### What changes were proposed in this pull request?
Support set off heap memory in `ExecutorResourceRequests`
### Why are the changes needed?
Support stage level scheduling
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Added UT in `ResourceProfileSuite` and `DAGSchedulerSuite`
Closes#28972 from warrenzhu25/30794.
Authored-by: Warren Zhu <zhonzh@microsoft.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
Updates to scalatest 3.2.0. Though it looks large, it is 99% changes to the new location of scalatest classes.
### Why are the changes needed?
3.2.0+ has a fix that is required for Scala 2.13.3+ compatibility.
### Does this PR introduce _any_ user-facing change?
No, only affects tests.
### How was this patch tested?
Existing tests.
Closes#29196 from srowen/SPARK-32398.
Authored-by: Sean Owen <srowen@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
See again the related PRs like https://github.com/apache/spark/pull/28971
This completes fixing compilation for 2.13 for all but `repl`, which is a separate task.
### Why are the changes needed?
Eventually, we need to support a Scala 2.13 build, perhaps in Spark 3.1.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Existing tests. (2.13 was not tested; this is about getting it to compile without breaking 2.12)
Closes#29147 from srowen/SPARK-29292.4.
Authored-by: Sean Owen <srowen@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR will remove references to these "blacklist" and "whitelist" terms besides the blacklisting feature as a whole, which can be handled in a separate JIRA/PR.
This touches quite a few files, but the changes are straightforward (variable/method/etc. name changes) and most quite self-contained.
### Why are the changes needed?
As per discussion on the Spark dev list, it will be beneficial to remove references to problematic language that can alienate potential community members. One such reference is "blacklist" and "whitelist". While it seems to me that there is some valid debate as to whether these terms have racist origins, the cultural connotations are inescapable in today's world.
### Does this PR introduce _any_ user-facing change?
In the test file `HiveQueryFileTest`, a developer has the ability to specify the system property `spark.hive.whitelist` to specify a list of Hive query files that should be tested. This system property has been renamed to `spark.hive.includelist`. The old property has been kept for compatibility, but will log a warning if used. I am open to feedback from others on whether keeping a deprecated property here is unnecessary given that this is just for developers running tests.
### How was this patch tested?
Existing tests should be suitable since no behavior changes are expected as a result of this PR.
Closes#28874 from xkrogen/xkrogen-SPARK-32036-rename-blacklists.
Authored-by: Erik Krogen <ekrogen@linkedin.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
This change replaces the world slave with alternatives matching the context.
### Why are the changes needed?
There is no need to call things slave, we might as well use better clearer names.
### Does this PR introduce _any_ user-facing change?
Yes, the ouput JSON does change. To allow backwards compatibility this is an additive change.
The shell scripts for starting & stopping workers are renamed, and for backwards compatibility old scripts are added to call through to the new ones while printing a deprecation message to stderr.
### How was this patch tested?
Existing tests.
Closes#28864 from holdenk/SPARK-32004-drop-references-to-slave.
Lead-authored-by: Holden Karau <hkarau@apple.com>
Co-authored-by: Holden Karau <holden@pigscanfly.ca>
Signed-off-by: Holden Karau <hkarau@apple.com>
### What changes were proposed in this pull request?
In IPv6 scenario, the current logic to split hostname and port is not correct.
### Why are the changes needed?
to support IPV6 deployment scenario
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
UT and IPV6 spark deployment with yarn
Closes#28931 from PavithraRamachandran/ipv6_issue.
Authored-by: Pavithraramachandran <pavi.rams@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
## What changes were proposed in this pull request?
When a Spark Job launched in Cluster mode with Yarn, Application Master sets spark.ui.port port to 0 which means Driver's web UI gets any random port even if we want to explicitly set the Port range for Driver's Web UI
## Why are the changes needed?
We access Spark Web UI via Knox Proxy, and there are firewall restrictions due to which we can not access Spark Web UI since Web UI port range gets random port even if we set explicitly.
This Change will check if there is a specified port range explicitly mentioned so that it does not assign a random port.
## Does this PR introduce any user-facing change?
No
## How was this patch tested?
Local Tested.
Closes#28880 from rajatahujaatinmobi/ahujarajat261/SPARK-32039-change-yarn-webui-port-range-with-property-latest-spark.
Authored-by: Rajat Ahuja <rahuja@twitter.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
#31029 # What changes were proposed in this pull request?
In YarnSchedulerBackend, we should avoid using the global execution context for its Future. Otherwise if user's Spark application also uses global execution context for its Future, the user is facing indeterministic behavior in terms of the thread's context class loader.
### Why are the changes needed?
When running tpc-ds test (https://github.com/databricks/spark-sql-perf), occasionally we see error related to class not found:
2020-02-04 20:00:26,673 ERROR yarn.ApplicationMaster: User class threw exception: scala.ScalaReflectionException: class com.databricks.spark.sql.perf.ExperimentRun in JavaMirror with
sun.misc.Launcher$AppClassLoader28ba21f3 of type class sun.misc.Launcher$AppClassLoader with classpath [...]
and parent being sun.misc.Launcher$ExtClassLoader3ff5d147 of type class sun.misc.Launcher$ExtClassLoader with classpath [...]
and parent being primordial classloader with boot classpath [...] not found.
This is the root cause for the problem:
Spark driver starts ApplicationMaster in the main thread, which starts a user thread and set MutableURLClassLoader to that thread's ContextClassLoader.
userClassThread = startUserApplication()
The main thread then setup YarnSchedulerBackend RPC endpoints, which handles these calls using scala Future with the default global ExecutionContext:
doRequestTotalExecutors
doKillExecutors
So for the main thread and user thread, whoever starts the future first get a chance to set ContextClassLoader to the default thread pool:
- If main thread starts a future to handle doKillExecutors() before user thread does then the default thread pool thread's ContextClassLoader would be the default (AppClassLoader).
- If user thread starts a future first then the thread pool thread will have MutableURLClassLoader.
Note that only MutableURLClassLoader can load user provided class for the Spark app, you will see errors related to class not found if the ContextClassLoader is AppClassLoader.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Existing unit tests and manual tests
Closes#27843 from shanyu/shanyu-31029.
Authored-by: Shanyu Zhao <shzhao@microsoft.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
If a Spark distribution has built-in hadoop runtime, Spark will not populate the hadoop classpath from `yarn.application.classpath` and `mapreduce.application.classpath` when a job is submitted to Yarn. Users can override this behavior by setting `spark.yarn.populateHadoopClasspath` to `true`.
### Why are the changes needed?
Without this, Spark will populate the hadoop classpath from `yarn.application.classpath` and `mapreduce.application.classpath` even Spark distribution has built-in hadoop. This results jar conflict and many unexpected behaviors in runtime.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Manually test with two builds, with-hadoop and no-hadoop builds.
Closes#28788 from dbtsai/yarn-classpath.
Authored-by: DB Tsai <d_tsai@apple.com>
Signed-off-by: DB Tsai <d_tsai@apple.com>
### What changes were proposed in this pull request?
Use spark-submit to submit a pyspark app on Yarn, and set this in spark-env.sh:
export PYSPARK_ARCHIVES_PATH=local:/opt/spark/python/lib/pyspark.zip,local:/opt/spark/python/lib/py4j-0.10.7-src.zip
You can see that these local archives are still uploaded to Yarn distributed cache:
yarn.Client: Uploading resource file:/opt/spark/python/lib/pyspark.zip -> hdfs://myhdfs/user/test1/.sparkStaging/application_1581024490249_0001/pyspark.zip
This PR fix this issue by checking the files specified in PYSPARK_ARCHIVES_PATH, if they are local archives, don't distribute to Yarn dist cache.
### Why are the changes needed?
For pyspark appp to support local pyspark archives set in PYSPARK_ARCHIVES_PATH.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Existing tests and manual tests.
Closes#27598 from shanyu/shanyu-30845.
Authored-by: Shanyu Zhao <shzhao@microsoft.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
Improve the check logic on if all node managers are really being backlisted.
### Why are the changes needed?
I observed when the AM is out of sync with ResourceManager, or RM is having issue report back with current number of available NMs, something like below happens:
...
20/05/13 09:01:21 INFO RetryInvocationHandler: java.io.EOFException: End of File Exception between local host is: "client.zyx.com/x.x.x.124"; destination host is: "rm.zyx.com":8030; : java.io.EOFException; For more details see: http://wiki.apache.org/hadoop/EOFException, while invoking ApplicationMasterProtocolPBClientImpl.allocate over rm543. Trying to failover immediately.
...
20/05/13 09:01:28 WARN AMRMClientImpl: ApplicationMaster is out of sync with ResourceManager, hence resyncing.
...
then the spark job would suddenly run into AllNodeBlacklisted state:
...
20/05/13 09:01:31 INFO ApplicationMaster: Final app status: FAILED, exitCode: 11, (reason: Due to executor failures all available nodes are blacklisted)
...
but actually there's no black listed nodes in currentBlacklistedYarnNodes, and I do not see any blacklisting message from:
https://github.com/apache/spark/blob/master/resource-managers/yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocatorBlacklistTracker.scala#L119
We should only return isAllNodeBlacklisted =true when we see there are >0 numClusterNodes AND 'currentBlacklistedYarnNodes.size >= numClusterNodes'.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
A minor change. No changes on tests.
Closes#28606 from cnZach/false_AllNodeBlacklisted_when_RM_is_having_issue.
Authored-by: Yuexin Zhang <zach.yx.zhang@gmail.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
This PR proposes to show the actual traceback when "handle large number of containers and tasks (SPARK-18750)" test fails in `LocalityPlacementStrategySuite`.
**It does not fully resolve the JIRA SPARK-31746 yet**. I tried to reproduce in my local by controlling the factors in the tests but I couldn't. I double checked the changes in SPARK-18750 are still valid.
