARK-32198] Reduce job failures during decommissioning This PR reduces the prospect of a job loss during decommissioning. It fixes two holes in the current decommissioning framework: - (a) Loss of decommissioned executors is not treated as a job failure: We know that the decommissioned executor would be dying soon, so its death is clearly not caused by the application. - (b) Shuffle files on the decommissioned host are cleared when the first fetch failure is detected from a decommissioned host: This is a bit tricky in terms of when to clear the shuffle state ? Ideally you want to clear it the millisecond before the shuffle service on the node dies (or the executor dies when there is no external shuffle service) -- too soon and it could lead to some wastage and too late would lead to fetch failures. The approach here is to do this clearing when the very first fetch failure is observed on the decommissioned block manager, without waiting for other blocks to also signal a failure. Without them decommissioning a lot of executors at a time leads to job failures. The task scheduler tracks the executors that were decommissioned along with their `ExecutorDecommissionInfo`. This information is used by: (a) For handling a `ExecutorProcessLost` error, or (b) by the `DAGScheduler` when handling a fetch failure. No Added a new unit test `DecommissionWorkerSuite` to test the new behavior by exercising the Master-Worker decommissioning. I chose to add a new test since the setup logic was quite different from the existing `WorkerDecommissionSuite`. I am open to changing the name of the newly added test suite :-) - Should I add a feature flag to guard these two behaviors ? They seem safe to me that they should only get triggered by decommissioning, but you never know :-). Closes apache#29014 from agrawaldevesh/decom_harden. Authored-by: Devesh Agrawal <devesh.agrawal@gmail.com> Signed-off-by: Holden Karau <hkarau@apple.com> [SPARK-32417] Fix flakyness of BlockManagerDecommissionIntegrationSuite This test tries to fix the flakyness of BlockManagerDecommissionIntegrationSuite. Make the block manager decommissioning test be less flaky An interesting failure happens when migrateDuring = true (and persist or shuffle is true): - We schedule the job with tasks on executors 0, 1, 2. - We wait 300 ms and decommission executor 0. - If the task is not yet done on executor 0, it will now fail because the block manager won't be able to save the block. This condition is easy to trigger on a loaded machine where the github checks run. - The task with retry on a different executor (1 or 2) and its shuffle blocks will land there. - No actual block migration happens here because the decommissioned executor technically failed before it could even produce a block. To remove the above race, this change replaces the fixed wait for 300 ms to wait for an actual task to succeed. When a task has succeeded, we know its blocks would have been written for sure and thus its executor would certainly be forced to migrate those blocks when it is decommissioned. The change always decommissions an executor on which a real task finished successfully instead of picking the first executor. Because the system may choose to schedule nothing on the first executor and instead run the two tasks on one executor. I have had bad luck with BlockManagerDecommissionIntegrationSuite and it has failed several times on my PRs. So fixing it. No, unit test only change. Github checks. Ran this test 100 times, 10 at a time in parallel in a script. Closes apache#29226 from agrawaldevesh/block-manager-decom-flaky. Authored-by: Devesh Agrawal <devesh.agrawal@gmail.com> Signed-off-by: Holden Karau <hkarau@apple.com> [SPARK-31197][CORE] Shutdown executor once we are done decommissioning Exit the executor when it has been asked to decommission and there is nothing left for it to do. This is a rebase of apache#28817 If we want to use decommissioning in Spark's own scale down we should terminate the executor once finished. Furthermore, in graceful shutdown it makes sense to release resources we no longer need if we've been asked to shutdown by the cluster manager instead of always holding the resources as long as possible. The decommissioned executors will exit and the end of decommissioning. This is sort of a user facing change, however decommissioning hasn't been in any releases yet. I changed the unit test to not send the executor exit message and still wait on the executor exited message. Closes apache#29211 from holdenk/SPARK-31197-exit-execs-redone. Authored-by: Holden Karau <hkarau@apple.com> Signed-off-by: Holden Karau <hkarau@apple.com> Connect decommissioning to dynamic scaling Because the mock always says there is an RDD we may replicate more than once, and now that there are independent threads Make