[SPARK-23243][CORE][2.3] Fix RDD.repartition() data correctness issue #22354
Conversation
This file contains 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
An alternative fix for apache#21698 When Spark rerun tasks for an RDD, there are 3 different behaviors: 1. determinate. Always return the same result with same order when rerun. 2. unordered. Returns same data set in random order when rerun. 3. indeterminate. Returns different result when rerun. Normally Spark doesn't need to care about it. Spark runs stages one by one, when a task is failed, just rerun it. Although the rerun task may return a different result, users will not be surprised. However, Spark may rerun a finished stage when seeing fetch failures. When this happens, Spark needs to rerun all the tasks of all the succeeding stages if the RDD output is indeterminate, because the input of the succeeding stages has been changed. If the RDD output is determinate, we only need to rerun the failed tasks of the succeeding stages, because the input doesn't change. If the RDD output is unordered, it's same as determinate, because shuffle partitioner is always deterministic(round-robin partitioner is not a shuffle partitioner that extends `org.apache.spark.Partitioner`), so the reducers will still get the same input data set. This PR fixed the failure handling for `repartition`, to avoid correctness issues. For `repartition`, it applies a stateful map function to generate a round-robin id, which is order sensitive and makes the RDD's output indeterminate. When the stage contains `repartition` reruns, we must also rerun all the tasks of all the succeeding stages. **future improvement:** 1. Currently we can't rollback and rerun a shuffle map stage, and just fail. We should fix it later. https://issues.apache.org/jira/browse/SPARK-25341 2. Currently we can't rollback and rerun a result stage, and just fail. We should fix it later. https://issues.apache.org/jira/browse/SPARK-25342 3. We should provide public API to allow users to tag the random level of the RDD's computing function. a new test case Closes apache#22112 from cloud-fan/repartition. Lead-authored-by: Wenchen Fan <wenchen@databricks.com> Co-authored-by: Xingbo Jiang <xingbo.jiang@databricks.com> Signed-off-by: Xiao Li <gatorsmile@gmail.com>
|
+1 thanks @cloud-fan |
|
Test build #95764 has finished for PR 22354 at commit
|
|
thats weird, swear it had already passed, test failure does not look related |
|
test this please |
|
Test build #95770 has finished for PR 22354 at commit
|
|
thanks, merging to 2.3! |
asfgit
pushed a commit
that referenced
this pull request
Sep 7, 2018
backport #22112 to 2.3 ------- An alternative fix for #21698 When Spark rerun tasks for an RDD, there are 3 different behaviors: 1. determinate. Always return the same result with same order when rerun. 2. unordered. Returns same data set in random order when rerun. 3. indeterminate. Returns different result when rerun. Normally Spark doesn't need to care about it. Spark runs stages one by one, when a task is failed, just rerun it. Although the rerun task may return a different result, users will not be surprised. However, Spark may rerun a finished stage when seeing fetch failures. When this happens, Spark needs to rerun all the tasks of all the succeeding stages if the RDD output is indeterminate, because the input of the succeeding stages has been changed. If the RDD output is determinate, we only need to rerun the failed tasks of the succeeding stages, because the input doesn't change. If the RDD output is unordered, it's same as determinate, because shuffle partitioner is always deterministic(round-robin partitioner is not a shuffle partitioner that extends `org.apache.spark.Partitioner`), so the reducers will still get the same input data set. This PR fixed the failure handling for `repartition`, to avoid correctness issues. For `repartition`, it applies a stateful map function to generate a round-robin id, which is order sensitive and makes the RDD's output indeterminate. When the stage contains `repartition` reruns, we must also rerun all the tasks of all the succeeding stages. **future improvement:** 1. Currently we can't rollback and rerun a shuffle map stage, and just fail. We should fix it later. https://issues.apache.org/jira/browse/SPARK-25341 2. Currently we can't rollback and rerun a result stage, and just fail. We should fix it later. https://issues.apache.org/jira/browse/SPARK-25342 3. We should provide public API to allow users to tag the random level of the RDD's computing function. a new test case Closes #22354 from cloud-fan/repartition. Authored-by: Wenchen Fan <wenchen@databricks.com> Signed-off-by: Wenchen Fan <wenchen@databricks.com>