### Why are the changes needed?
This test is flaky for an unknown reason (see https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/122768/testReport/org.apache.spark.deploy.yarn/LocalityPlacementStrategySuite/handle_large_number_of_containers_and_tasks__SPARK_18750_/):
```
sbt.ForkMain$ForkError: org.scalatest.exceptions.TestFailedException: java.lang.StackOverflowError did not equal null
at org.scalatest.Assertions.newAssertionFailedException(Assertions.scala:530)
at org.scalatest.Assertions.newAssertionFailedException$(Assertions.scala:529)
at org.scalatest.FunSuite.newAssertionFailedException(FunSuite.scala:1560)
at org.scalatest.Assertions$AssertionsHelper.macroAssert(Assertions.scala:503)
```
After this PR, it will help to investigate the root cause:
**Before**:
```
[info] - handle large number of containers and tasks (SPARK-18750) *** FAILED *** (824 milliseconds)
[info] java.lang.StackOverflowError did not equal null (LocalityPlacementStrategySuite.scala:49)
[info] org.scalatest.exceptions.TestFailedException:
[info] at org.scalatest.Assertions.newAssertionFailedException(Assertions.scala:530)
[info] at org.scalatest.Assertions.newAssertionFailedException$(Assertions.scala:529)
[info] at org.scalatest.FunSuite.newAssertionFailedException(FunSuite.scala:1560)
[info] at org.scalatest.Assertions$AssertionsHelper.macroAssert(Assertions.scala:503)
[info] at org.apache.spark.deploy.yarn.LocalityPlacementStrategySuite.$anonfun$new$1(LocalityPlacementStrategySuite.scala:49)
[info] at org.scalatest.OutcomeOf.outcomeOf(OutcomeOf.scala:85)
[info] at org.scalatest.OutcomeOf.outcomeOf$(OutcomeOf.scala:83)
[info] at org.scalatest.OutcomeOf$.outcomeOf(OutcomeOf.scala:104)
[info] at org.scalatest.Transformer.apply(Transformer.scala:22)
[info] at org.scalatest.Transformer.apply(Transformer.scala:20)
[info] at org.scalatest.FunSuiteLike$$anon$1.apply(FunSuiteLike.scala:186)
[info] at org.apache.spark.SparkFunSuite.withFixture(SparkFunSuite.scala:157)
[info] at org.scalatest.FunSuiteLike.invokeWithFixture$1(FunSuiteLike.scala:184)
[info] at org.scalatest.FunSuiteLike.$anonfun$runTest$1(FunSuiteLike.scala:196)
[info] at org.scalatest.SuperEngine.runTestImpl(Engine.scala:286)
[info] at org.scalatest.FunSuiteLike.runTest(FunSuiteLike.scala:196)
[info] at org.scalatest.FunSuiteLike.runTest$(FunSuiteLike.scala:178)
...
```
**After**:
```
[info] - handle large number of containers and tasks (SPARK-18750) *** FAILED *** (825 milliseconds)
[info] StackOverflowError should not be thrown; however, got:
[info]
[info] java.lang.StackOverflowError
[info] at scala.collection.TraversableLike.$anonfun$filterImpl$1(TraversableLike.scala:256)
[info] at scala.collection.MapLike$MappedValues.$anonfun$foreach$3(MapLike.scala:256)
[info] at scala.collection.TraversableLike$WithFilter.$anonfun$foreach$1(TraversableLike.scala:877)
[info] at scala.collection.MapLike$MappedValues.$anonfun$foreach$3(MapLike.scala:256)
[info] at scala.collection.TraversableLike$WithFilter.$anonfun$foreach$1(TraversableLike.scala:877)
[info] at scala.collection.MapLike$MappedValues.$anonfun$foreach$3(MapLike.scala:256)
[info] at scala.collection.TraversableLike$WithFilter.$anonfun$foreach$1(TraversableLike.scala:877)
[info] at scala.collection.MapLike$MappedValues.$anonfun$foreach$3(MapLike.scala:256)
[info] at scala.collection.TraversableLike$WithFilter.$anonfun$foreach$1(TraversableLike.scala:877)
[info] at scala.collection.MapLike$MappedValues.$anonfun$foreach$3(MapLike.scala:256)
[info] at scala.collection.TraversableLike$WithFilter.$anonfun$foreach$1(TraversableLike.scala:877)
[info] at scala.collection.MapLike$MappedValues.$anonfun$foreach$3(MapLike.scala:256)
[info] at scala.collection.TraversableLike$WithFilter.$anonfun$foreach$1(TraversableLike.scala:877)
[info] at scala.collection.MapLike$MappedValues.$anonfun$foreach$3(MapLike.scala:256)
[info] at scala.collection.TraversableLike$WithFilter.$anonfun$foreach$1(TraversableLike.scala:877)
[info] at scala.collection.MapLike$MappedValues.$anonfun$foreach$3(MapLike.scala:256)
...
```
### Does this PR introduce _any_ user-facing change?
No, dev-only.
### How was this patch tested?
Manually tested by reverting 76db394f2b locally.
Closes#28566 from HyukjinKwon/SPARK-31746.
Authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
This PR aims to recover Hadoop-3.2 profile jobs on `master` branch by disabling a UT added by SPARK-31235 in Hadoop 3.2 temporarily. The target UT is not a flaky test. It always fail on Hadoop-3.2 profile currently although it works in Hadoop 2.7 profile. So, in this PR, we keep the test coverage in Hadoop 2.7 and ignore the test in Hadoop 3.2 temporarily to unblock the other PRs.
### Why are the changes needed?
SPARK-31235 added a test case which is breaking Hadoop 3.2 and there are two follow-up to fix it. Although two follow-ups can fix the UT in Hadoop 3.2 environment. The side-effect on Hadoop classes cause some random UT failures in the other suites.
- https://github.com/apache/spark/pull/28456
- https://github.com/apache/spark/pull/28550
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Pass the Jenkins with SBT/Maven.
Closes#28552 from dongjoon-hyun/SPARK-31235-2.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This patch re-obtain tokens at the start of AM for yarn cluster mode, if principal and keytab are available. It basically transfers the credentials from the original user, so this patch puts the new tokens into credentials from the original user via overwriting.
To obtain tokens from providers in user application, this patch leverages the user classloader as context classloader while initializing token manager in the startup of AM.
### Why are the changes needed?
Submitter will obtain delegation tokens for yarn-cluster mode, and add these credentials to the launch context. AM will be launched with these credentials, and AM and driver are able to leverage these tokens.
In Yarn cluster mode, driver is launched in AM, which in turn initializes token manager (while initializing SparkContext) and obtain delegation tokens (+ schedule to renew) if both principal and keytab are available.
That said, even we provide principal and keytab to run application with yarn-cluster mode, AM always starts with initial tokens from launch context until token manager runs and obtains delegation tokens.
So there's a "gap", and if user codes (driver) access to external system with delegation tokens (e.g. HDFS) before initializing SparkContext, it cannot leverage the tokens token manager will obtain. It will make the application fail if AM is killed "after" the initial tokens are expired and relaunched.
This is even a regression: see below codes in branch-2.4:
https://github.com/apache/spark/blob/branch-2.4/resource-managers/yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scalahttps://github.com/apache/spark/blob/branch-2.4/resource-managers/yarn/src/main/scala/org/apache/spark/deploy/yarn/security/AMCredentialRenewer.scala
In Spark 2.4.x, AM runs AMCredentialRenewer at initialization, and AMCredentialRenew obtains tokens and merge with credentials being provided with launch context of AM. So it guarantees new tokens in driver run.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
Manually tested with specifically crafted application (simple reproducer) - https://github.com/HeartSaVioR/spark-delegation-token-experiment/blob/master/src/main/scala/net/heartsavior/spark/example/LongRunningAppWithHDFSConfig.scala
Before this patch, new AM attempt failed when I killed AM after the expiration of tokens. After this patch the new AM attempt runs fine.
Closes#28336 from HeartSaVioR/SPARK-31559.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan.opensource@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@apache.org>
### What changes were proposed in this pull request?
Register `NettyMemoryMetrics` into Node Manager's metrics system through `YarnShuffleServiceMetrics`.
- usedDirectMemory
- usedHeapMemory
### Why are the changes needed?
Such that `NettyMemoryMetrics` can be exposed through Node Manager's JMX.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Update UT to ensure NettyMemoryMetrics are registered into Node Manager's metrics system.
Closes#28416 from manuzhang/spark-31611.
Authored-by: tianlzhang <tianlzhang@ebay.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
Update the input parameters for instantiating `RMAppManager` and `ClientRMService`
### Why are the changes needed?
For hadoop3.2, if `RMAppManager` is not created correctly, the following exception will occur:
```
java.lang.RuntimeException: java.lang.NullPointerException
at org.apache.hadoop.util.ReflectionUtils.newInstance(ReflectionUtils.java:135)
at org.apache.hadoop.yarn.security.YarnAuthorizationProvider.getInstance(YarnAuthorizationProvider.java:55)
at org.apache.hadoop.yarn.server.resourcemanager.RMAppManager.<init>(RMAppManager.java:117)
```
### How was this patch tested?
UTs
Closes#28456 from wang-zhun/Fix-SPARK-31235.
Authored-by: wang-zhun <wangzhun6103@gmail.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
Remove unused `registeredConnections` counter from `ExternalBlockHandler#ShuffleMetrics`
This was added by SPARK-25642 at 3.0.0
- 8dd29fe36b
### Why are the changes needed?
It's `registeredConnections` counter created in `TransportContext` that's really counting the numbers and it's misleading for people who want to add new metrics like `registeredConnections`.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Add UTs to ensure all expected metrics are registered for `ExternalShuffleService` and `YarnShuffleService`
Closes#28457 from manuzhang/spark-31611-pre.
Lead-authored-by: tianlzhang <tianlzhang@ebay.com>
Co-authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR adds `spark.yarn.applicationType` to identify the application type
### Why are the changes needed?
The current application defaults to the SPARK type.
