## What changes were proposed in this pull request?
We propose not to hard-code the RPC port in the AM registration.
## How was this patch tested?
Tested application reports from a pseudo-distributed cluster
```
18/04/10 14:56:21 INFO Client:
client token: N/A
diagnostics: N/A
ApplicationMaster host: localhost
ApplicationMaster RPC port: 58338
queue: default
start time: 1523397373659
final status: UNDEFINED
tracking URL: http://localhost:8088/proxy/application_1523370127531_0016/
```
Author: Gera Shegalov <gera@apache.org>
Closes#21047 from gerashegalov/gera/am-to-rm-nmhost.
## What changes were proposed in this pull request?
`YarnAllocator` uses `numExecutorsRunning` to track the number of running executor. `numExecutorsRunning` is used to check if there're executors missing and need to allocate more.
In current code, `numExecutorsRunning` can be negative when driver asks to kill a same idle executor multiple times.
## How was this patch tested?
UT added
Author: jinxing <jinxing6042@126.com>
Closes#20781 from jinxing64/SPARK-23637.
Currently, the Spark AM relies on the initial set of tokens created by
the submission client to be able to talk to HDFS and other services that
require delegation tokens. This means that after those tokens expire, a
new AM will fail to start (e.g. when there is an application failure and
re-attempts are enabled).
This PR makes it so that the first thing the AM does when the user provides
a principal and keytab is to create new delegation tokens for use. This
makes sure that the AM can be started irrespective of how old the original
token set is. It also allows all of the token management to be done by the
AM - there is no need for the submission client to set configuration values
to tell the AM when to renew tokens.
Note that even though in this case the AM will not be using the delegation
tokens created by the submission client, those tokens still need to be provided
to YARN, since they are used to do log aggregation.
To be able to re-use the code in the AMCredentialRenewal for the above
purposes, I refactored that class a bit so that it can fetch tokens into
a pre-defined UGI, insted of always logging in.
Another issue with re-attempts is that, after the fix that allows the AM
to restart correctly, new executors would get confused about when to
update credentials, because the credential updater used the update time
initially set up by the submission code. This could make the executor
fail to update credentials in time, since that value would be very out
of date in the situation described in the bug.
To fix that, I changed the YARN code to use the new RPC-based mechanism
for distributing tokens to executors. This allowed the old credential
updater code to be removed, and a lot of code in the renewer to be
simplified.
I also made two currently hardcoded values (the renewal time ratio, and
the retry wait) configurable; while this probably never needs to be set
by anyone in a production environment, it helps with testing; that's also
why they're not documented.
Tested on real cluster with a specially crafted application to test this
functionality: checked proper access to HDFS, Hive and HBase in cluster
mode with token renewal on and AM restarts. Tested things still work in
client mode too.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#20657 from vanzin/SPARK-23361.
## What changes were proposed in this pull request?
Yarn throws the following exception in cluster mode when the application is really small:
```
18/03/07 23:34:22 WARN netty.NettyRpcEnv: Ignored failure: java.util.concurrent.RejectedExecutionException: Task java.util.concurrent.ScheduledThreadPoolExecutor$ScheduledFutureTask7c974942 rejected from java.util.concurrent.ScheduledThreadPoolExecutor1eea9d2d[Terminated, pool size = 0, active threads = 0, queued tasks = 0, completed tasks = 0]
18/03/07 23:34:22 ERROR yarn.ApplicationMaster: Uncaught exception:
org.apache.spark.SparkException: Exception thrown in awaitResult:
at org.apache.spark.util.ThreadUtils$.awaitResult(ThreadUtils.scala:205)
at org.apache.spark.rpc.RpcTimeout.awaitResult(RpcTimeout.scala:75)
at org.apache.spark.rpc.RpcEndpointRef.askSync(RpcEndpointRef.scala:92)
at org.apache.spark.rpc.RpcEndpointRef.askSync(RpcEndpointRef.scala:76)
at org.apache.spark.deploy.yarn.YarnAllocator.<init>(YarnAllocator.scala:102)
at org.apache.spark.deploy.yarn.YarnRMClient.register(YarnRMClient.scala:77)
at org.apache.spark.deploy.yarn.ApplicationMaster.registerAM(ApplicationMaster.scala:450)
at org.apache.spark.deploy.yarn.ApplicationMaster.runDriver(ApplicationMaster.scala:493)
at org.apache.spark.deploy.yarn.ApplicationMaster.org$apache$spark$deploy$yarn$ApplicationMaster$$runImpl(ApplicationMaster.scala:345)
at org.apache.spark.deploy.yarn.ApplicationMaster$$anonfun$run$2.apply$mcV$sp(ApplicationMaster.scala:260)
at org.apache.spark.deploy.yarn.ApplicationMaster$$anonfun$run$2.apply(ApplicationMaster.scala:260)
at org.apache.spark.deploy.yarn.ApplicationMaster$$anonfun$run$2.apply(ApplicationMaster.scala:260)
at org.apache.spark.deploy.yarn.ApplicationMaster$$anon$5.run(ApplicationMaster.scala:810)
at java.security.AccessController.doPrivileged(Native Method)
at javax.security.auth.Subject.doAs(Subject.java:422)
at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1920)
at org.apache.spark.deploy.yarn.ApplicationMaster.doAsUser(ApplicationMaster.scala:809)
at org.apache.spark.deploy.yarn.ApplicationMaster.run(ApplicationMaster.scala:259)
at org.apache.spark.deploy.yarn.ApplicationMaster$.main(ApplicationMaster.scala:834)
at org.apache.spark.deploy.yarn.ApplicationMaster.main(ApplicationMaster.scala)
Caused by: org.apache.spark.rpc.RpcEnvStoppedException: RpcEnv already stopped.