|
cc @jerryshao |
bersprockets
pushed a commit
to bersprockets/spark
that referenced
this pull request
Sep 7, 2018
backport apache#22112 to 2.2 ------- An alternative fix for apache#21698 When Spark rerun tasks for an RDD, there are 3 different behaviors: 1. determinate. Always return the same result with same order when rerun. 2. unordered. Returns same data set in random order when rerun. 3. indeterminate. Returns different result when rerun. Normally Spark doesn't need to care about it. Spark runs stages one by one, when a task is failed, just rerun it. Although the rerun task may return a different result, users will not be surprised. However, Spark may rerun a finished stage when seeing fetch failures. When this happens, Spark needs to rerun all the tasks of all the succeeding stages if the RDD output is indeterminate, because the input of the succeeding stages has been changed. If the RDD output is determinate, we only need to rerun the failed tasks of the succeeding stages, because the input doesn't change. If the RDD output is unordered, it's same as determinate, because shuffle partitioner is always deterministic(round-robin partitioner is not a shuffle partitioner that extends `org.apache.spark.Partitioner`), so the reducers will still get the same input data set. This PR fixed the failure handling for `repartition`, to avoid correctness issues. For `repartition`, it applies a stateful map function to generate a round-robin id, which is order sensitive and makes the RDD's output indeterminate. When the stage contains `repartition` reruns, we must also rerun all the tasks of all the succeeding stages. **future improvement:** 1. Currently we can't rollback and rerun a shuffle map stage, and just fail. We should fix it later. https://issues.apache.org/jira/browse/SPARK-25341 2. Currently we can't rollback and rerun a result stage, and just fail. We should fix it later. https://issues.apache.org/jira/browse/SPARK-25342 3. We should provide public API to allow users to tag the random level of the RDD's computing function. a new test case Closes apache#22354 from cloud-fan/repartition. Authored-by: Wenchen Fan <wenchen@databricks.com> Signed-off-by: Wenchen Fan <wenchen@databricks.com>
asfgit
pushed a commit
that referenced
this pull request
Sep 11, 2018
…ectness issue ## What changes were proposed in this pull request? Back port of #22354 and #17955 to 2.2 (#22354 depends on methods introduced by #17955). ------- An alternative fix for #21698 When Spark rerun tasks for an RDD, there are 3 different behaviors: 1. determinate. Always return the same result with same order when rerun. 2. unordered. Returns same data set in random order when rerun. 3. indeterminate. Returns different result when rerun. Normally Spark doesn't need to care about it. Spark runs stages one by one, when a task is failed, just rerun it. Although the rerun task may return a different result, users will not be surprised. However, Spark may rerun a finished stage when seeing fetch failures. When this happens, Spark needs to rerun all the tasks of all the succeeding stages if the RDD output is indeterminate, because the input of the succeeding stages has been changed. If the RDD output is determinate, we only need to rerun the failed tasks of the succeeding stages, because the input doesn't change. If the RDD output is unordered, it's same as determinate, because shuffle partitioner is always deterministic(round-robin partitioner is not a shuffle partitioner that extends `org.apache.spark.Partitioner`), so the reducers will still get the same input data set. This PR fixed the failure handling for `repartition`, to avoid correctness issues. For `repartition`, it applies a stateful map function to generate a round-robin id, which is order sensitive and makes the RDD's output indeterminate. When the stage contains `repartition` reruns, we must also rerun all the tasks of all the succeeding stages. **future improvement:** 1. Currently we can't rollback and rerun a shuffle map stage, and just fail. We should fix it later. https://issues.apache.org/jira/browse/SPARK-25341 2. Currently we can't rollback and rerun a result stage, and just fail. We should fix it later. https://issues.apache.org/jira/browse/SPARK-25342 3. We should provide public API to allow users to tag the random level of the RDD's computing function. ## How was this patch tested? a new test case Closes #22382 from bersprockets/SPARK-23243-2.2. Lead-authored-by: Bruce Robbins <bersprockets@gmail.com> Co-authored-by: Josh Rosen <joshrosen@databricks.com> Co-authored-by: Wenchen Fan <wenchen@databricks.com> Signed-off-by: Wenchen Fan <wenchen@databricks.com>