Spark's dynamic allocation use decommissioning Track the decommissioning executors in the core dynamic scheduler so we don't scale down too low, update the streaming ExecutorAllocationManager to also delegate to decommission Fix up executor add for resource profile Fix our exiting and cleanup thread for better debugging next time. Cleanup the locks we use in decommissioning and clarify some more bits. Verify executors decommissioned, then killed by external external cluster manager are re-launched Verify some additional calls are not occuring in the executor allocation manager suite. Dont' close the watcher until the end of the test Use decommissionExecutors and set adjustTargetNumExecutors to false so that we can match the pattern for killExecutor/killExecutors bump numparts up to 6 Revert "bump numparts up to 6" This reverts commit daf96dd. Small coment & visibility cleanup CR feedback/cleanup Cleanup the merge [SPARK-21040][CORE] Speculate tasks which are running on decommission executors This PR adds functionality to consider the running tasks on decommission executors based on some config. In spark-on-cloud , we sometimes already know that an executor won't be alive for more than fix amount of time. Ex- In AWS Spot nodes, once we get the notification, we know that a node will be gone in 120 seconds. So if the running tasks on the decommissioning executors may run beyond currentTime+120 seconds, then they are candidate for speculation. Currently when an executor is decommission, we stop scheduling new tasks on those executors but the already running tasks keeps on running on them. Based on the cloud, we might know beforehand that an executor won't be alive for more than a preconfigured time. Different cloud providers gives different timeouts before they take away the nodes. For Ex- In case of AWS spot nodes, an executor won't be alive for more than 120 seconds. We can utilize this information in cloud environments and take better decisions about speculating the already running tasks on decommission executors. Yes. This PR adds a new config "spark.executor.decommission.killInterval" which they can explicitly set based on the cloud environment where they are running. Added UT. Closes apache#28619 from prakharjain09/SPARK-21040-speculate-decommission-exec-tasks. Authored-by: Prakhar Jain <prakharjain09@gmail.com> Signed-off-by: Holden Karau <hkarau@apple.com> [SPARK-20629][CORE][K8S] Copy shuffle data when nodes are being shutdown This pull request adds the ability to migrate shuffle files during Spark's decommissioning. The design document associated with this change is at https://docs.google.com/document/d/1xVO1b6KAwdUhjEJBolVPl9C6sLj7oOveErwDSYdT-pE . To allow this change the `MapOutputTracker` has been extended to allow the location of shuffle files to be updated with `updateMapOutput`. When a shuffle block is put, a block update message will be sent which triggers the `updateMapOutput`. Instead of rejecting remote puts of shuffle blocks `BlockManager` delegates the storage of shuffle blocks to it's shufflemanager's resolver (if supported). A new, experimental, trait is added for shuffle resolvers to indicate they handle remote putting of blocks. The existing block migration code is moved out into a separate file, and a producer/consumer model is introduced for migrating shuffle files from the host as quickly as possible while not overwhelming other executors. Recomputting shuffle blocks can be expensive, we should take advantage of our decommissioning time to migrate these blocks. This PR introduces two new configs parameters, `spark.storage.decommission.shuffleBlocks.enabled` & `spark.storage.decommission.rddBlocks.enabled` that control which blocks should be migrated during storage decommissioning. New unit test & expansion of the Spark on K8s decom test to assert that decommisioning with shuffle block migration means that the results are not recomputed even when the original executor is terminated. This PR is a cleaned-up version of the previous WIP PR I made apache#28331 (thanks to attilapiros for his very helpful reviewing on it :)). Closes apache#28708 from holdenk/SPARK-20629-copy-shuffle-data-when-nodes-are-being-shutdown-cleaned-up. Lead-authored-by: Holden Karau <hkarau@apple.com> Co-authored-by: Holden Karau <holden@pigscanfly.ca> Co-authored-by: “attilapiros” <piros.attila.zsolt@gmail.com> Co-authored-by: Attila Zsolt Piros <attilazsoltpiros@apiros-mbp16.lan> Signed-off-by: Holden Karau <hkarau@apple.com> [SPARK-24266][K8S] Restart the watcher when we receive a version changed from k8s Restart the watcher when it failed with a HTTP_GONE code from the kubernetes api. Which means a resource version has changed. For more relevant information see here: fabric8io/kubernetes-client#1075 No Running spark-submit to a k8s