at org.apache.spark.rpc.netty.Dispatcher.postMessage(Dispatcher.scala:158)
at org.apache.spark.rpc.netty.Dispatcher.postLocalMessage(Dispatcher.scala:135)
at org.apache.spark.rpc.netty.NettyRpcEnv.ask(NettyRpcEnv.scala:229)
at org.apache.spark.rpc.netty.NettyRpcEndpointRef.ask(NettyRpcEnv.scala:523)
at org.apache.spark.rpc.RpcEndpointRef.askSync(RpcEndpointRef.scala:91)
... 17 more
18/03/07 23:34:22 INFO yarn.ApplicationMaster: Final app status: FAILED, exitCode: 13, (reason: Uncaught exception: org.apache.spark.SparkException: Exception thrown in awaitResult: )
```
Example application:
```
object ExampleApp {
def main(args: Array[String]): Unit = {
val conf = new SparkConf().setAppName("ExampleApp")
val sc = new SparkContext(conf)
try {
// Do nothing
} finally {
sc.stop()
}
}
```
This PR pauses user class thread after `SparkContext` created and keeps it so until application master initialises properly.
## How was this patch tested?
Automated: Existing unit tests
Manual: Application submitted into small cluster
Author: Gabor Somogyi <gabor.g.somogyi@gmail.com>
Closes#20807 from gaborgsomogyi/SPARK-23660.
## What changes were proposed in this pull request?
In the current Spark on YARN code, AM always will copy and overwrite its env variables to executors, so we cannot set different values for executors.
To reproduce issue, user could start spark-shell like:
```
./bin/spark-shell --master yarn-client --conf spark.executorEnv.SPARK_ABC=executor_val --conf spark.yarn.appMasterEnv.SPARK_ABC=am_val
```
Then check executor env variables by
```
sc.parallelize(1 to 1).flatMap \{ i => sys.env.toSeq }.collect.foreach(println)
```
We will always get `am_val` instead of `executor_val`. So we should not let AM to overwrite specifically set executor env variables.
## How was this patch tested?
Added UT and tested in local cluster.
Author: jerryshao <sshao@hortonworks.com>
Closes#20799 from jerryshao/SPARK-23635.
This change restores functionality that was inadvertently removed as part
of the fix for SPARK-22372.
Also modified an existing unit test to make sure the feature works as intended.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#20776 from vanzin/SPARK-23630.
## What changes were proposed in this pull request?
Include stacktrace in the diagnostics message upon abnormal unregister from RM
## How was this patch tested?
Tested with a failing job, and confirmed a stacktrace in the client output and YARN webUI.
Author: Gera Shegalov <gera@apache.org>
Closes#20470 from gerashegalov/gera/stacktrace-diagnostics.
Third time is the charm?
There was still a race that was left in previous attempts. If the handle
closes the connection, the close() implementation would clean up state
that would prevent the thread from waiting on the connection thread to
finish. That could cause the race causing the test flakiness reported
in the bug.
The fix is to move the "wait for connection thread" code to a separate
close method that is used by the handle; that also simplifies the code
a bit and makes it also easier to follow.
I included an unrelated, but correct, change to a YARN test so that
it triggers when the PR is built.
Tested by inserting a sleep in the connection thread to mimic the race;
test failed reliably with the sleep, passes now. (Sleep not included in
the patch.) Also ran YARN tests to make sure.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#20388 from vanzin/SPARK-23020.
The race in the code is because the handle might update
its state to the wrong state if the connection handling
thread is still processing incoming data; so the handle
needs to wait for the connection to finish up before
checking the final state.
The race in the test is because when waiting for a handle
to reach a final state, the waitFor() method needs to wait
until all handle state is updated (which also includes
waiting for the connection thread above to finish).
Otherwise, waitFor() may return too early, which would cause
a bunch of different races (like the listener not being yet
notified of the state change, or being in the middle of
being notified, or the handle not being properly disposed
and causing postChecks() to assert).
On top of that I found, by code inspection, a couple of
potential races that could make a handle end up in the
wrong state when being killed.
The original version of this fix introduced the flipped
version of the first race described above; the connection
closing might override the handle state before the
handle might have a chance to do cleanup. The fix there
is to only dispose of the handle from the connection
when there is an error, and let the handle dispose
itself in the normal case.
The fix also caused a bug in YarnClusterSuite to be surfaced;
the code was checking for a file in the classpath that was
not expected to be there in client mode. Because of the above
issues, the error was not propagating correctly and the (buggy)
test was incorrectly passing.
Tested by running the existing unit tests a lot (and not
seeing the errors I was seeing before).
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#20297 from vanzin/SPARK-23020.
## What changes were proposed in this pull request?
This patch bumps the master branch version to `2.4.0-SNAPSHOT`.
## How was this patch tested?
N/A
Author: gatorsmile <gatorsmile@gmail.com>
Closes#20222 from gatorsmile/bump24.
## What changes were proposed in this pull request?
Two filesystems comparing does not consider the authority of URI. This is specific for
WASB file storage system, where userInfo is honored to differentiate filesystems.
For example: wasbs://user1xyz.net, wasbs://user2xyz.net would consider as two filesystem.
