### 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?
Document the resource scheduling feature - https://issues.apache.org/jira/browse/SPARK-24615
Add general docs, yarn, kubernetes, and standalone cluster specific ones.
### Why are the changes needed?
Help users understand the feature
### Does this PR introduce any user-facing change?
docs
### How was this patch tested?
N/A
Closes#25698 from tgravescs/SPARK-27492-gpu-sched-docs.
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 reverts the logic which was introduced as a part of SPARK-24149 and a subsequent followup PR.
With existing logic:
- Spark fails to launch with HDFS federation enabled while trying to get a path to a logical nameservice.
- It gets tokens for unrelated namespaces if they are used in HDFS Federation
- Automatic namespace discovery is supported only if these are on the same cluster.
Rationale for change:
- For accessing data from related namespaces, viewfs should handle getting tokens for spark
- For accessing data from unrelated namespaces(user explicitly specifies them using existing configs) as these could be on the same or different cluster.
(Please fill in changes proposed in this fix)
Revert the changes.
## How was this patch tested?
Ran few manual tests and unit test.
Closes#24785 from dhruve/bug/SPARK-27937.
Authored-by: Dhruve Ashar <dhruveashar@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.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?
Add AL2 license to metadata of all .md files.
This seemed to be the tidiest way as it will get ignored by .md renderers and other tools. Attempts to write them as markdown comments revealed that there is no such standard thing.
## How was this patch tested?
Doc build
Closes#24243 from srowen/SPARK-26918.
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?
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?
`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?
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>
## What changes were proposed in this pull request?
Clarify documentation about security.
## How was this patch tested?
None, just documentation
Closes#22852 from tgravescs/SPARK-25023.
Authored-by: Thomas Graves <tgraves@thirteenroutine.corp.gq1.yahoo.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
## What changes were proposed in this pull request?
This PR adds CLI support for YARN custom resources, e.g. GPUs and any other resources YARN defines.
The custom resources are defined with Spark properties, no additional CLI arguments were introduced.
The properties can be defined in the following form:
**AM resources, client mode:**
Format: `spark.yarn.am.resource.<resource-name>`
The property name follows the naming convention of YARN AM cores / memory properties: `spark.yarn.am.memory and spark.yarn.am.cores
`
**Driver resources, cluster mode:**
Format: `spark.yarn.driver.resource.<resource-name>`
The property name follows the naming convention of driver cores / memory properties: `spark.driver.memory and spark.driver.cores.`
**Executor resources:**
Format: `spark.yarn.executor.resource.<resource-name>`
The property name follows the naming convention of executor cores / memory properties: `spark.executor.memory / spark.executor.cores`.
For the driver resources (cluster mode) and executor resources properties, we use the `yarn` prefix here as custom resource types are specific to YARN, currently.
**Validation:**
Please note that a validation logic is added to avoid having requested resources defined in 2 ways, for example defining the following configs:
```
"--conf", "spark.driver.memory=2G",
"--conf", "spark.yarn.driver.resource.memory=1G"
```
will not start execution and will print an error message.
## How was this patch tested?
Unit tests + manual execution with Hadoop2 and Hadoop 3 builds.
Testing have been performed on a real cluster with Spark and YARN configured:
Cluster and client mode
Request Resource Types with lowercase and uppercase units
Start Spark job with only requesting standard resources (mem / cpu)
Error handling cases:
- Request unknown resource type
- Request Resource type (either memory / cpu) with duplicate configs at the same time (e.g. with this config:
```
--conf spark.yarn.am.resource.memory=1G \
--conf spark.yarn.driver.resource.memory=2G \
--conf spark.yarn.executor.resource.memory=3G \
```
), ResourceTypeValidator handles these cases well, so it is not permitted
- Request standard resource (memory / cpu) with the new style configs, e.g. --conf spark.yarn.am.resource.memory=1G, this is not permitted and handled well.