In fact, different types of applications have different characteristics and are suitable for different scenarios.For example: SPAKR-SQL, SPARK-STREAMING.
I recommend distinguishing them by the parameter `spark.yarn.applicationType` so that we can more easily manage and maintain different types of applications.
### How was this patch tested?
1.add UT
2.Tested by verifying Yarn-UI `ApplicationType` in the following cases:
- client and cluster mode
Need additional explanation:
limit cannot exceed 20 characters, can be empty or space
The reasons are as follows:
```
// org.apache.hadoop.yarn.server.resourcemanager.submitApplication.
if (submissionContext.getApplicationType() == null) {
submissionContext
.setApplicationType(YarnConfiguration.DEFAULT_APPLICATION_TYPE);
} else {
// APPLICATION_TYPE_LENGTH = 20
if (submissionContext.getApplicationType().length() > YarnConfiguration.APPLICATION_TYPE_LENGTH) {
submissionContext.setApplicationType(submissionContext
.getApplicationType().substring(0,
YarnConfiguration.APPLICATION_TYPE_LENGTH));
}
}
```
Closes#28009 from wang-zhun/SPARK-31235.
Authored-by: wang-zhun <wangzhun6103@gmail.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
We are adding a new Spark Yarn configuration, `spark.yarn.populateHadoopClasspath` to not populate Hadoop classpath from `yarn.application.classpath` and `mapreduce.application.classpath`.
### Why are the changes needed?
Spark Yarn client populates extra Hadoop classpath from `yarn.application.classpath` and `mapreduce.application.classpath` when a job is submitted to a Yarn Hadoop cluster.
However, for `with-hadoop` Spark build that embeds Hadoop runtime, it can cause jar conflicts because Spark distribution can contain different version of Hadoop jars.
One case we have is when a user uses an Apache Spark distribution with its-own embedded hadoop, and submits a job to Cloudera or Hortonworks Yarn clusters, because of two different incompatible Hadoop jars in the classpath, it runs into errors.
By not populating the Hadoop classpath from the clusters can address this issue.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
An UT is added, but very hard to add a new integration test since this requires using different incompatible versions of Hadoop.
We also manually tested this PR, and we are able to submit a Spark job using Spark distribution built with Apache Hadoop 2.10 to CDH 5.6 without populating CDH classpath.
Closes#28376 from dbtsai/yarn-classpath.
Authored-by: DB Tsai <d_tsai@apple.com>
Signed-off-by: DB Tsai <d_tsai@apple.com>
### What changes were proposed in this pull request?
This PR fixes a typo in deploy/yarn/LocalityPreferredContainerPlacementStrategy.scala file.
### Why are the changes needed?
To deliver correct explanation about how the placement policy works.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
UT as specified, although shouldn't influence any functionality since it's in the comment.
Closes#28267 from asclepiusaka/master.
Authored-by: Cong Du <asclepius1993@gmail.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
Yarn side changes for Stage level scheduling. The previous PR for dynamic allocation changes was https://github.com/apache/spark/pull/27313
Modified the data structures to store things on a per ResourceProfile basis.
I tried to keep the code changes to a minimum, the main loop that requests just goes through each Resourceprofile and the logic inside for each one stayed very close to the same.
On submission we now have to give each ResourceProfile a separate yarn Priority because yarn doesn't support asking for containers with different resources at the same Priority. We just use the profile id as the priority level.
Using a different Priority actually makes things easier when the containers come back to match them again which ResourceProfile they were requested for.
The expectation is that yarn will only give you a container with resource amounts you requested or more. It should never give you a container if it doesn't satisfy your resource requests.
If you want to see the full feature changes you can look at https://github.com/apache/spark/pull/27053/files for reference
### Why are the changes needed?
For stage level scheduling YARN support.
### Does this PR introduce any user-facing change?
no
### How was this patch tested?
Tested manually on YARN cluster and then unit tests.
Closes#27583 from tgravescs/SPARK-29149.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
This patch is to bump the master branch version to 3.1.0-SNAPSHOT.
### Why are the changes needed?
N/A
### Does this PR introduce any user-facing change?
N/A
### How was this patch tested?
N/A
Closes#27698 from gatorsmile/updateVersion.
Authored-by: gatorsmile <gatorsmile@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
This is another PR for stage level scheduling. In particular this adds changes to the dynamic allocation manager and the scheduler backend to be able to track what executors are needed per ResourceProfile. Note the api is still private to Spark until the entire feature gets in, so this functionality will be there but only usable by tests for profiles other then the DefaultProfile.
The main changes here are simply tracking things on a ResourceProfile basis as well as sending the executor requests to the scheduler backend for all ResourceProfiles.
I introduce a ResourceProfileManager in this PR that will track all the actual ResourceProfile objects so that we can keep them all in a single place and just pass around and use in datastructures the resource profile id. The resource profile id can be used with the ResourceProfileManager to get the actual ResourceProfile contents.
There are various places in the code that use executor "slots" for things. The ResourceProfile adds functionality to keep that calculation in it. This logic is more complex then it should due to standalone mode and mesos coarse grained not setting the executor cores config. They default to all cores on the worker, so calculating slots is harder there.
This PR keeps the functionality to make the cores the limiting resource because the scheduler still uses that for "slots" for a few things.
This PR does also add the resource profile id to the Stage and stage info classes to be able to test things easier. That full set of changes will come with the scheduler PR that will be after this one.
The PR stops at the scheduler backend pieces for the cluster manager and the real YARN support hasn't been added in this PR, that again will be in a separate PR, so this has a few of the API changes up to the cluster manager and then just uses the default profile requests to continue.
The code for the entire feature is here for reference: https://github.com/apache/spark/pull/27053/files although it needs to be upmerged again as well.
### Why are the changes needed?
Needed for stage level scheduling feature.
### Does this PR introduce any user-facing change?
No user facing api changes added here.
### How was this patch tested?
Lots of unit tests and manually testing. I tested on yarn, k8s, standalone, local modes. Ran both failure and success cases.
Closes#27313 from tgravescs/SPARK-29148.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
This change is to allow custom resource scheduler (GPUs,FPGAs,etc) resource discovery to be more flexible. Users are asking for it to work with hadoop 2.x versions that do not support resource scheduling in YARN and/or also they may not run in an isolated environment.
This change creates a plugin api that users can write their own resource discovery class that allows a lot more flexibility. The user can chain plugins for different resource types. The user specified plugins execute in the order specified and will fall back to use the discovery script plugin if they don't return information for a particular resource.
I had to open up a few of the classes to be public and change them to not be case classes and make them developer api in order for the the plugin to get enough information it needs.
I also relaxed the yarn side so that if yarn isn't configured for resource scheduling we just warn and go on. This helps users that have yarn 3.1 but haven't configured the resource scheduling side on their cluster yet, or aren't running in isolated environment.
The user would configured this like:
--conf spark.resources.discovery.plugin="org.apache.spark.resource.ResourceDiscoveryFPGAPlugin, org.apache.spark.resource.ResourceDiscoveryGPUPlugin"
Note the executor side had to be wrapped with a classloader to make sure we include the user classpath for jars they specified on submission.
Note this is more flexible because the discovery script has limitations such as spawning it in a separate process. This means if you are trying to allocate resources in that process they might be released when the script returns. Other things are the class makes it more flexible to be able to integrate with existing systems and solutions for assigning resources.
### Why are the changes needed?
to more easily use spark resource scheduling with older versions of hadoop or in non-isolated enivronments.
### Does this PR introduce any user-facing change?
Yes a plugin api
### How was this patch tested?
Unit tests added and manual testing done on yarn and standalone modes.
Closes#27410 from tgravescs/hadoop27spark3.
Lead-authored-by: Thomas Graves <tgraves@nvidia.com>
Co-authored-by: Thomas Graves <tgraves@apache.org>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
This is the second PR for the Stage Level Scheduling. This is adding in the necessary executor side changes:
1) executors to know what ResourceProfile they should be using
2) handle parsing the resource profile settings - these are not in the global configs
3) then reporting back to the driver what resource profile it was started with.
This PR adds all the piping for YARN to pass the information all the way to executors, but it just uses the default ResourceProfile (which is the global applicatino level configs).
At a high level these changes include:
1) adding a new --resourceProfileId option to the CoarseGrainedExecutorBackend
2) Add the ResourceProfile settings to new internal confs that gets passed into the Executor
3) Executor changes that use the resource profile id passed in to read the corresponding ResourceProfile confs and then parse those requests and discover resources as necessary
4) Executor registers to Driver with the Resource profile id so that the ExecutorMonitor can track how many executor with each profile are running
5) YARN side changes to show that passing the resource profile id and confs actually works. Just uses the DefaultResourceProfile for now.
I also removed a check from the CoarseGrainedExecutorBackend that used to check to make sure there were task requirements before parsing any custom resource executor requests. With the resource profiles this becomes much more expensive because we would then have to pass the task requests to each executor and the check was just a short cut and not really needed. It was much cleaner just to remove it.
Note there were some changes to the ResourceProfile, ExecutorResourceRequests, and TaskResourceRequests in this PR as well because I discovered some issues with things not being immutable. That api now look like:
val rpBuilder = new ResourceProfileBuilder()
val ereq = new ExecutorResourceRequests()
val treq = new TaskResourceRequests()
ereq.cores(2).memory("6g").memoryOverhead("2g").pysparkMemory("2g").resource("gpu", 2, "/home/tgraves/getGpus")
treq.cpus(2).resource("gpu", 2)
val resourceProfile = rpBuilder.require(ereq).require(treq).build
This makes is so that ResourceProfile is immutable and Spark can use it directly without worrying about the user changing it.
### Why are the changes needed?
These changes are needed for the executor to report which ResourceProfile they are using so that ultimately the dynamic allocation manager can use that information to know how many with a profile are running and how many more it needs to request. Its also needed to get the resource profile confs to the executor so that it can run the appropriate discovery script if needed.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Unit tests and manually on YARN.