Therefore, we have to add the authority to compare two filesystem, and two filesystem with different authority can not be the same FS.
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: Mingjie Tang <mtang@hortonworks.com>
Closes#19885 from merlintang/EAR-7377.
## What changes were proposed in this pull request?
Register spark.history.ui.port as a known spark conf to be used in substitution expressions even if it's not set explicitly.
## How was this patch tested?
Added unit test to demonstrate the issue
Author: Gera Shegalov <gera@apache.org>
Author: Gera Shegalov <gshegalov@salesforce.com>
Closes#20098 from gerashegalov/gera/register-SHS-port-conf.
This change adds a new launcher that allows applications to be run
in a separate thread in the same process as the calling code. To
achieve that, some code from the child process implementation was
moved to abstract classes that implement the common functionality,
and the new launcher inherits from those.
The new launcher was added as a new class, instead of implemented
as a new option to the existing SparkLauncher, to avoid ambigous
APIs. For example, SparkLauncher has ways to set the child app's
environment, modify SPARK_HOME, or control the logging of the
child process, none of which apply to in-process apps.
The in-process launcher has limitations: it needs Spark in the
context class loader of the calling thread, and it's bound by
Spark's current limitation of a single client-mode application
per JVM. It also relies on the recently added SparkApplication
trait to make sure different apps don't mess up each other's
configuration, so config isolation is currently limited to cluster mode.
I also chose to keep the same socket-based communication for in-process
apps, even though it might be possible to avoid it for in-process
mode. That helps both implementations share more code.
Tested with new and existing unit tests, and with a simple app that
uses the launcher; also made sure the app ran fine with older launcher
jar to check binary compatibility.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#19591 from vanzin/SPARK-11035.
This PR contains implementation of the basic submission client for the cluster mode of Spark on Kubernetes. It's step 2 from the step-wise plan documented [here](https://github.com/apache-spark-on-k8s/spark/issues/441#issuecomment-330802935).
This addition is covered by the [SPIP](http://apache-spark-developers-list.1001551.n3.nabble.com/SPIP-Spark-on-Kubernetes-td22147.html) vote which passed on Aug 31.
This PR and #19468 together form a MVP of Spark on Kubernetes that allows users to run Spark applications that use resources locally within the driver and executor containers on Kubernetes 1.6 and up. Some changes on pom and build/test setup are copied over from #19468 to make this PR self contained and testable.
The submission client is mainly responsible for creating the Kubernetes pod that runs the Spark driver. It follows a step-based approach to construct the driver pod, as the code under the `submit.steps` package shows. The steps are orchestrated by `DriverConfigurationStepsOrchestrator`. `Client` creates the driver pod and waits for the application to complete if it's configured to do so, which is the case by default.
This PR also contains Dockerfiles of the driver and executor images. They are included because some of the environment variables set in the code would not make sense without referring to the Dockerfiles.
* The patch contains unit tests which are passing.
* Manual testing: ./build/mvn -Pkubernetes clean package succeeded.
* It is a subset of the entire changelist hosted at http://github.com/apache-spark-on-k8s/spark which is in active use in several organizations.
* There is integration testing enabled in the fork currently hosted by PepperData which is being moved over to RiseLAB CI.
* Detailed documentation on trying out the patch in its entirety is in: https://apache-spark-on-k8s.github.io/userdocs/running-on-kubernetes.html
cc rxin felixcheung mateiz (shepherd)
k8s-big-data SIG members & contributors: mccheah foxish ash211 ssuchter varunkatta kimoonkim erikerlandson tnachen ifilonenko liyinan926
Author: Yinan Li <liyinan926@gmail.com>
Closes#19717 from liyinan926/spark-kubernetes-4.
## What changes were proposed in this pull request?
I see the two instances where the exception is occurring.
**Instance 1:**
```
17/11/10 15:49:32 ERROR util.Utils: Uncaught exception in thread driver-revive-thread
org.apache.spark.SparkException: Could not find CoarseGrainedScheduler.
at org.apache.spark.rpc.netty.Dispatcher.postMessage(Dispatcher.scala:160)
at org.apache.spark.rpc.netty.Dispatcher.postOneWayMessage(Dispatcher.scala:140)
at org.apache.spark.rpc.netty.NettyRpcEnv.send(NettyRpcEnv.scala:187)
at org.apache.spark.rpc.netty.NettyRpcEndpointRef.send(NettyRpcEnv.scala:521)
at org.apache.spark.scheduler.cluster.CoarseGrainedSchedulerBackend$DriverEndpoint$$anon$1$$anonfun$run$1$$anonfun$apply$mcV$sp$1.apply(CoarseGrainedSchedulerBackend.scala:125)
at org.apache.spark.scheduler.cluster.CoarseGrainedSchedulerBackend$DriverEndpoint$$anon$1$$anonfun$run$1$$anonfun$apply$mcV$sp$1.apply(CoarseGrainedSchedulerBackend.scala:125)
at scala.Option.foreach(Option.scala:257)
at org.apache.spark.scheduler.cluster.CoarseGrainedSchedulerBackend$DriverEndpoint$$anon$1$$anonfun$run$1.apply$mcV$sp(CoarseGrainedSchedulerBackend.scala:125)
at org.apache.spark.util.Utils$.tryLogNonFatalError(Utils.scala:1344)
at org.apache.spark.scheduler.cluster.CoarseGrainedSchedulerBackend$DriverEndpoint$$anon$1.run(CoarseGrainedSchedulerBackend.scala:124)
at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511)
at java.util.concurrent.FutureTask.runAndReset(FutureTask.java:308)
at java.util.concurrent.ScheduledThreadPoolExecutor$ScheduledFutureTask.access$301(ScheduledThreadPoolExecutor.java:180)
at java.util.concurrent.ScheduledThreadPoolExecutor$ScheduledFutureTask.run(ScheduledThreadPoolExecutor.java:294)
at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
at java.lang.Thread.run(Thread.java:745)
```
In CoarseGrainedSchedulerBackend.scala, driver-revive-thread starts with DriverEndpoint.onStart() and keeps sending the ReviveOffers messages periodically till it gets shutdown as part DriverEndpoint.onStop(). There is no proper coordination between the driver-revive-thread(shutdown) and the RpcEndpoint unregister, RpcEndpoint unregister happens first and then driver-revive-thread shuts down as part of DriverEndpoint.onStop(), In-between driver-revive-thread may try to send the ReviveOffers message which is leading to the above exception.