An example about how I ran the testcases:
```
cd ~;export HADOOP_CONF_DIR=/opt/hadoop/etc/hadoop/;
./spark-2.4.0-SNAPSHOT-bin-custom-spark/bin/spark-submit \
--class org.apache.spark.examples.SparkPi \
--master yarn \
--deploy-mode cluster \
--driver-memory 1G \
--driver-cores 1 \
--executor-memory 1G \
--executor-cores 1 \
--conf spark.logConf=true \
--conf spark.yarn.executor.resource.gpu=3G \
--verbose \
./spark-2.4.0-SNAPSHOT-bin-custom-spark/examples/jars/spark-examples_2.11-2.4.0-SNAPSHOT.jar \
10;
```
Closes#20761 from szyszy/SPARK-20327.
Authored-by: Szilard Nemeth <snemeth@cloudera.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
Remove Hadoop 2.6 references and make 2.7 the default.
Obviously, this is for master/3.0.0 only.
After this we can also get rid of the separate test jobs for Hadoop 2.6.
## How was this patch tested?
Existing tests
Closes#22615 from srowen/SPARK-25016.
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?
SparkSubmit already logs in the user if a keytab is provided, the only issue is that it uses the existing configs which have "yarn" in their name. As such, the configs were changed to:
`spark.kerberos.keytab` and `spark.kerberos.principal`.
## How was this patch tested?
Will be tested with K8S tests, but needs to be tested with Yarn
- [x] K8S Secure HDFS tests
- [x] Yarn Secure HDFS tests vanzin
Closes#22362 from ifilonenko/SPARK-25372.
Authored-by: Ilan Filonenko <if56@cornell.edu>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
Fix a few broken links and typos, and, nit, use HTTPS more consistently esp. on scripts and Apache links
## How was this patch tested?
Doc build
Closes#22172 from srowen/DocTypo.
Authored-by: Sean Owen <sean.owen@databricks.com>
Signed-off-by: hyukjinkwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
In this PR metrics are introduced for YARN. As up to now there was no metrics in the YARN module a new metric system is created with the name "applicationMaster".
To support both client and cluster mode the metric system lifecycle is bound to the AM.
## How was this patch tested?
Both client and cluster mode was tested manually.
Before the test on one of the YARN node spark-core was removed to cause the allocation failure.
Spark was started as (in case of client mode):
```
spark2-submit \
--class org.apache.spark.examples.SparkPi \
--conf "spark.yarn.blacklist.executor.launch.blacklisting.enabled=true" --conf "spark.blacklist.application.maxFailedExecutorsPerNode=2" --conf "spark.dynamicAllocation.enabled=true" --conf "spark.metrics.conf.*.sink.console.class=org.apache.spark.metrics.sink.ConsoleSink" \
--master yarn \
--deploy-mode client \
original-spark-examples_2.11-2.4.0-SNAPSHOT.jar \
1000
```
In both cases the YARN logs contained the new metrics as:
```
$ yarn logs --applicationId application_1529926424933_0015
...
-- Gauges ----------------------------------------------------------------------
application_1531751594108_0046.applicationMaster.numContainersPendingAllocate
value = 0
application_1531751594108_0046.applicationMaster.numExecutorsFailed
value = 3
application_1531751594108_0046.applicationMaster.numExecutorsRunning
value = 9
application_1531751594108_0046.applicationMaster.numLocalityAwareTasks
value = 0
application_1531751594108_0046.applicationMaster.numReleasedContainers
value = 0
...
```
Author: “attilapiros” <piros.attila.zsolt@gmail.com>
Author: Attila Zsolt Piros <2017933+attilapiros@users.noreply.github.com>
Closes#21635 from attilapiros/SPARK-24594.
## What changes were proposed in this pull request?
In the case of getting tokens via customized `ServiceCredentialProvider`, it is required that `ServiceCredentialProvider` be available in local spark-submit process classpath. In this case, all the configured remote sources should be forced to download to local.
For the ease of using this configuration, here propose to add wildcard '*' support to `spark.yarn.dist.forceDownloadSchemes`, also clarify the usage of this configuration.
## How was this patch tested?
New UT added.
Author: jerryshao <sshao@hortonworks.com>
Closes#21633 from jerryshao/SPARK-21917-followup.
## What changes were proposed in this pull request?
This change extends YARN resource allocation handling with blacklisting functionality.