Closes#26682 from tgravescs/SPARK-29306.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
Remove `executorsPendingToRemove.clear()` from `CoarseGrainedSchedulerBackend.reset()`.
### Why are the changes needed?
Clear `executorsPendingToRemove` before remove executors will cause all tasks running on those "pending to remove" executors to count failures. But that's not true for the case of `executorsPendingToRemove(execId)=true`.
Besides, `executorsPendingToRemove` will be cleaned up within `removeExecutor()` at the end just as same as `executorsPendingLossReason`.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Added a new test in `TaskSetManagerSuite`.
Closes#27017 from Ngone51/dont-clear-eptr-in-reset.
Authored-by: yi.wu <yi.wu@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
ShutdownHook of YarnClientSchedulerBackend prints just "Stopped" which can be improved to "YarnClientSchedulerBackend Stopped" for better understanding.
### Why are the changes needed?
While stopping or gracefully exiting the spark-shell/spark-sql --master yarn, only printing `stopped` is useless.
### Does this PR introduce any user-facing change?
Yes. Log info message change.
### How was this patch tested?
Manually
Closes#27049 from jobitmathew/imp_stop_message.
Authored-by: Jobit Mathew <jobit.mathew@huawei.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
1. Revert "Preparing development version 3.0.1-SNAPSHOT": 56dcd79
2. Revert "Preparing Spark release v3.0.0-preview2-rc2": c216ef1
### Why are the changes needed?
Shouldn't change master.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
manual test:
https://github.com/apache/spark/compare/5de5e46..wangyum:revert-masterCloses#26915 from wangyum/revert-master.
Authored-by: Yuming Wang <yumwang@ebay.com>
Signed-off-by: Yuming Wang <wgyumg@gmail.com>
### What changes were proposed in this pull request?
When set `spark.yarn.jars=/xxx/xxx` which is just a no schema path, spark will throw a NullPointerException.
The reason is hdfs will return null if pathFs.globStatus(path) is not exist, and spark just use `pathFs.globStatus(path).filter(_.isFile())` without check it.
### Why are the changes needed?
Avoid NullPointerException.
### Does this PR introduce any user-facing change?
Yes. User will get a FileNotFoundException instead NullPointerException when `spark.yarn.jars` does not have schema and not exists.
### How was this patch tested?
Add UT.
Closes#26462 from ulysses-you/check-yarn-jars-path-exist.
Authored-by: ulysses <youxiduo@weidian.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
### What changes were proposed in this pull request?
With this change, executor's bindAddress is passed as an input parameter for RPCEnv.create.
A previous PR https://github.com/apache/spark/pull/21261 which addressed the same, was using a Spark Conf property to get the bindAddress which wouldn't have worked for multiple executors.
This PR is to enable anyone overriding CoarseGrainedExecutorBackend with their custom one to be able to invoke CoarseGrainedExecutorBackend.main() along with the option to configure bindAddress.
### Why are the changes needed?
This is required when Kernel-based Virtual Machine (KVM)'s are used inside Linux container where the hostname is not the same as container hostname.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Tested by running jobs with executors on KVMs inside a linux container.
Closes#26331 from nishchalv/SPARK-29670.
Lead-authored-by: Nishchal Venkataramana <nishchal@apple.com>
Co-authored-by: nishchal <nishchal@apple.com>
Signed-off-by: DB Tsai <d_tsai@apple.com>
### What changes were proposed in this pull request?
Priority for YARN to define pending applications ordering policy, those with higher priority have a better opportunity to be activated. YARN CapacityScheduler only.
### Why are the changes needed?
Ordering pending spark apps
### Does this PR introduce any user-facing change?
add a conf
### How was this patch tested?
add ut
Closes#26255 from yaooqinn/SPARK-29603.
Authored-by: Kent Yao <yaooqinn@hotmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
### What changes were proposed in this pull request?
To push the built jars to maven release repository, we need to remove the 'SNAPSHOT' tag from the version name.
Made the following changes in this PR:
* Update all the `3.0.0-SNAPSHOT` version name to `3.0.0-preview`
* Update the sparkR version number check logic to allow jvm version like `3.0.0-preview`
**Please note those changes were generated by the release script in the past, but this time since we manually add tags on master branch, we need to manually apply those changes too.**
We shall revert the changes after 3.0.0-preview release passed.
### Why are the changes needed?
To make the maven release repository to accept the built jars.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
N/A
### What changes were proposed in this pull request?
To push the built jars to maven release repository, we need to remove the 'SNAPSHOT' tag from the version name.
Made the following changes in this PR:
* Update all the `3.0.0-SNAPSHOT` version name to `3.0.0-preview`
* Update the PySpark version from `3.0.0.dev0` to `3.0.0`
**Please note those changes were generated by the release script in the past, but this time since we manually add tags on master branch, we need to manually apply those changes too.**
We shall revert the changes after 3.0.0-preview release passed.
### Why are the changes needed?
To make the maven release repository to accept the built jars.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
N/A
Closes#26243 from jiangxb1987/3.0.0-preview-prepare.
Lead-authored-by: Xingbo Jiang <xingbo.jiang@databricks.com>
Co-authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: Xingbo Jiang <xingbo.jiang@databricks.com>
### What changes were proposed in this pull request?
This PR aims to remove `scalatest` deprecation warnings with the following changes.
- `org.scalatest.mockito.MockitoSugar` -> `org.scalatestplus.mockito.MockitoSugar`
- `org.scalatest.selenium.WebBrowser` -> `org.scalatestplus.selenium.WebBrowser`
- `org.scalatest.prop.Checkers` -> `org.scalatestplus.scalacheck.Checkers`
- `org.scalatest.prop.GeneratorDrivenPropertyChecks` -> `org.scalatestplus.scalacheck.ScalaCheckDrivenPropertyChecks`
### Why are the changes needed?
According to the Jenkins logs, there are 118 warnings about this.
```
grep "is deprecated" ~/consoleText | grep scalatest | wc -l
118
```
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
After Jenkins passes, we need to check the Jenkins log.
Closes#25982 from dongjoon-hyun/SPARK-29307.
Authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
Scala 2.13 emits a deprecation warning for procedure-like declarations:
```
def foo() {
...
```
This is equivalent to the following, so should be changed to avoid a warning:
```
def foo(): Unit = {
...
```
### Why are the changes needed?
It will avoid about a thousand compiler warnings when we start to support Scala 2.13. I wanted to make the change in 3.0 as there are less likely to be back-ports from 3.0 to 2.4 than 3.1 to 3.0, for example, minimizing that downside to touching so many files.
Unfortunately, that makes this quite a big change.
### Does this PR introduce any user-facing change?
No behavior change at all.
### How was this patch tested?
Existing tests.
Closes#25968 from srowen/SPARK-29291.
Authored-by: Sean Owen <sean.owen@databricks.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
In `ApplicationMaster.Reporter` thread, fatal exception information is swallowed. It's better to expose it.
We found our thrift server was shutdown due to a fatal exception but no useful information from log.
> 19/09/16 06:59:54,498 INFO [Reporter] yarn.ApplicationMaster:54 : Final app status: FAILED, exitCode: 12, (reason: Exception was thrown 1 time(s) from Reporter thread.)
19/09/16 06:59:54,500 ERROR [Driver] thriftserver.HiveThriftServer2:91 : Error starting HiveThriftServer2
java.lang.InterruptedException: sleep interrupted
at java.lang.Thread.sleep(Native Method)
at org.apache.spark.sql.hive.thriftserver.HiveThriftServer2$.main(HiveThriftServer2.scala:160)
at org.apache.spark.sql.hive.thriftserver.HiveThriftServer2.main(HiveThriftServer2.scala)
at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)
at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
at java.lang.reflect.Method.invoke(Method.java:498)
at org.apache.spark.deploy.yarn.ApplicationMaster$$anon$4.run(ApplicationMaster.scala:708)
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Manual test
Closes#25810 from LantaoJin/SPARK-29112.
Authored-by: LantaoJin <jinlantao@gmail.com>
Signed-off-by: jerryshao <jerryshao@tencent.com>
### What changes were proposed in this pull request?
Make r file extension check case insensitive for spark-submit.
### Why are the changes needed?
spark-submit does not accept `.r` files as R scripts. Some codebases have r files that end with lowercase file extensions. It is inconvenient to use spark-submit with lowercase extension R files. The error is not very clear (https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/deploy/SparkSubmitArguments.scala#L232).
```
$ ./bin/spark-submit examples/src/main/r/dataframe.r
Exception in thread "main" org.apache.spark.SparkException: Cannot load main class from JAR file:/Users/dongjoon/APACHE/spark-release/spark-2.4.4-bin-hadoop2.7/examples/src/main/r/dataframe.r
```
### Does this PR introduce any user-facing change?
Yes. spark-submit can now be used to run R scripts with `.r` file extension.
### How was this patch tested?
Manual.
```
$ mv examples/src/main/r/dataframe.R examples/src/main/r/dataframe.r
$ ./bin/spark-submit examples/src/main/r/dataframe.r
```
Closes#25778 from Loquats/r-case.
Authored-by: Andy Zhang <yue.zhang@databricks.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
- Remove SQLContext.createExternalTable and Catalog.createExternalTable, deprecated in favor of createTable since 2.2.0, plus tests of deprecated methods
- Remove HiveContext, deprecated in 2.0.0, in favor of `SparkSession.builder.enableHiveSupport`
- Remove deprecated KinesisUtils.createStream methods, plus tests of deprecated methods, deprecate in 2.2.0
- Remove deprecated MLlib (not Spark ML) linear method support, mostly utility constructors and 'train' methods, and associated docs. This includes methods in LinearRegression, LogisticRegression, Lasso, RidgeRegression. These have been deprecated since 2.0.0
- Remove deprecated Pyspark MLlib linear method support, including LogisticRegressionWithSGD, LinearRegressionWithSGD, LassoWithSGD
- Remove 'runs' argument in KMeans.train() method, which has been a no-op since 2.0.0
- Remove deprecated ChiSqSelector isSorted protected method
- Remove deprecated 'yarn-cluster' and 'yarn-client' master argument in favor of 'yarn' and deploy mode 'cluster', etc
Notes:
- I was not able to remove deprecated DataFrameReader.json(RDD) in favor of DataFrameReader.json(Dataset); the former was deprecated in 2.2.0, but, it is still needed to support Pyspark's .json() method, which can't use a Dataset.