Willymontaz
pushed a commit
to criteo-forks/spark
that referenced
this pull request
Sep 25, 2019
…ectness issue ## What changes were proposed in this pull request? Back port of apache#22354 and apache#17955 to 2.2 (apache#22354 depends on methods introduced by apache#17955). ------- An alternative fix for apache#21698 When Spark rerun tasks for an RDD, there are 3 different behaviors: 1. determinate. Always return the same result with same order when rerun. 2. unordered. Returns same data set in random order when rerun. 3. indeterminate. Returns different result when rerun. Normally Spark doesn't need to care about it. Spark runs stages one by one, when a task is failed, just rerun it. Although the rerun task may return a different result, users will not be surprised. However, Spark may rerun a finished stage when seeing fetch failures. When this happens, Spark needs to rerun all the tasks of all the succeeding stages if the RDD output is indeterminate, because the input of the succeeding stages has been changed. If the RDD output is determinate, we only need to rerun the failed tasks of the succeeding stages, because the input doesn't change. If the RDD output is unordered, it's same as determinate, because shuffle partitioner is always deterministic(round-robin partitioner is not a shuffle partitioner that extends `org.apache.spark.Partitioner`), so the reducers will still get the same input data set. This PR fixed the failure handling for `repartition`, to avoid correctness issues. For `repartition`, it applies a stateful map function to generate a round-robin id, which is order sensitive and makes the RDD's output indeterminate. When the stage contains `repartition` reruns, we must also rerun all the tasks of all the succeeding stages. **future improvement:** 1. Currently we can't rollback and rerun a shuffle map stage, and just fail. We should fix it later. https://issues.apache.org/jira/browse/SPARK-25341 2. Currently we can't rollback and rerun a result stage, and just fail. We should fix it later. https://issues.apache.org/jira/browse/SPARK-25342 3. We should provide public API to allow users to tag the random level of the RDD's computing function. ## How was this patch tested? a new test case Closes apache#22382 from bersprockets/SPARK-23243-2.2. Lead-authored-by: Bruce Robbins <bersprockets@gmail.com> Co-authored-by: Josh Rosen <joshrosen@databricks.com> Co-authored-by: Wenchen Fan <wenchen@databricks.com> Signed-off-by: Wenchen Fan <wenchen@databricks.com>
Willymontaz
pushed a commit
to criteo-forks/spark
that referenced
this pull request
Sep 26, 2019
…ectness issue ## What changes were proposed in this pull request? Back port of apache#22354 and apache#17955 to 2.2 (apache#22354 depends on methods introduced by apache#17955). ------- An alternative fix for apache#21698 When Spark rerun tasks for an RDD, there are 3 different behaviors: 1. determinate. Always return the same result with same order when rerun. 2. unordered. Returns same data set in random order when rerun. 3. indeterminate. Returns different result when rerun. Normally Spark doesn't need to care about it. Spark runs stages one by one, when a task is failed, just rerun it. Although the rerun task may return a different result, users will not be surprised. However, Spark may rerun a finished stage when seeing fetch failures. When this happens, Spark needs to rerun all the tasks of all the succeeding stages if the RDD output is indeterminate, because the input of the succeeding stages has been changed. If the RDD output is determinate, we only need to rerun the failed tasks of the succeeding stages, because the input doesn't change. If the RDD output is unordered, it's same as determinate, because shuffle partitioner is always deterministic(round-robin partitioner is not a shuffle partitioner that extends `org.apache.spark.Partitioner`), so the reducers will still get the same input data set. This PR fixed the failure handling for `repartition`, to avoid correctness issues. For `repartition`, it applies a stateful map function to generate a round-robin id, which is order sensitive and makes the RDD's output indeterminate. When the stage contains `repartition` reruns, we must also rerun all the tasks of all the succeeding stages. **future improvement:** 1. Currently we can't rollback and rerun a shuffle map stage, and just fail. We should fix it later. https://issues.apache.org/jira/browse/SPARK-25341 2. Currently we can't rollback and rerun a result stage, and just fail. We should fix it later. https://issues.apache.org/jira/browse/SPARK-25342 3. We should provide public API to allow users to tag the random level of the RDD's computing function. ## How was this patch tested? a new test case Closes apache#22382 from bersprockets/SPARK-23243-2.2. Lead-authored-by: Bruce Robbins <bersprockets@gmail.com> Co-authored-by: Josh Rosen <joshrosen@databricks.com> Co-authored-by: Wenchen Fan <wenchen@databricks.com> Signed-off-by: Wenchen Fan <wenchen@databricks.com>