To fix this issue, this PR moves the shutting down of driver-revive-thread to CoarseGrainedSchedulerBackend.stop() which executes before the DriverEndpoint unregister.
**Instance 2:**
```
17/11/10 16:31:38 ERROR cluster.YarnSchedulerBackend$YarnSchedulerEndpoint: Error requesting driver to remove executor 1 for reason Executor for container container_1508535467865_0226_01_000002 exited because of a YARN event (e.g., pre-emption) and not because of an error in the running job.
org.apache.spark.SparkException: Could not find CoarseGrainedScheduler.
at org.apache.spark.rpc.netty.Dispatcher.postMessage(Dispatcher.scala:160)
at org.apache.spark.rpc.netty.Dispatcher.postLocalMessage(Dispatcher.scala:135)
at org.apache.spark.rpc.netty.NettyRpcEnv.ask(NettyRpcEnv.scala:229)
at org.apache.spark.rpc.netty.NettyRpcEndpointRef.ask(NettyRpcEnv.scala:516)
at org.apache.spark.rpc.RpcEndpointRef.ask(RpcEndpointRef.scala:63)
at org.apache.spark.scheduler.cluster.YarnSchedulerBackend$YarnSchedulerEndpoint$$anonfun$receive$1.applyOrElse(YarnSchedulerBackend.scala:269)
at org.apache.spark.rpc.netty.Inbox$$anonfun$process$1.apply$mcV$sp(Inbox.scala:117)
at org.apache.spark.rpc.netty.Inbox.safelyCall(Inbox.scala:205)
at org.apache.spark.rpc.netty.Inbox.process(Inbox.scala:101)
at org.apache.spark.rpc.netty.Dispatcher$MessageLoop.run(Dispatcher.scala:221)
at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
at java.lang.Thread.run(Thread.java:745)
```
Here YarnDriverEndpoint tries to send remove executor messages after the Yarn scheduler backend service stop, which is leading to the above exception. To avoid the above exception,
1) We may add a condition(which checks whether service has stopped or not) before sending executor remove message
2) Add a warn log message in onFailure case when the service is already stopped
In this PR, chosen the 2) option which adds a log message in the case of onFailure without the exception stack trace since the option 1) would need to to go through for every remove executor message.
## How was this patch tested?
I verified it manually, I don't see these exceptions with the PR changes.
Author: Devaraj K <devaraj@apache.org>
Closes#19741 from devaraj-kavali/SPARK-14228.
The main goal of this change is to allow multiple cluster-mode
submissions from the same JVM, without having them end up with
mixed configuration. That is done by extending the SparkApplication
trait, and doing so was reasonably trivial for standalone and
mesos modes.
For YARN mode, there was a complication. YARN used a "SPARK_YARN_MODE"
system property to control behavior indirectly in a whole bunch of
places, mainly in the SparkHadoopUtil / YarnSparkHadoopUtil classes.
Most of the changes here are removing that.
Since we removed support for Hadoop 1.x, some methods that lived in
YarnSparkHadoopUtil can now live in SparkHadoopUtil. The remaining
methods don't need to be part of the class, and can be called directly
from the YarnSparkHadoopUtil object, so now there's a single
implementation of SparkHadoopUtil.
There were two places in the code that relied on SPARK_YARN_MODE to
make decisions about YARN-specific functionality, and now explicitly check
the master from the configuration for that instead:
* fetching the external shuffle service port, which can come from the YARN
configuration.
* propagation of the authentication secret using Hadoop credentials. This also
was cleaned up a little to not need so many methods in `SparkHadoopUtil`.
With those out of the way, actually changing the YARN client
to extend SparkApplication was easy.
Tested with existing unit tests, and also by running YARN apps
with auth and kerberos both on and off in a real cluster.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#19631 from vanzin/SPARK-22372.
## What changes were proposed in this pull request?
This is a stripped down version of the `KubernetesClusterSchedulerBackend` for Spark with the following components:
- Static Allocation of Executors
- Executor Pod Factory
- Executor Recovery Semantics
It's step 1 from the step-wise plan documented [here](https://github.com/apache-spark-on-k8s/spark/issues/441#issuecomment-330802935).
This addition is covered by the [SPIP vote](http://apache-spark-developers-list.1001551.n3.nabble.com/SPIP-Spark-on-Kubernetes-td22147.html) which passed on Aug 31 .
## How was this patch tested?
- The patch contains unit tests which are passing.
- Manual testing: `./build/mvn -Pkubernetes clean package` succeeded.