- Looks like SQLContext.createExternalTable was not actually deprecated in Pyspark, but, almost certainly was meant to be? Catalog.createExternalTable was.
- I afterwards noted that the toDegrees, toRadians functions were almost removed fully in SPARK-25908, but Felix suggested keeping just the R version as they hadn't been technically deprecated. I'd like to revisit that. Do we really want the inconsistency? I'm not against reverting it again, but then that implies leaving SQLContext.createExternalTable just in Pyspark too, which seems weird.
- I *kept* LogisticRegressionWithSGD, LinearRegressionWithSGD, LassoWithSGD, RidgeRegressionWithSGD in Pyspark, though deprecated, as it is hard to remove them (still used by StreamingLogisticRegressionWithSGD?) and they are not fully removed in Scala. Maybe should not have been deprecated.
### Why are the changes needed?
Deprecated items are easiest to remove in a major release, so we should do so as much as possible for Spark 3. This does not target items deprecated 'recently' as of Spark 2.3, which is still 18 months old.
### Does this PR introduce any user-facing change?
Yes, in that deprecated items are removed from some public APIs.
### How was this patch tested?
Existing tests.
Closes#25684 from srowen/SPARK-28980.
Lead-authored-by: Sean Owen <sean.owen@databricks.com>
Co-authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
If MEMORY_OFFHEAP_ENABLED is true, add MEMORY_OFFHEAP_SIZE to resource requested for executor to ensure instance has enough memory to use.
In this pr add a helper method `executorOffHeapMemorySizeAsMb` in `YarnSparkHadoopUtil`.
## How was this patch tested?
Add 3 new test suite to test `YarnSparkHadoopUtil#executorOffHeapMemorySizeAsMb`
Closes#25309 from LuciferYang/spark-28577.
Authored-by: yangjie01 <yangjie01@baidu.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
## What changes were proposed in this pull request?
This fixes issues that can arise when the jars for different hadoop versions mix, and short-circuits the case where we are running with a spark that was not built for yarn 3 (resource support).
## How was this patch tested?
I tested it manually.
Closes#25403 from abellina/SPARK-28679.
Authored-by: Alessandro Bellina <abellina@nvidia.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
This change makes the client mode AM ignore any login configuration,
which is now always handled by the driver. The previous code tried
to achieve that by modifying the configuration visible to the AM, but
that missed the case where old configuration names were being used.
Tested in real cluster with reproduction provided in the bug.
Closes#25467 from vanzin/SPARK-28634.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
After SPARK-27677, the shuffle client not only handles the shuffle block but also responsible for local persist RDD blocks. For better code scalability and precise semantics(as the [discussion](https://github.com/apache/spark/pull/24892#discussion_r300173331)), here we did several changes:
- Rename ShuffleClient to BlockStoreClient.
- Correspondingly rename the ExternalShuffleClient to ExternalBlockStoreClient, also change the server-side class from ExternalShuffleBlockHandler to ExternalBlockHandler.
- Move MesosExternalBlockStoreClient to Mesos package.
Note, we still keep the name of BlockTransferService, because the `Service` contains both client and server, also the name of BlockTransferService is not referencing shuffle client only.
## How was this patch tested?
Existing UT.
Closes#25327 from xuanyuanking/SPARK-28593.
Lead-authored-by: Yuanjian Li <xyliyuanjian@gmail.com>
Co-authored-by: Yuanjian Li <yuanjian.li@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
we are adding in generic resource support into spark where we have suffix for the amount of the resource so that we could support other configs.
Spark on yarn already had added configs to request resources via the configs spark.yarn.{executor/driver/am}.resource=<some amount>, where the <some amount> is value and unit together. We should change those configs to have a `.amount` suffix on them to match the spark configs and to allow future configs to be more easily added. YARN itself already supports tags and attributes so if we want the user to be able to pass those from spark at some point having a suffix makes sense. it would allow for a spark.yarn.{executor/driver/am}.resource.{resource}.tag= type config.
## How was this patch tested?
Tested via unit tests and manually on a yarn 3.x cluster with GPU resources configured on.
Closes#24989 from tgravescs/SPARK-27959-yarn-resourceconfigs.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
SparkRackResolver generates an INFO message every time is called with 0 arguments.
In this PR I've deleted it because it's too verbose.
## How was this patch tested?
Existing unit tests + spark-shell.
Closes#24935 from gaborgsomogyi/SPARK-28005.
Authored-by: Gabor Somogyi <gabor.g.somogyi@gmail.com>
Signed-off-by: Imran Rashid <irashid@cloudera.com>
## What changes were proposed in this pull request?
Continue the work from https://github.com/apache/spark/pull/24821. Refactor resource handling code to make the code more readable. Major changes:
* Moved resource-related classes to `spark.resource` from `spark`.
* Added ResourceUtils and helper classes so we don't need to directly deal with Spark conf.
* ResourceID: resource identifier and it provides conf keys
* ResourceRequest/Allocation: abstraction for requested and allocated resources
* Added `TestResourceIDs` to reference commonly used resource IDs in tests like `spark.executor.resource.gpu`.
cc: tgravescs jiangxb1987 Ngone51
## How was this patch tested?
Unit tests for added utils and existing unit tests.
Closes#24856 from mengxr/SPARK-27823.
Lead-authored-by: Xiangrui Meng <meng@databricks.com>
Co-authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Xingbo Jiang <xingbo.jiang@databricks.com>
## What changes were proposed in this pull request?
Change the resource config spark.{executor/driver}.resource.{resourceName}.count to .amount to allow future usage of containing both a count and a unit. Right now we only support counts - # of gpus for instance, but in the future we may want to support units for things like memory - 25G. I think making the user only have to specify a single config .amount is better then making them specify 2 separate configs of a .count and then a .unit. Change it now since its a user facing config.
Amount also matches how the spark on yarn configs are setup.
## How was this patch tested?
Unit tests and manually verified on yarn and local cluster mode
Closes#24810 from tgravescs/SPARK-27760-amount.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
## What changes were proposed in this pull request?
This PR targets to deduplicate hardcoded `py4j-0.10.8.1-src.zip` in order to make py4j upgrade easier.
## How was this patch tested?
N/A
Closes#24770 from HyukjinKwon/minor-py4j-dedup.
Authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
Add a metric for number of exceptions caught in the `ExternalShuffleBlockHandler`, the idea being that spikes in this metric over some time window (or more desirably, the lack thereof) can be used as an indicator of the health of an external shuffle service. (Where "health" refers to its ability to successfully respond to client requests.)
## How was this patch tested?
Deployed a build of this PR to a YARN cluster, and confirmed that the NodeManagers' JMX metrics include `numCaughtExceptions`.
Closes#24645 from sjrand/SPARK-27773.
Authored-by: Steven Rand <srand@palantir.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
Add yarn support for GPU-aware scheduling. Since SPARK-20327 already added yarn custom resource support, this jira is really just making sure the spark resource configs get mapped into the yarn resource configs and user doesn't specify both yarn and spark config for the known types of resources (gpu and fpga are the known types on yarn).
You can find more details on the design and requirements documented: https://issues.apache.org/jira/browse/SPARK-27376
Note that the running on yarn docs already state to use it, it must be yarn 3.0+. We will add any further documentation under SPARK-20327
## How was this patch tested?
Unit tests and manually testing on yarn cluster
Closes#24634 from tgravescs/SPARK-27361.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
avoid hardcoded configs in `SparkConf` and `SparkSubmit` and test
## How was this patch tested?
N/A
Closes#24631 from wenxuanguan/minor-fix.
Authored-by: wenxuanguan <choose_home@126.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
Add in GPU and generic resource type allocation to the executors.
Note this is part of a bigger feature for gpu-aware scheduling and is just how the executor find the resources. The general flow :
- users ask for a certain set of resources, for instance number of gpus - each cluster manager has a specific way to do this.
- cluster manager allocates a container or set of resources (standalone mode)
- When spark launches the executor in that container, the executor either has to be told what resources it has or it has to auto discover them.
- Executor has to register with Driver and tell the driver the set of resources it has so the scheduler can use that to schedule tasks that requires a certain amount of each of those resources
In this pr I added configs and arguments to the executor to be able discover resources. The argument to the executor is intended to be used by standalone mode or other cluster managers that don't have isolation so that it can assign specific resources to specific executors in case there are multiple executors on a node. The argument is a file contains JSON Array of ResourceInformation objects.
The discovery script is meant to be used in an isolated environment where the executor only sees the resources it should use.
Note that there will be follow on PRs to add other parts like the scheduler part. See the epic high level jira: https://issues.apache.org/jira/browse/SPARK-24615
## How was this patch tested?
Added unit tests and manually tested.
Please review http://spark.apache.org/contributing.html before opening a pull request.
Closes#24406 from tgravescs/gpu-sched-executor-clean.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
## What changes were proposed in this pull request?
Fixed the `spark-<version>-yarn-shuffle.jar` artifact packaging to shade the native netty libraries:
- shade the `META-INF/native/libnetty_*` native libraries when packagin
the yarn shuffle service jar. This is required as netty library loader
derives that based on shaded package name.
- updated the `org/spark_project` shade package prefix to `org/sparkproject`
(i.e. removed underscore) as the former breaks the netty native lib loading.