cluster. Not sure how to make an automated test for this. If someone can help me out that would be great. Closes apache#28423 from stijndehaes/bugfix/k8s-submit-resource-version-change. Authored-by: Stijn De Haes <stijndehaes@gmail.com> Signed-off-by: Holden Karau <hkarau@apple.com> [SPARK-32217] Plumb whether a worker would also be decommissioned along with executor This PR is a giant plumbing PR that plumbs an `ExecutorDecommissionInfo` along with the DecommissionExecutor message. The primary motivation is to know whether a decommissioned executor would also be loosing shuffle files -- and thus it is important to know whether the host would also be decommissioned. In the absence of this PR, the existing code assumes that decommissioning an executor does not loose the whole host with it, and thus does not clear the shuffle state if external shuffle service is enabled. While this may hold in some cases (like K8s decommissioning an executor pod, or YARN container preemption), it does not hold in others like when the cluster is managed by a Standalone Scheduler (Master). This is similar to the existing `workerLost` field in the `ExecutorProcessLost` message. In the future, this `ExecutorDecommissionInfo` can be embellished for knowing how long the executor has to live for scenarios like Cloud spot kills (or Yarn preemption) and the like. No Tweaked an existing unit test in `AppClientSuite` Closes apache#29032 from agrawaldevesh/plumb_decom_info. Authored-by: Devesh Agrawal <devesh.agrawal@gmail.com> Signed-off-by: Holden Karau <hkarau@apple.com> [SPARK-32199][SPARK-32198] Reduce job failures during decommissioning This PR reduces the prospect of a job loss during decommissioning. It fixes two holes in the current decommissioning framework: - (a) Loss of decommissioned executors is not treated as a job failure: We know that the decommissioned executor would be dying soon, so its death is clearly not caused by the application. - (b) Shuffle files on the decommissioned host are cleared when the first fetch failure is detected from a decommissioned host: This is a bit tricky in terms of when to clear the shuffle state ? Ideally you want to clear it the millisecond before the shuffle service on the node dies (or the executor dies when there is no external shuffle service) -- too soon and it could lead to some wastage and too late would lead to fetch failures. The approach here is to do this clearing when the very first fetch failure is observed on the decommissioned block manager, without waiting for other blocks to also signal a failure. Without them decommissioning a lot of executors at a time leads to job failures. The task scheduler tracks the executors that were decommissioned along with their `ExecutorDecommissionInfo`. This information is used by: (a) For handling a `ExecutorProcessLost` error, or (b) by the `DAGScheduler` when handling a fetch failure. No Added a new unit test `DecommissionWorkerSuite` to test the new behavior by exercising the Master-Worker decommissioning. I chose to add a new test since the setup logic was quite different from the existing `WorkerDecommissionSuite`. I am open to changing the name of the newly added test suite :-) - Should I add a feature flag to guard these two behaviors ? They seem safe to me that they should only get triggered by decommissioning, but you never know :-). Closes apache#29014 from agrawaldevesh/decom_harden. Authored-by: Devesh Agrawal <devesh.agrawal@gmail.com> Signed-off-by: Holden Karau <hkarau@apple.com> [SPARK-32417] Fix flakyness of BlockManagerDecommissionIntegrationSuite This test tries to fix the flakyness of BlockManagerDecommissionIntegrationSuite. Make the block manager decommissioning test be less flaky An interesting failure happens when migrateDuring = true (and persist or shuffle is true): - We schedule the job with tasks on executors 0, 1, 2. - We wait 300 ms and decommission executor 0. - If the task is not yet done on executor 0, it will now fail because the block manager won't be able to save the block. This condition is easy to trigger on a loaded machine where the github checks run. - The task with retry on a different executor (1 or 2) and its shuffle blocks will land there. - No actual block migration happens here because the decommissioned executor technically failed before it could even produce a block. To remove the above race, this change replaces the fixed wait for 300 ms to wait for an actual task to succeed. When a task has succeeded, we know its blocks would have been written for sure and thus its executor would certainly be forced to migrate those blocks when it is decommissioned. The change always decommissions an executor on which a real task finished successfully instead of picking the first executor. Because the system may choose to schedule