- It is a **subset** of the entire changelist hosted in http://github.com/apache-spark-on-k8s/spark which is in active use in several organizations.
- There is integration testing enabled in the fork currently [hosted by PepperData](spark-k8s-jenkins.pepperdata.org:8080) which is being moved over to RiseLAB CI.
- Detailed documentation on trying out the patch in its entirety is in: https://apache-spark-on-k8s.github.io/userdocs/running-on-kubernetes.html
cc rxin felixcheung mateiz (shepherd)
k8s-big-data SIG members & contributors: mccheah ash211 ssuchter varunkatta kimoonkim erikerlandson liyinan926 tnachen ifilonenko
Author: Yinan Li <liyinan926@gmail.com>
Author: foxish <ramanathana@google.com>
Author: mcheah <mcheah@palantir.com>
Closes#19468 from foxish/spark-kubernetes-3.
## What changes were proposed in this pull request?
tl;dr: Add a class, `MesosHadoopDelegationTokenManager` that updates delegation tokens on a schedule on the behalf of Spark Drivers. Broadcast renewed credentials to the executors.
## The problem
We recently added Kerberos support to Mesos-based Spark jobs as well as Secrets support to the Mesos Dispatcher (SPARK-16742, SPARK-20812, respectively). However the delegation tokens have a defined expiration. This poses a problem for long running Spark jobs (e.g. Spark Streaming applications). YARN has a solution for this where a thread is scheduled to renew the tokens they reach 75% of their way to expiration. It then writes the tokens to HDFS for the executors to find (uses a monotonically increasing suffix).
## This solution
We replace the current method in `CoarseGrainedSchedulerBackend` which used to discard the token renewal time with a protected method `fetchHadoopDelegationTokens`. Now the individual cluster backends are responsible for overriding this method to fetch and manage token renewal. The delegation tokens themselves, are still part of the `CoarseGrainedSchedulerBackend` as before.
In the case of Mesos renewed Credentials are broadcasted to the executors. This maintains all transfer of Credentials within Spark (as opposed to Spark-to-HDFS). It also does not require any writing of Credentials to disk. It also does not require any GC of old files.
## How was this patch tested?
Manually against a Kerberized HDFS cluster.
Thank you for the reviews.
Author: ArtRand <arand@soe.ucsc.edu>
Closes#19272 from ArtRand/spark-21842-450-kerberos-ticket-renewal.
## What changes were proposed in this pull request?
Removed the unnecessary stagingDirPath null check in ApplicationMaster.cleanupStagingDir().
## How was this patch tested?
I verified with the existing test cases.
Author: Devaraj K <devaraj@apache.org>
Closes#19749 from devaraj-kavali/SPARK-22519.
## What changes were proposed in this pull request?
When I ran self contained sql apps, such as
```scala
import org.apache.spark.sql.SparkSession
object ShowHiveTables {
def main(args: Array[String]): Unit = {
val spark = SparkSession
.builder()
.appName("Show Hive Tables")
.enableHiveSupport()
.getOrCreate()
spark.sql("show tables").show()
spark.stop()
}
}
```
with **yarn cluster** mode and `hive-site.xml` correctly within `$SPARK_HOME/conf`,they failed to connect the right hive metestore for not seeing hive-site.xml in AM/Driver's classpath.
Although submitting them with `--files/--jars local/path/to/hive-site.xml` or puting it to `$HADOOP_CONF_DIR/YARN_CONF_DIR` can make these apps works well in cluster mode as client mode, according to the official doc, see http://spark.apache.org/docs/latest/sql-programming-guide.html#hive-tables
> Configuration of Hive is done by placing your hive-site.xml, core-site.xml (for security configuration), and hdfs-site.xml (for HDFS configuration) file in conf/.
We may respect these configuration files too or modify the doc for hive-tables in cluster mode.
## How was this patch tested?
cc cloud-fan gatorsmile
Author: Kent Yao <yaooqinn@hotmail.com>
Closes#19663 from yaooqinn/SPARK-21888.
The bug was introduced in SPARK-22290, which changed how the app's user
is impersonated in the AM. The changed missed an initialization function
that needs to be run as the app owner (who has the right credentials to
read from HDFS).
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#19566 from vanzin/SPARK-22341.
Hive delegation tokens are only needed when the Spark driver has no access
to the kerberos TGT. That happens only in two situations:
- when using a proxy user
- when using cluster mode without a keytab
This change modifies the Hive provider so that it only generates delegation
tokens in those situations, and tweaks the YARN AM so that it makes the proper
user visible to the Hive code when running with keytabs, so that the TGT
can be used instead of a delegation token.
The effect of this change is that now it's possible to initialize multiple,
non-concurrent SparkContext instances in the same JVM. Before, the second
invocation would fail to fetch a new Hive delegation token, which then could
make the second (or third or...) application fail once the token expired.
With this change, the TGT will be used to authenticate to the HMS instead.
This change also avoids polluting the current logged in user's credentials
when launching applications. The credentials are copied only when running
applications as a proxy user. This makes it possible to implement SPARK-11035
later, where multiple threads might be launching applications, and each app
should have its own set of credentials.
Tested by verifying HDFS and Hive access in following scenarios:
- client and cluster mode
- client and cluster mode with proxy user
- client and cluster mode with principal / keytab
- long-running cluster app with principal / keytab
- pyspark app that creates (and stops) multiple SparkContext instances
through its lifetime
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#19509 from vanzin/SPARK-22290.