This handles cases when node is messed up or misconfigured such that a container won't launch on it. Before this change backlisting only focused on task execution but this change introduces YarnAllocatorBlacklistTracker which tracks allocation failures per host (when enabled via "spark.yarn.blacklist.executor.launch.blacklisting.enabled").
## How was this patch tested?
### With unit tests
Including a new suite: YarnAllocatorBlacklistTrackerSuite.
#### Manually
It was tested on a cluster by deleting the Spark jars on one of the node.
#### Behaviour before these changes
Starting Spark as:
```
spark2-shell --master yarn --deploy-mode client --num-executors 4 --conf spark.executor.memory=4g --conf "spark.yarn.max.executor.failures=6"
```
Log is:
```
18/04/12 06:49:36 INFO yarn.ApplicationMaster: Final app status: FAILED, exitCode: 11, (reason: Max number of executor failures (6) reached)
18/04/12 06:49:39 INFO yarn.ApplicationMaster: Unregistering ApplicationMaster with FAILED (diag message: Max number of executor failures (6) reached)
18/04/12 06:49:39 INFO impl.AMRMClientImpl: Waiting for application to be successfully unregistered.
18/04/12 06:49:39 INFO yarn.ApplicationMaster: Deleting staging directory hdfs://apiros-1.gce.test.com:8020/user/systest/.sparkStaging/application_1523459048274_0016
18/04/12 06:49:39 INFO util.ShutdownHookManager: Shutdown hook called
```
#### Behaviour after these changes
Starting Spark as:
```
spark2-shell --master yarn --deploy-mode client --num-executors 4 --conf spark.executor.memory=4g --conf "spark.yarn.max.executor.failures=6" --conf "spark.yarn.blacklist.executor.launch.blacklisting.enabled=true"
```
And the log is:
```
18/04/13 05:37:43 INFO yarn.YarnAllocator: Will request 1 executor container(s), each with 1 core(s) and 4505 MB memory (including 409 MB of overhead)
18/04/13 05:37:43 INFO yarn.YarnAllocator: Submitted 1 unlocalized container requests.
18/04/13 05:37:43 INFO yarn.YarnAllocator: Launching container container_1523459048274_0025_01_000008 on host apiros-4.gce.test.com for executor with ID 6
18/04/13 05:37:43 INFO yarn.YarnAllocator: Received 1 containers from YARN, launching executors on 1 of them.
18/04/13 05:37:43 INFO yarn.YarnAllocator: Completed container container_1523459048274_0025_01_000007 on host: apiros-4.gce.test.com (state: COMPLETE, exit status: 1)
18/04/13 05:37:43 INFO yarn.YarnAllocatorBlacklistTracker: blacklisting host as YARN allocation failed: apiros-4.gce.test.com
18/04/13 05:37:43 INFO yarn.YarnAllocatorBlacklistTracker: adding nodes to YARN application master's blacklist: List(apiros-4.gce.test.com)
18/04/13 05:37:43 WARN yarn.YarnAllocator: Container marked as failed: container_1523459048274_0025_01_000007 on host: apiros-4.gce.test.com. Exit status: 1. Diagnostics: Exception from container-launch.
Container id: container_1523459048274_0025_01_000007
Exit code: 1
Stack trace: ExitCodeException exitCode=1:
at org.apache.hadoop.util.Shell.runCommand(Shell.java:604)
at org.apache.hadoop.util.Shell.run(Shell.java:507)
at org.apache.hadoop.util.Shell$ShellCommandExecutor.execute(Shell.java:789)
at org.apache.hadoop.yarn.server.nodemanager.DefaultContainerExecutor.launchContainer(DefaultContainerExecutor.java:213)
at org.apache.hadoop.yarn.server.nodemanager.containermanager.launcher.ContainerLaunch.call(ContainerLaunch.java:302)
at org.apache.hadoop.yarn.server.nodemanager.containermanager.launcher.ContainerLaunch.call(ContainerLaunch.java:82)
at java.util.concurrent.FutureTask.run(FutureTask.java:266)
at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
at java.lang.Thread.run(Thread.java:748)
```
Where the most important part is:
```
18/04/13 05:37:43 INFO yarn.YarnAllocatorBlacklistTracker: blacklisting host as YARN allocation failed: apiros-4.gce.test.com
18/04/13 05:37:43 INFO yarn.YarnAllocatorBlacklistTracker: adding nodes to YARN application master's blacklist: List(apiros-4.gce.test.com)
```
And execution was continued (no shutdown called).