This was causing the yarn external shuffle service to fail
when spark.shuffle.io.mode=EPOLL
## How was this patch tested?
Manual tests
Closes#24502 from amuraru/SPARK-27610_master.
Authored-by: Adi Muraru <amuraru@adobe.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
This PR avoids usage of reflective calls in Scala. It removes the import that suppresses the warnings and rewrites code in small ways to avoid accessing methods that aren't technically accessible.
## How was this patch tested?
Existing tests.
Closes#24463 from srowen/SPARK-27571.
Authored-by: Sean Owen <sean.owen@databricks.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
This patch makes several test flakiness fixes.
## How was this patch tested?
N/A
Closes#24434 from gatorsmile/fixFlakyTest.
Lead-authored-by: gatorsmile <gatorsmile@gmail.com>
Co-authored-by: Hyukjin Kwon <gurwls223@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
Fix build warnings -- see some details below.
But mostly, remove use of postfix syntax where it causes warnings without the `scala.language.postfixOps` import. This is mostly in expressions like "120000 milliseconds". Which, I'd like to simplify to things like "2.minutes" anyway.
## How was this patch tested?
Existing tests.
Closes#24314 from srowen/SPARK-27404.
Authored-by: Sean Owen <sean.owen@databricks.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
When you submit a stage on a large cluster, rack resolving takes a long time when initializing TaskSetManager because a script is invoked to resolve the rack of each host, one by one.
Based on current implementation, it takes 30~40 seconds to resolve the racks in our 5000 nodes' cluster. After applied the patch, it decreased to less than 15 seconds.
YARN-9332 has added an interface to handle multiple hosts in one invocation to save time. But before upgrading to the newest Hadoop, we could construct the same tool in Spark to resolve this issue.
## How was this patch tested?
UT and manually testing on a 5000 node cluster.
Closes#24245 from squito/SPARK-13704_update.
Lead-authored-by: LantaoJin <jinlantao@gmail.com>
Co-authored-by: Imran Rashid <irashid@cloudera.com>
Signed-off-by: Imran Rashid <irashid@cloudera.com>
## What changes were proposed in this pull request?
Fix testing issues with `yarn` module in Hadoop-3:
1. Upgrade jersey-1 to `1.19` to fix ```Cause: java.lang.NoClassDefFoundError: com/sun/jersey/spi/container/servlet/ServletContainer```.
2. Copy `ServerSocketUtil` from hadoop-common-project/hadoop-common/src/test/java/org/apache/hadoop/net/ServerSocketUtil.java to fix ```java.lang.NoClassDefFoundError: org/apache/hadoop/net/ServerSocketUtil```.
3. Adapte `SessionHandler` from jetty-9.3.25.v20180904/jetty-server/src/main/java/org/eclipse/jetty/server/session/SessionHandler.java to fix ```java.lang.NoSuchMethodError: org.eclipse.jetty.server.session.SessionHandler.getSessionManager()Lorg/eclipse/jetty/server/SessionManager```.
## How was this patch tested?
manual tests:
```shell
build/sbt yarn/test -Pyarn
build/sbt yarn/test -Phadoop-3.2 -Pyarn
build/mvn -Dtest=none -DwildcardSuites=org.apache.spark.deploy.yarn.YarnClusterSuite -pl resource-managers/yarn test -Pyarn
build/mvn -Dtest=none -DwildcardSuites=org.apache.spark.deploy.yarn.YarnClusterSuite -pl resource-managers/yarn test -Pyarn -Phadoop-3.2
```
Closes#24115 from wangyum/hadoop3-yarn.
Authored-by: Yuming Wang <yumwang@ebay.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
Use Single Abstract Method syntax where possible (and minor related cleanup). Comments below. No logic should change here.
## How was this patch tested?
Existing tests.
Closes#24241 from srowen/SPARK-27323.
Authored-by: Sean Owen <sean.owen@databricks.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
There is some hardcode configs in code, I think it best to modify。
## How was this patch tested?
Existing tests
Closes#24103 from wangjiaochun/yarnHardCode.
Authored-by: 10087686 <wang.jiaochun@zte.com.cn>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
To avoid the case where the YARN libraries would swallow the exception and
prevent YarnAllocator from shutting down, call the offending code in a
separate thread, so that the parent thread can respond appropriately to
the shut down.
As a safeguard, also explicitly stop the executor launch thread pool when
shutting down the application, to prevent new executors from coming up
after the application started its shutdown.
Tested with unit tests + some internal tests on real cluster.
Closes#24017 from vanzin/SPARK-27094.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
When we run YarnSchedulerBackendSuite, the class path seems to be made from the classes folder(resource-managers/yarn/target/scala-2.12/classes) instead of jar (resource-managers/yarn/target/spark-yarn_2.12-3.0.0-SNAPSHOT.jar) . ui.getHandlers is in spark-core and its loaded from spark-core.jar which is shaded and hence refers to org.spark_project.jetty.servlet.ServletContextHandler
Here in org.apache.spark.scheduler.cluster.YarnSchedulerBackend, as its not shaded, it expects org.eclipse.jetty.servlet.ServletContextHandler
Refer discussion https://issues.apache.org/jira/browse/SPARK-27122?focusedCommentId=16792318&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-16792318
Hence as a fix, org.apache.spark.ui.WebUI must only return a wrapper class instance or references so that Jetty classes can be avoided in getters which are accessed outside spark-core
## How was this patch tested?
Existing UT can pass
Closes#24088 from ajithme/shadebug.
Authored-by: Ajith <ajith2489@gmail.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
Currently, when enabled streaming dynamic allocation for streaming applications, the maxNumExecutorFailures in ApplicationMaster is still computed with `spark.dynamicAllocation.maxExecutors`.
Actually, we should consider `spark.streaming.dynamicAllocation.maxExecutors` instead.
Related codes:
f87153a3ac/resource-managers/yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala (L101)
## How was this patch tested?
NA
Please review http://spark.apache.org/contributing.html before opening a pull request.
Closes#23845 from liupc/Fix-incorrect-maxNumExecutorFailures-for-streaming.
Lead-authored-by: Liupengcheng <liupengcheng@xiaomi.com>
Co-authored-by: liupengcheng <liupengcheng@xiaomi.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
This pr adds 2 maven properties to help us upgrade the built-in Hive.
| Property Name | Default | In future |
| ------ | ------ | ------ |
| hive.classifier | (none) | core |
| hive.parquet.group | com.twitter | org.apache.parquet |
## How was this patch tested?
existing tests
Closes#23996 from wangyum/add_2_maven_properties.
Authored-by: Yuming Wang <yumwang@ebay.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
Introducing new config for initially blacklisted YARN nodes.
## How was this patch tested?
With existing and a new unit test.
Closes#23616 from attilapiros/SPARK-26688.
Lead-authored-by: “attilapiros” <piros.attila.zsolt@gmail.com>
Co-authored-by: Attila Zsolt Piros <2017933+attilapiros@users.noreply.github.com>
Signed-off-by: Imran Rashid <irashid@cloudera.com>
## What changes were proposed in this pull request?
`MetricsSystem` instance creations have a scattered distribution in the project code. So do their names. It may cause some inconvenience for browsing and management.
This PR tries to put them together. In this way, we can have a uniform location for adding or removing them, and have a overall view of `MetircsSystem `instances in current project.
It's also helpful for maintaining user documents by avoiding missing something.
## How was this patch tested?
Existing unit tests.
Closes#23869 from SongYadong/metrics_system_inst_manage.
Lead-authored-by: SongYadong <song.yadong1@zte.com.cn>
Co-authored-by: walter2001 <ydsong2007@163.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
Comparing whether Boolean expression is equal to true is redundant
For example:
The datatype of `a` is boolean.
Before:
if (a == true)
After:
if (a)
## How was this patch tested?
N/A
Closes#23884 from 10110346/simplifyboolean.
Authored-by: liuxian <liu.xian3@zte.com.cn>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
This patch applies redaction to command line arguments before logging them. This applies to two resource managers: standalone cluster and YARN.
This patch only concerns about arguments starting with `-D` since Spark is likely passing the Spark configuration to command line arguments as `-Dspark.blabla=blabla`. More change is necessary if we also want to handle the case of `--conf spark.blabla=blabla`.
## How was this patch tested?
Added UT for redact logic. This patch only touches how to log so not easy to add UT regarding it.
Closes#23820 from HeartSaVioR/MINOR-redact-command-line-args-for-running-driver-executor.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
Since the yarn module is actually private to Spark, this interface was never
actually "public". Since it has no use inside of Spark, let's avoid adding
a yarn-specific extension that isn't public, and point any potential users
are more general solutions (like using a SparkListener).
Closes#23839 from vanzin/SPARK-26788.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
This PR aims to fix some outdated comments about task schedulers.
1. Change "ClusterScheduler" to "YarnScheduler" in comments of `YarnClusterScheduler`
According to [SPARK-1140 Remove references to ClusterScheduler](https://issues.apache.org/jira/browse/SPARK-1140), ClusterScheduler is not used anymore.
I also searched "ClusterScheduler" within the whole project, no other occurrences are found in comments or test cases. Note classes like `YarnClusterSchedulerBackend` or `MesosClusterScheduler` are not relevant.
2. Update comments about `statusUpdate` from `TaskSetManager`
`statusUpdate` has been moved to `TaskSchedulerImpl`. StatusUpdate event handling is delegated to `handleSuccessfulTask`/`handleFailedTask`.
## How was this patch tested?
N/A. Fix comments only.
Closes#23844 from seancxmao/taskscheduler-comments.
Authored-by: seancxmao <seancxmao@gmail.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
Currently, when running applications on yarn mode, the app staging directory of is controlled by `spark.yarn.stagingDir` config if specified, and this directory cannot separate different users, sometimes, it's inconvenient for file and quota management for users.
Sometimes, there might be an unexpected increasing of the staging files, two possible reasons are:
1. The `spark.yarn.preserve.staging.files` provided can be misused by users
2. cron task constantly starting new applications on non-existent yarn queue(wrong configuration).