## What changes were proposed in this pull request?
When the libraries temp directory(i.e. __spark_libs__*.zip dir) file system and staging dir(destination) file systems are the same then the __spark_libs__*.zip is not copying to the staging directory. But after making this decision the libraries zip file is getting deleted immediately and becoming unavailable for the Node Manager's localization.
With this change, client copies the files to remote always when the source scheme is "file".
## How was this patch tested?
I have verified it manually in yarn/cluster and yarn/client modes with hdfs and local file systems.
Author: Devaraj K <devaraj@apache.org>
Closes#19141 from devaraj-kavali/SPARK-21384.
The live listener bus now cleans up after itself and releases listeners
after stopping, so code cannot get references to listeners after the
Spark context is stopped.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#19297 from vanzin/SPARK-18838.hotfix.
…build; fix some things that will be warnings or errors in 2.12; restore Scala 2.12 profile infrastructure
## What changes were proposed in this pull request?
This change adds back the infrastructure for a Scala 2.12 build, but does not enable it in the release or Python test scripts.
In order to make that meaningful, it also resolves compile errors that the code hits in 2.12 only, in a way that still works with 2.11.
It also updates dependencies to the earliest minor release of dependencies whose current version does not yet support Scala 2.12. This is in a sense covered by other JIRAs under the main umbrella, but implemented here. The versions below still work with 2.11, and are the _latest_ maintenance release in the _earliest_ viable minor release.
- Scalatest 2.x -> 3.0.3
- Chill 0.8.0 -> 0.8.4
- Clapper 1.0.x -> 1.1.2
- json4s 3.2.x -> 3.4.2
- Jackson 2.6.x -> 2.7.9 (required by json4s)
This change does _not_ fully enable a Scala 2.12 build:
- It will also require dropping support for Kafka before 0.10. Easy enough, just didn't do it yet here
- It will require recreating `SparkILoop` and `Main` for REPL 2.12, which is SPARK-14650. Possible to do here too.
What it does do is make changes that resolve much of the remaining gap without affecting the current 2.11 build.
## How was this patch tested?
Existing tests and build. Manually tested with `./dev/change-scala-version.sh 2.12` to verify it compiles, modulo the exceptions above.
Author: Sean Owen <sowen@cloudera.com>
Closes#18645 from srowen/SPARK-14280.
In the current code, if NM recovery is not enabled then `YarnShuffleService` will write shuffle metadata to NM local dir-1, if this local dir-1 is on bad disk, then `YarnShuffleService` will be failed to start. So to solve this issue, in Spark side if NM recovery is not enabled, then Spark will not persist data into leveldb, in that case yarn shuffle service can still be served but lose the ability for recovery, (it is fine because the failure of NM will kill the containers as well as applications).
Tested in the local cluster with NM recovery off and on to see if folder is created or not. MiniCluster UT isn't added because in MiniCluster NM will always set port to 0, but NM recovery requires non-ephemeral port.
Author: jerryshao <sshao@hortonworks.com>
Closes#19032 from jerryshao/SPARK-17321.
Change-Id: I8f2fe73d175e2ad2c4e380caede3873e0192d027
## What changes were proposed in this pull request?
Fix typos
## How was this patch tested?
Existing tests
Author: Andrew Ash <andrew@andrewash.com>
Closes#18996 from ash211/patch-2.
## What changes were proposed in this pull request?
Add Kerberos Support to Mesos. This includes kinit and --keytab support, but does not include delegation token renewal.
## How was this patch tested?
Manually against a Secure DC/OS Apache HDFS cluster.
Author: ArtRand <arand@soe.ucsc.edu>
Author: Michael Gummelt <mgummelt@mesosphere.io>
Closes#18519 from mgummelt/SPARK-16742-kerberos.
This version fixes a few issues in the import order checker; it provides
better error messages, and detects more improper ordering (thus the need
to change a lot of files in this patch). The main fix is that it correctly
complains about the order of packages vs. classes.
As part of the above, I moved some "SparkSession" import in ML examples
inside the "$example on$" blocks; that didn't seem consistent across
different source files to start with, and avoids having to add more on/off blocks
around specific imports.
The new scalastyle also seems to have a better header detector, so a few
license headers had to be updated to match the expected indentation.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#18943 from vanzin/SPARK-21731.
The main goal of this change is to avoid the situation described
in the bug, where an AM restart in the middle of a job may cause
no new executors to be allocated because of faulty logic in the
reset path.
The change does two things:
- fixes the executor alloc manager's reset() so that it does not
stop allocation after a reset() in the middle of a job
- re-orders the initialization of the YarnAllocator class so that
it fetches the current executor ID before triggering the reset()
above.
This ensures both that the new allocator gets new requests for executors,
and that it starts from the correct executor id.
Tested with unit tests and by manually causing AM restarts while
running jobs using spark-shell in YARN mode.
Closes#17882
Author: Marcelo Vanzin <vanzin@cloudera.com>
Author: Guoqiang Li <witgo@qq.com>
Closes#18663 from vanzin/SPARK-20079.
If you run a spark job without creating the SparkSession or SparkContext, the spark job logs says it succeeded but yarn says it fails and retries 3 times. Also, since, Application Master unregisters with Resource Manager and exits successfully, it deletes the spark staging directory, so when yarn makes subsequent retries, it fails to find the staging directory and thus, the retries fail.
Added a flag to check whether user has initialized SparkContext. If it is true, we let Application Master unregister with Resource Manager else, we do not let AM unregister with RM.