Willymontaz
pushed a commit
to criteo-forks/spark
that referenced
this pull request
Sep 27, 2019
…ectness issue ## What changes were proposed in this pull request? Back port of apache#22354 and apache#17955 to 2.2 (apache#22354 depends on methods introduced by apache#17955). ------- An alternative fix for apache#21698 When Spark rerun tasks for an RDD, there are 3 different behaviors: 1. determinate. Always return the same result with same order when rerun. 2. unordered. Returns same data set in random order when rerun. 3. indeterminate. Returns different result when rerun. Normally Spark doesn't need to care about it. Spark runs stages one by one, when a task is failed, just rerun it. Although the rerun task may return a different result, users will not be surprised. However, Spark may rerun a finished stage when seeing fetch failures. When this happens, Spark needs to rerun all the tasks of all the succeeding stages if the RDD output is indeterminate, because the input of the succeeding stages has been changed. If the RDD output is determinate, we only need to rerun the failed tasks of the succeeding stages, because the input doesn't change. If the RDD output is unordered, it's same as determinate, because shuffle partitioner is always deterministic(round-robin partitioner is not a shuffle partitioner that extends `org.apache.spark.Partitioner`), so the reducers will still get the same input data set. This PR fixed the failure handling for `repartition`, to avoid correctness issues. For `repartition`, it applies a stateful map function to generate a round-robin id, which is order sensitive and makes the RDD's output indeterminate. When the stage contains `repartition` reruns, we must also rerun all the tasks of all the succeeding stages. **future improvement:** 1. Currently we can't rollback and rerun a shuffle map stage, and just fail. We should fix it later. https://issues.apache.org/jira/browse/SPARK-25341 2. Currently we can't rollback and rerun a result stage, and just fail. We should fix it later. https://issues.apache.org/jira/browse/SPARK-25342 3. We should provide public API to allow users to tag the random level of the RDD's computing function. ## How was this patch tested? a new test case Closes apache#22382 from bersprockets/SPARK-23243-2.2. Lead-authored-by: Bruce Robbins <bersprockets@gmail.com> Co-authored-by: Josh Rosen <joshrosen@databricks.com> Co-authored-by: Wenchen Fan <wenchen@databricks.com> Signed-off-by: Wenchen Fan <wenchen@databricks.com>
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.
backport #22112 to 2.3
An alternative fix for #21698
When Spark rerun tasks for an RDD, there are 3 different behaviors:
Normally Spark doesn't need to care about it. Spark runs stages one by one, when a task is failed, just rerun it. Although the rerun task may return a different result, users will not be surprised.
However, Spark may rerun a finished stage when seeing fetch failures. When this happens, Spark needs to rerun all the tasks of all the succeeding stages if the RDD output is indeterminate, because the input of the succeeding stages has been changed.
If the RDD output is determinate, we only need to rerun the failed tasks of the succeeding stages, because the input doesn't change.
If the RDD output is unordered, it's same as determinate, because shuffle partitioner is always deterministic(round-robin partitioner is not a shuffle partitioner that extends
org.apache.spark.Partitioner), so the reducers will still get the same input data set.This PR fixed the failure handling for
repartition, to avoid correctness issues.For
repartition, it applies a stateful map function to generate a round-robin id, which is order sensitive and makes the RDD's output indeterminate. When the stage containsrepartitionreruns, we must also rerun all the tasks of all the succeeding stages.future improvement:
a new test case