But now, we are not easy to find out the which user obtains the most HDFS files or spaces.
what's more, even we want set HDFS name quota or space quota for each user to limit the increase is impossible.
So I propose to add user sub directories under this app staging directory which is more clear.
existing UT
Closes#23786 from liupc/Support-user-level-app-staging-dir.
Authored-by: Liupengcheng <liupengcheng@xiaomi.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
The test "RequestExecutors reflects node blacklist and is serializable" is flaky because of multi threaded access of the mock task scheduler. For details check [Mockito FAQ (occasional exceptions like: WrongTypeOfReturnValue)](https://github.com/mockito/mockito/wiki/FAQ#is-mockito-thread-safe). So instead of mocking the task scheduler in the test TaskSchedulerImpl is simply subclassed.
This multithreaded access of the `nodeBlacklist()` method is coming from:
1) the unit test thread via calling of the method `prepareRequestExecutors()`
2) the `DriverEndpoint.onStart` which runs a periodic task that ends up calling this method
Existing unittest.
Closes#23801 from attilapiros/SPARK-26891.
Authored-by: “attilapiros” <piros.attila.zsolt@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
`HadoopDelegationTokenProvider` has basically the same functionality just like `ServiceCredentialProvider` so the interfaces can be merged.
`YARNHadoopDelegationTokenManager` now loads `ServiceCredentialProvider`s in one step. The drawback of this if one provider fails all others are not loaded. `HadoopDelegationTokenManager` loads `HadoopDelegationTokenProvider`s independently so it provides more robust behaviour.
In this PR I've I've made the following changes:
* Deleted `YARNHadoopDelegationTokenManager` and `ServiceCredentialProvider`
* Made `HadoopDelegationTokenProvider` a `DeveloperApi`
## How was this patch tested?
Existing unit tests.
Closes#23686 from gaborgsomogyi/SPARK-26772.
Authored-by: Gabor Somogyi <gabor.g.somogyi@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
This patch proposes to change the approach on extracting log urls as well as attributes from YARN executor:
- AS-IS: extract information from `Container` API and include them to container launch context
- TO-BE: let YARN executor self-extracting information
This approach leads us to populate more attributes like nodemanager's IPC port which can let us configure custom log url to JHS log url directly.
## How was this patch tested?
Existing unit tests.
Closes#23706 from HeartSaVioR/SPARK-26790.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
In the PR, I propose to use `System.nanoTime()` instead of `System.currentTimeMillis()` in measurements of time intervals.
`System.currentTimeMillis()` returns current wallclock time and will follow changes to the system clock. Thus, negative wallclock adjustments can cause timeouts to "hang" for a long time (until wallclock time has caught up to its previous value again). This can happen when ntpd does a "step" after the network has been disconnected for some time. The most canonical example is during system bootup when DHCP takes longer than usual. This can lead to failures that are really hard to understand/reproduce. `System.nanoTime()` is guaranteed to be monotonically increasing irrespective of wallclock changes.
## How was this patch tested?
By existing test suites.
Closes#23727 from MaxGekk/system-nanotime.
Lead-authored-by: Maxim Gekk <max.gekk@gmail.com>
Co-authored-by: Maxim Gekk <maxim.gekk@databricks.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
Delegation token providers interface now has a parameter `fileSystems` but this is needed only for `HadoopFSDelegationTokenProvider`.
In this PR I've addressed this issue in the following way:
* Removed `fileSystems` parameter from `HadoopDelegationTokenProvider`
* Moved `YarnSparkHadoopUtil.hadoopFSsToAccess` into `HadoopFSDelegationTokenProvider`
* Moved `spark.yarn.stagingDir` into core
* Moved `spark.yarn.access.namenodes` into core and renamed to `spark.kerberos.access.namenodes`
* Moved `spark.yarn.access.hadoopFileSystems` into core and renamed to `spark.kerberos.access.hadoopFileSystems`
## How was this patch tested?
Existing unit tests.
Closes#23698 from gaborgsomogyi/SPARK-26766.
Authored-by: Gabor Somogyi <gabor.g.somogyi@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
This patch proposes adding a new configuration on SHS: custom executor log URL pattern. This will enable end users to replace executor logs to other than RM provide, like external log service, which enables to serve executor logs when NodeManager becomes unavailable in case of YARN.
End users can build their own of custom executor log URLs with pre-defined patterns which would be vary on each resource manager. This patch adds some patterns to YARN resource manager. (For others, there's even no executor log url available so cannot define patterns as well.)
Please refer the doc change as well as added UTs in this patch to see how to set up the feature.
## How was this patch tested?
Added UT, as well as manual test with YARN cluster
Closes#23260 from HeartSaVioR/SPARK-26311.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
This change addes a new mode for credential renewal that does not require
a keytab; it uses the local ticket cache instead, so it works while the
user keeps the cache valid.
This can be useful for, e.g., people running long spark-shell sessions where
their kerberos login is kept up-to-date.
The main change to enable this behavior is in HadoopDelegationTokenManager,
with a small change in the HDFS token provider. The other changes are to avoid
creating duplicate tokens when submitting the application to YARN; they allow
the tokens from the scheduler to be sent to the YARN AM, reducing the round trips
to HDFS.
For that, the scheduler initialization code was changed a little bit so that
the tokens are available when the YARN client is initialized. That basically
takes care of a long-standing TODO that was in the code to clean up configuration
propagation to the driver's RPC endpoint (in CoarseGrainedSchedulerBackend).
Tested with an app designed to stress this functionality, with both keytab and
cache-based logins. Some basic kerberos tests on k8s also.
Closes#23525 from vanzin/SPARK-26595.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
Providing a new configuration "spark.yarn.un-managed-am" (defaults to false) to enable the Unmanaged AM Application in Yarn Client mode which launches the Application Master service as part of the Client. It utilizes the existing code for communicating between the Application Master <-> Task Scheduler for the container requests/allocations/launch, and eliminates these,
1. Allocating and launching the Application Master container
2. Remote Node/Process communication between Application Master <-> Task Scheduler
## How was this patch tested?
I verified manually running the applications in yarn-client mode with "spark.yarn.un-managed-am" enabled, and also ensured that there is no impact to the existing execution flows.
I would like to hear others feedback/thoughts on this.
Closes#19616 from devaraj-kavali/SPARK-22404.
Authored-by: Devaraj K <devaraj@apache.org>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
The PR makes hardcoded `spark.dynamicAllocation`, `spark.scheduler`, `spark.rpc`, `spark.task`, `spark.speculation`, and `spark.cleaner` configs to use `ConfigEntry`.
## How was this patch tested?
Existing tests
Closes#23416 from kiszk/SPARK-26463.
Authored-by: Kazuaki Ishizaki <ishizaki@jp.ibm.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
The PR makes hardcoded configs below to use `ConfigEntry`.
* spark.kryo
* spark.kryoserializer
* spark.serializer
* spark.jars
* spark.files
* spark.submit
* spark.deploy
* spark.worker
This patch doesn't change configs which are not relevant to SparkConf (e.g. system properties).
## How was this patch tested?
Existing tests.
Closes#23532 from HeartSaVioR/SPARK-26466-v2.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan@gmail.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
The PR makes hardcoded configs below to use `ConfigEntry`.
* spark.ui
* spark.ssl
* spark.authenticate
* spark.master.rest
* spark.master.ui
* spark.metrics
* spark.admin
* spark.modify.acl
This patch doesn't change configs which are not relevant to SparkConf (e.g. system properties).
## How was this patch tested?
Existing tests.
Closes#23423 from HeartSaVioR/SPARK-26466.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
Currently, `Client#createConfArchive` do not handle IOException, and some detail info is not provided in logs. Sometimes, this may delay the time of locating the root cause of io error.
This PR will add debug logs for confArchive when preparing local resource.
## How was this patch tested?
unittest
Closes#23444 from liupc/Add-logs-for-IOException-when-preparing-local-resource.
Authored-by: Liupengcheng <liupengcheng@xiaomi.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
Currently there is code scattered in a bunch of places to do different
things related to HTTP security, such as access control, setting
security-related headers, and filtering out bad content. This makes it
really easy to miss these things when writing new UI code.
This change creates a new filter that does all of those things, and
makes sure that all servlet handlers that are attached to the UI get
the new filter and any user-defined filters consistently. The extent
of the actual features should be the same as before.
The new filter is added at the end of the filter chain, because authentication
is done by custom filters and thus needs to happen first. This means that
custom filters see unfiltered HTTP requests - which is actually the current
behavior anyway.
As a side-effect of some of the code refactoring, handlers added after
the initial set also get wrapped with a GzipHandler, which didn't happen
before.
Tested with added unit tests and in a history server with SPNEGO auth
configured.
Closes#23302 from vanzin/SPARK-24522.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Imran Rashid <irashid@cloudera.com>
This change modifies the behavior of the delegation token code when running
on YARN, so that the driver controls the renewal, in both client and cluster
mode. For that, a few different things were changed:
* The AM code only runs code that needs DTs when DTs are available.
In a way, this restores the AM behavior to what it was pre-SPARK-23361, but
keeping the fix added in that bug. Basically, all the AM code is run in a
"UGI.doAs()" block; but code that needs to talk to HDFS (basically the
distributed cache handling code) was delayed to the point where the driver
is up and running, and thus when valid delegation tokens are available.
* SparkSubmit / ApplicationMaster now handle user login, not the token manager.
The previous AM code was relying on the token manager to keep the user
logged in when keytabs are used. This required some odd APIs in the token
manager and the AM so that the right UGI was exposed and used in the right
places.
After this change, the logged in user is handled separately from the token
manager, so the API was cleaned up, and, as explained above, the whole AM
runs under the logged in user, which also helps with simplifying some more code.
* Distributed cache configs are sent separately to the AM.