## How was this patch tested?
Manually tested the fix.
Before:
<img width="1253" alt="screen shot-before" src="https://user-images.githubusercontent.com/22228190/28647214-69bf81e2-722b-11e7-9ed0-d416d2bf23be.png">
After:
<img width="1319" alt="screen shot-after" src="https://user-images.githubusercontent.com/22228190/28647220-70f9eea2-722b-11e7-85c6-e56276b15614.png">
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: pgandhi <pgandhi@yahoo-inc.com>
Author: pgandhi999 <parthkgandhi9@gmail.com>
Closes#18741 from pgandhi999/SPARK-21541.
There was some code based on the old SASL handler in the new auth client that
was incorrectly using the SASL user as the user to authenticate against the
external shuffle service. This caused the external service to not be able to
find the correct secret to authenticate the connection, failing the connection.
In the course of debugging, I found that some log messages from the YARN shuffle
service were a little noisy, so I silenced some of them, and also added a couple
of new ones that helped find this issue. On top of that, I found that a check
in the code that records app secrets was wrong, causing more log spam and also
using an O(n) operation instead of an O(1) call.
Also added a new integration suite for the YARN shuffle service with auth on,
and verified it failed before, and passes now.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#18706 from vanzin/SPARK-21494.
When NodeManagers launching Executors,
the `missing` value will exceed the
real value when the launch is slow, this can lead to YARN allocates more resource.
We add the `numExecutorsRunning` when calculate the `missing` to avoid this.
Test by experiment.
Author: DjvuLee <lihu@bytedance.com>
Closes#18651 from djvulee/YarnAllocate.
Instead of using the host's cpu count, use the number of cores allocated
for the Spark process when sizing the RPC dispatch thread pool. This avoids
creating large thread pools on large machines when the number of allocated
cores is small.
Tested by verifying number of threads with spark.executor.cores set
to 1 and 4; same thing for YARN AM.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#18639 from vanzin/SPARK-21408.
## What changes were proposed in this pull request?
In the current `YARNHadoopDelegationTokenManager`, `FileSystem` to which to get tokens are created out of KDC logged UGI, using these `FileSystem` to get new tokens will lead to exception. The main thing is that Spark code trying to get new tokens from the FS created with token auth-ed UGI, but Hadoop can only grant new tokens in kerberized UGI. To fix this issue, we should lazily create these FileSystem within KDC logged UGI.
## How was this patch tested?
Manual verification in secure cluster.
CC vanzin mgummelt please help to review, thanks!
Author: jerryshao <sshao@hortonworks.com>
Closes#18633 from jerryshao/SPARK-21411.
## What changes were proposed in this pull request?
In this issue we have a long running Spark application with secure HBase, which requires `HBaseCredentialProvider` to get tokens periodically, we specify HBase related jars with `--packages`, but these dependencies are not added into AM classpath, so when `HBaseCredentialProvider` tries to initialize HBase connections to get tokens, it will be failed.
Currently because jars specified with `--jars` or `--packages` are not added into AM classpath, the only way to extend AM classpath is to use "spark.driver.extraClassPath" which supposed to be used in yarn cluster mode.
So in this fix, we proposed to use/reuse a classloader for `AMCredentialRenewer` to acquire new tokens.
Also in this patch, we fixed AM cannot get tokens from HDFS issue, it is because FileSystem is gotten before kerberos logged, so using this FS to get tokens will throw exception.
## How was this patch tested?
Manual verification.
Author: jerryshao <sshao@hortonworks.com>
Closes#18616 from jerryshao/SPARK-21377.
## What changes were proposed in this pull request?
The current code is very verbose on shutdown.
The changes I propose is to change the log level when the driver is shutting down and the RPC connections are closed (RpcEnvStoppedException).
## How was this patch tested?
Tested with word count(deploy-mode = cluster, master = yarn, num-executors = 4) with 300GB of data.
Author: John Lee <jlee2@yahoo-inc.com>
Closes#18547 from yoonlee95/SPARK-21321.
When localizing the gateway config files in a YARN application, avoid
overwriting final configs by distributing the gateway files to a separate
directory, and explicitly loading them into the Hadoop config, instead
of placing those files before the cluster's files in the classpath.
This is done by saving the gateway's config to a separate XML file
distributed with the rest of the Spark app's config, and loading that
file when creating a new config through `YarnSparkHadoopUtil`.
Tested with existing unit tests, and by verifying the behavior in a YARN
cluster (final values are not overridden, non-final values are).
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#18370 from vanzin/SPARK-9825.
## What changes were proposed in this pull request?
This issue happens in long running application with yarn cluster mode, because yarn#client doesn't sync token with AM, so it will always keep the initial token, this token may be expired in the long running scenario, so when yarn#client tries to clean up staging directory after application finished, it will use this expired token and meet token expire issue.
## How was this patch tested?
Manual verification is secure cluster.
Author: jerryshao <sshao@hortonworks.com>
Closes#18617 from jerryshao/SPARK-21376.
Currently the code monitoring the launch of the client AM uses the value of
spark.yarn.report.interval as the interval for polling the RM; if someone
has that value to a really large interval, it would take that long to detect
that the client AM has started, which is not expected.
Instead, have a separate config for the interval to use when the client AM is
starting. The other config is still used in cluster mode, and to detect the
status of the client AM after it is already running.
Tested by running client and cluster mode apps with a modified value of
spark.yarn.report.interval, verifying client AM launch is detected before
that interval elapses.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#18380 from vanzin/SPARK-16019.