nothing on the first executor and instead run the two tasks on one executor. I have had bad luck with BlockManagerDecommissionIntegrationSuite and it has failed several times on my PRs. So fixing it. No, unit test only change. Github checks. Ran this test 100 times, 10 at a time in parallel in a script. Closes apache#29226 from agrawaldevesh/block-manager-decom-flaky. Authored-by: Devesh Agrawal <devesh.agrawal@gmail.com> Signed-off-by: Holden Karau <hkarau@apple.com> [SPARK-31197][CORE] Shutdown executor once we are done decommissioning Exit the executor when it has been asked to decommission and there is nothing left for it to do. This is a rebase of apache#28817 If we want to use decommissioning in Spark's own scale down we should terminate the executor once finished. Furthermore, in graceful shutdown it makes sense to release resources we no longer need if we've been asked to shutdown by the cluster manager instead of always holding the resources as long as possible. The decommissioned executors will exit and the end of decommissioning. This is sort of a user facing change, however decommissioning hasn't been in any releases yet. I changed the unit test to not send the executor exit message and still wait on the executor exited message. Closes apache#29211 from holdenk/SPARK-31197-exit-execs-redone. Authored-by: Holden Karau <hkarau@apple.com> Signed-off-by: Holden Karau <hkarau@apple.com> Connect decommissioning to dynamic scaling Because the mock always says there is an RDD we may replicate more than once, and now that there are independent threads Make Spark's dynamic allocation use decommissioning Track the decommissioning executors in the core dynamic scheduler so we don't scale down too low, update the streaming ExecutorAllocationManager to also delegate to decommission Fix up executor add for resource profile Fix our exiting and cleanup thread for better debugging next time. Cleanup the locks we use in decommissioning and clarify some more bits. Verify executors decommissioned, then killed by external external cluster manager are re-launched Verify some additional calls are not occuring in the executor allocation manager suite. Dont' close the watcher until the end of the test Use decommissionExecutors and set adjustTargetNumExecutors to false so that we can match the pattern for killExecutor/killExecutors bump numparts up to 6 Revert "bump numparts up to 6" This reverts commit daf96dd. Small coment & visibility cleanup CR feedback/cleanup Fix up the merge CR feedback, move adjustExecutors to a common utility function Exclude some non-public APIs Remove junk More CR feedback Fix adjustExecutors backport This test fails for me locally and from what I recall it's because we use a different method of resolving the bind address than upstream so disabling the test This test fails for me locally and from what I recall it's because we use a different method of resolving the bind address than upstream so disabling the test Cleanup and drop watcher changes from the backport
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
What changes were proposed in this pull request?
This PR reduces the prospect of a job loss during decommissioning. It
fixes two holes in the current decommissioning framework:
(a) Loss of decommissioned executors is not treated as a job failure:
We know that the decommissioned executor would be dying soon, so its death is
clearly not caused by the application.
(b) Shuffle files on the decommissioned host are cleared when the
first fetch failure is detected from a decommissioned host: This is a
bit tricky in terms of when to clear the shuffle state ? Ideally you
want to clear it the millisecond before the shuffle service on the node
dies (or the executor dies when there is no external shuffle service) --
too soon and it could lead to some wastage and too late would lead to
fetch failures.
The approach here is to do this clearing when the very first fetch
failure is observed on the decommissioned block manager, without waiting for
other blocks to also signal a failure.
Why are the changes needed?
Without them decommissioning a lot of executors at a time leads to job failures.
Code overview
The task scheduler tracks the executors that were decommissioned along with their
ExecutorDecommissionInfo. This information is used by: (a) For handling aExecutorProcessLosterror, or (b) by theDAGSchedulerwhen handling a fetch failure.Does this PR introduce any user-facing change?
No
How was this patch tested?
Added a new unit test
DecommissionWorkerSuiteto test the new behavior by exercising the Master-Worker decommissioning. I chose to add a new test since the setup logic was quite different from the existingWorkerDecommissionSuite. I am open to changing the name of the newly added test suite :-)Questions for reviewers