Because of the delayed initialization of the cached resources in the AM, it
became easier to write the cache config to a separate properties file instead
of bundling it with the rest of the Spark config. This also avoids having
to modify the SparkConf to hide things from the UI.
* Finally, the AM doesn't manage the token manager anymore.
The above changes allow the token manager to be completely handled by the
driver's scheduler backend code also in YARN mode (whether client or cluster),
making it similar to other RMs. To maintain the fix added in SPARK-23361 also
in client mode, the AM now sends an extra message to the driver on initialization
to fetch delegation tokens; and although it might not really be needed, the
driver also keeps the running AM updated when new tokens are created.
Tested in a kerberized cluster with the same tests used to validate SPARK-23361,
in both client and cluster mode. Also tested with a non-kerberized cluster.
Closes#23338 from vanzin/SPARK-25689.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Imran Rashid <irashid@cloudera.com>
## What changes were proposed in this pull request?
This PR upgrades Mockito from 1.10.19 to 2.23.4. The following changes are required.
- Replace `org.mockito.Matchers` with `org.mockito.ArgumentMatchers`
- Replace `anyObject` with `any`
- Replace `getArgumentAt` with `getArgument` and add type annotation.
- Use `isNull` matcher in case of `null` is invoked.
```scala
saslHandler.channelInactive(null);
- verify(handler).channelInactive(any(TransportClient.class));
+ verify(handler).channelInactive(isNull());
```
- Make and use `doReturn` wrapper to avoid [SI-4775](https://issues.scala-lang.org/browse/SI-4775)
```scala
private def doReturn(value: Any) = org.mockito.Mockito.doReturn(value, Seq.empty: _*)
```
## How was this patch tested?
Pass the Jenkins with the existing tests.
Closes#23452 from dongjoon-hyun/SPARK-26536.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
## What changes were proposed in this pull request?
The PR makes hardcoded spark.driver, spark.executor, and spark.cores.max configs to use `ConfigEntry`.
Note that some config keys are from `SparkLauncher` instead of defining in the config package object because the string is already defined in it and it does not depend on core module.
## How was this patch tested?
Existing tests.
Closes#23415 from ueshin/issues/SPARK-26445/hardcoded_driver_executor_configs.
Authored-by: Takuya UESHIN <ueshin@databricks.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
The PR makes hardcoded configs below to use ConfigEntry.
* spark.pyspark
* spark.python
* spark.r
This patch doesn't change configs which are not relevant to SparkConf (e.g. system properties, python source code)
## How was this patch tested?
Existing tests.
Closes#23428 from HeartSaVioR/SPARK-26489.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
Recently, the ability to expose the metrics for YARN Shuffle Service was added as part of [SPARK-18364](https://github.com/apache/spark/pull/22485). We need to add some metrics to be able to determine the number of active connections as well as open connections to the external shuffle service to benchmark network and connection issues on large cluster environments.
Added two more shuffle server metrics for Spark Yarn shuffle service: numRegisteredConnections which indicate the number of registered connections to the shuffle service and numActiveConnections which indicate the number of active connections to the shuffle service at any given point in time.
If these metrics are outputted to a file, we get something like this:
1533674653489 default.shuffleService: Hostname=server1.abc.com, openBlockRequestLatencyMillis_count=729, openBlockRequestLatencyMillis_rate15=0.7110833548897356, openBlockRequestLatencyMillis_rate5=1.657808981793011, openBlockRequestLatencyMillis_rate1=2.2404486061620474, openBlockRequestLatencyMillis_rateMean=0.9242558551196706,
numRegisteredConnections=35,
blockTransferRateBytes_count=2635880512, blockTransferRateBytes_rate15=2578547.6094160094, blockTransferRateBytes_rate5=6048721.726302424, blockTransferRateBytes_rate1=8548922.518223226, blockTransferRateBytes_rateMean=3341878.633637769, registeredExecutorsSize=5, registerExecutorRequestLatencyMillis_count=5, registerExecutorRequestLatencyMillis_rate15=0.0027973949328659836, registerExecutorRequestLatencyMillis_rate5=0.0021278007987206426, registerExecutorRequestLatencyMillis_rate1=2.8270296777387467E-6, registerExecutorRequestLatencyMillis_rateMean=0.006339206380043053, numActiveConnections=35
Closes#22498 from pgandhi999/SPARK-18364.
Authored-by: pgandhi <pgandhi@oath.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
As I mentioned in jira [SPARK-26269](https://issues.apache.org/jira/browse/SPARK-26269), in order to maxmize the use of cluster resource, this pr try to make `YarnAllocator` have the same blacklist behaviour with YARN.
## How was this patch tested?
Added.
Closes#23223 from Ngone51/dev-YarnAllocator-should-have-same-blacklist-behaviour-with-YARN.
Lead-authored-by: wuyi <ngone_5451@163.com>
Co-authored-by: Ngone51 <ngone_5451@163.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
## What changes were proposed in this pull request?
Right now, we cancel pending allocate requests by its sending order. I thing we can take
locality preference into account when do this to perfom least impact on task locality preference.
## How was this patch tested?
N.A.
Closes#23344 from Ngone51/dev-cancel-pending-allocate-requests-by-taking-locality-preference-into-account.
Authored-by: Ngone51 <ngone_5451@163.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
This change hooks up the k8s backed to the updated HadoopDelegationTokenManager,
so that delegation tokens are also available in client mode, and keytab-based token
renewal is enabled.
The change re-works the k8s feature steps related to kerberos so
that the driver does all the credential management and provides all
the needed information to executors - so nothing needs to be added
to executor pods. This also makes cluster mode behave a lot more
similarly to client mode, since no driver-related config steps are run
in the latter case.
The main two things that don't need to happen in executors anymore are:
- adding the Hadoop config to the executor pods: this is not needed
since the Spark driver will serialize the Hadoop config and send
it to executors when running tasks.
- mounting the kerberos config file in the executor pods: this is
not needed once you remove the above. The Hadoop conf sent by
the driver with the tasks is already resolved (i.e. has all the
kerberos names properly defined), so executors do not need access
to the kerberos realm information anymore.
The change also avoids creating delegation tokens unnecessarily.
This means that they'll only be created if a secret with tokens
was not provided, and if a keytab is not provided. In either of
those cases, the driver code will handle delegation tokens: in
cluster mode by creating a secret and stashing them, in client
mode by using existing mechanisms to send DTs to executors.
One last feature: the change also allows defining a keytab with
a "local:" URI. This is supported in client mode (although that's
the same as not saying "local:"), and in k8s cluster mode. This
allows the keytab to be mounted onto the image from a pre-existing
secret, for example.
Finally, the new code always sets SPARK_USER in the driver and
executor pods. This is in line with how other resource managers
behave: the submitting user reflects which user will access
Hadoop services in the app. (With kerberos, that's overridden
by the logged in user.) That user is unrelated to the OS user
the app is running as inside the containers.
Tested:
- client and cluster mode with kinit
- cluster mode with keytab
- cluster mode with local: keytab
- YARN cluster with keytab (to make sure it isn't broken)
Closes#22911 from vanzin/SPARK-25815.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
User specified filters are not applied to SQL tab in yarn mode, as it is overridden by the yarn AmIp filter.
So we need to append user provided filters (spark.ui.filters) with yarn filter.
## How was this patch tested?
【Test step】：
1) Launch spark sql with authentication filter as below:
2) spark-sql --master yarn --conf spark.ui.filters=org.apache.hadoop.security.authentication.server.AuthenticationFilter --conf spark.org.apache.hadoop.security.authentication.server.AuthenticationFilter.params="type=simple"
3) Go to Yarn application list UI link
4) Launch the application master for the Spark-SQL app ID and access all the tabs by appending tab name.
5) It will display an error for all tabs including SQL tab.(before able to access SQL tab,as Authentication filter is not applied for SQL tab)
6) Also can be verified with info logs,that Authentication filter applied to SQL tab.(before it is not applied).
I have attached the behaviour below in following order..
1) Command used
2) Before fix (logs and UI)
3) After fix (logs and UI)
**1) COMMAND USED**:
launching spark-sql with authentication filter.
**2) BEFORE FIX:**
**UI result:**
able to access SQL tab.
**logs**:
authentication filter not applied to SQL tab.
**3) AFTER FIX:**
**UI result**:
Not able to access SQL tab.
**in logs**:
Both yarn filter and Authentication filter applied to SQL tab.
Closes#23312 from chakravarthiT/SPARK-26255_ui.
Authored-by: chakravarthi <tcchakra@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
YARN applicationMaster metrics registration introduced in SPARK-24594 causes further registration of static metrics (Codegenerator and HiveExternalCatalog) and of JVM metrics, which I believe do not belong in this context.
This looks like an unintended side effect of using the start method of [[MetricsSystem]].
A possible solution proposed here, is to introduce startNoRegisterSources to avoid these additional registrations of static sources and of JVM sources in the case of YARN applicationMaster metrics (this could be useful for other metrics that may be added in the future).
## How was this patch tested?
Manually tested on a YARN cluster,
Closes#22279 from LucaCanali/YarnMetricsRemoveExtraSourceRegistration.
Lead-authored-by: Luca Canali <luca.canali@cern.ch>
Co-authored-by: LucaCanali <luca.canali@cern.ch>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
… incorrect.
## What changes were proposed in this pull request?
In the reported heartbeat information, the unit of the memory data is bytes, which is converted by the formatBytes() function in the utils.js file before being displayed in the interface. The cardinality of the unit conversion in the formatBytes function is 1000, which should be 1024.
Change the cardinality of the unit conversion in the formatBytes function to 1024.
## How was this patch tested?
manual tests
Please review http://spark.apache.org/contributing.html before opening a pull request.
Closes#22683 from httfighter/SPARK-25696.
Lead-authored-by: 韩田田00222924 <han.tiantian@zte.com.cn>
Co-authored-by: han.tiantian@zte.com.cn <han.tiantian@zte.com.cn>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