## What changes were proposed in this pull request?
This PR aims to bump Py4J in order to fix the following float/double bug.
Py4J 0.10.5 fixes this (https://github.com/bartdag/py4j/issues/272) and the latest Py4J is 0.10.6.
**BEFORE**
```
>>> df = spark.range(1)
>>> df.select(df['id'] + 17.133574204226083).show()
+--------------------+
|(id + 17.1335742042)|
+--------------------+
| 17.1335742042|
+--------------------+
```
**AFTER**
```
>>> df = spark.range(1)
>>> df.select(df['id'] + 17.133574204226083).show()
+-------------------------+
|(id + 17.133574204226083)|
+-------------------------+
| 17.133574204226083|
+-------------------------+
```
## How was this patch tested?
Manual.
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#18546 from dongjoon-hyun/SPARK-21278.
## What changes were proposed in this pull request?
When I set different clusters for "spark.hadoop.fs.defaultFS" and "spark.yarn.stagingDir" as follows：
```
spark.hadoop.fs.defaultFS hdfs://tl-nn-tdw.tencent-distribute.com:54310
spark.yarn.stagingDir hdfs://ss-teg-2-v2/tmp/spark
```
The staging dir can not be deleted, it will prompt following message:
```
java.lang.IllegalArgumentException: Wrong FS: hdfs://ss-teg-2-v2/tmp/spark/.sparkStaging/application_1496819138021_77618, expected: hdfs://tl-nn-tdw.tencent-distribute.com:54310
```
## How was this patch tested?
Existing tests
Author: sharkdtu <sharkdtu@tencent.com>
Closes#18352 from sharkdtu/master.
## What changes were proposed in this pull request?
Move Hadoop delegation token code from `spark-yarn` to `spark-core`, so that other schedulers (such as Mesos), may use it. In order to avoid exposing Hadoop interfaces in spark-core, the new Hadoop delegation token classes are kept private. In order to provider backward compatiblity, and to allow YARN users to continue to load their own delegation token providers via Java service loading, the old YARN interfaces, as well as the client code that uses them, have been retained.
Summary:
- Move registered `yarn.security.ServiceCredentialProvider` classes from `spark-yarn` to `spark-core`. Moved them into a new, private hierarchy under `HadoopDelegationTokenProvider`. Client code in `HadoopDelegationTokenManager` now loads credentials from a whitelist of three providers (`HadoopFSDelegationTokenProvider`, `HiveDelegationTokenProvider`, `HBaseDelegationTokenProvider`), instead of service loading, which means that users are not able to implement their own delegation token providers, as they are in the `spark-yarn` module.
- The `yarn.security.ServiceCredentialProvider` interface has been kept for backwards compatibility, and to continue to allow YARN users to implement their own delegation token provider implementations. Client code in YARN now fetches tokens via the new `YARNHadoopDelegationTokenManager` class, which fetches tokens from the core providers through `HadoopDelegationTokenManager`, as well as service loads them from `yarn.security.ServiceCredentialProvider`.
Old Hierarchy:
```
yarn.security.ServiceCredentialProvider (service loaded)
HadoopFSCredentialProvider
HiveCredentialProvider
HBaseCredentialProvider
yarn.security.ConfigurableCredentialManager
```
New Hierarchy:
```
HadoopDelegationTokenManager
HadoopDelegationTokenProvider (not service loaded)
HadoopFSDelegationTokenProvider
HiveDelegationTokenProvider
HBaseDelegationTokenProvider
yarn.security.ServiceCredentialProvider (service loaded)
yarn.security.YARNHadoopDelegationTokenManager
```
## How was this patch tested?
unit tests
Author: Michael Gummelt <mgummelt@mesosphere.io>
Author: Dr. Stefan Schimanski <sttts@mesosphere.io>
Closes#17723 from mgummelt/SPARK-20434-refactor-kerberos.
In Spark on YARN, when configuring "spark.yarn.jars" with local jars (jars started with "local" scheme), we will get inaccurate classpath for AM and containers. This is because we don't remove "local" scheme when concatenating classpath. It is OK to run because classpath is separated with ":" and java treat "local" as a separate jar. But we could improve it to remove the scheme.
Updated `ClientSuite` to check "local" is not in the classpath.
cc jerryshao
Author: Li Yichao <lyc@zhihu.com>
Author: Li Yichao <liyichao.good@gmail.com>
Closes#18129 from liyichao/SPARK-20365.
## What changes were proposed in this pull request?
Deleted generated JARs archive after distribution to HDFS
## How was this patch tested?
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: Lior Regev <lioregev@gmail.com>
Closes#17986 from liorregev/master.
## What changes were proposed in this pull request?
After SPARK-10997, client mode Netty RpcEnv doesn't require to start server, so port configurations are not used any more, here propose to remove these two configurations: "spark.executor.port" and "spark.am.port".
## How was this patch tested?
Existing UTs.
Author: jerryshao <sshao@hortonworks.com>
Closes#17866 from jerryshao/SPARK-20605.
## What changes were proposed in this pull request?
Currently, `spark.executor.instances` is deprecated in `spark-env.sh`, because we suggest config it in `spark-defaults.conf` or other config file. And also this parameter is useless even if you set it in `spark-env.sh`, so remove it in this patch.
## How was this patch tested?
Existing tests.
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: Xianyang Liu <xianyang.liu@intel.com>
Closes#17881 from ConeyLiu/deprecatedParam.