### Testing the backlisting of the whole cluster
Starting Spark with YARN blacklisting enabled then removing a the Spark core jar one by one from all the cluster nodes. Then executing a simple spark job which fails checking the yarn log the expected exit status is contained:
```
18/06/15 01:07:10 INFO yarn.ApplicationMaster: Final app status: FAILED, exitCode: 11, (reason: Due to executor failures all available nodes are blacklisted)
18/06/15 01:07:13 INFO util.ShutdownHookManager: Shutdown hook called
```
Author: “attilapiros” <piros.attila.zsolt@gmail.com>
Closes#21068 from attilapiros/SPARK-16630.
## What changes were proposed in this pull request?
Hadoop 3 introduces HDFS federation. This means that multiple namespaces are allowed on the same HDFS cluster. In Spark, we need to ask the delegation token for all the namenodes (for each namespace), otherwise accessing any other namespace different from the default one (for which we already fetch the delegation token) fails.
The PR adds the automatic discovery of all the namenodes related to all the namespaces available according to the configs in hdfs-site.xml.
## How was this patch tested?
manual tests in dockerized env
Author: Marco Gaido <marcogaido91@gmail.com>
Closes#21216 from mgaido91/SPARK-24149.
Instead of always throwing a generic exception when the AM fails,
print a generic error and throw the exception with the YARN
diagnostics containing the reason for the failure.
There was an issue with YARN sometimes providing a generic diagnostic
message, even though the AM provides a failure reason when
unregistering. That was happening because the AM was registering
too late, and if errors happened before the registration, YARN would
just create a generic "ExitCodeException" which wasn't very helpful.
Since most errors in this path are a result of not being able to
connect to the driver, this change modifies the AM registration
a bit so that the AM is registered before the connection to the
driver is established. That way, errors are properly propagated
through YARN back to the driver.
As part of that, I also removed the code that retried connections
to the driver from the client AM. At that point, the driver should
already be up and waiting for connections, so it's unlikely that
retrying would help - and in case it does, that means a flaky
network, which would mean problems would probably show up again.
The effect of that is that connection-related errors are reported
back to the driver much faster now (through the YARN report).
One thing to note is that there seems to be a race on the YARN
side that causes a report to be sent to the client without the
corresponding diagnostics string from the AM; the diagnostics are
available later from the RM web page. For that reason, the generic
error messages are kept in the Spark scheduler code, to help
guide users to a way of debugging their failure.
Also of note is that if YARN's max attempts configuration is lower
than Spark's, Spark will not unregister the AM with a proper
diagnostics message. Unfortunately there seems to be no way to
unregister the AM and still allow further re-attempts to happen.
Testing:
- existing unit tests
- some of our integration tests
- hardcoded an invalid driver address in the code and verified
the error in the shell. e.g.
```
scala> 18/05/04 15:09:34 ERROR cluster.YarnClientSchedulerBackend: YARN application has exited unexpectedly with state FAILED! Check the YARN application logs for more details.
18/05/04 15:09:34 ERROR cluster.YarnClientSchedulerBackend: Diagnostics message: Uncaught exception: org.apache.spark.SparkException: Exception thrown in awaitResult:
<AM stack trace>
Caused by: java.io.IOException: Failed to connect to localhost/127.0.0.1:1234
<More stack trace>
```
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#21243 from vanzin/SPARK-24182.
## What changes were proposed in this pull request?
Easy fix in the documentation.
## How was this patch tested?
N/A
Closes#20948
Author: Daniel Sakuma <dsakuma@gmail.com>
Closes#20928 from dsakuma/fix_typo_configuration_docs.
This change basically rewrites the security documentation so that it's
up to date with new features, more correct, and more complete.
Because security is such an important feature, I chose to move all the
relevant configuration documentation to the security page, instead of
having them peppered all over the place in the configuration page. This
allows an almost one-stop shop for security configuration in Spark. The
only exceptions are some YARN-specific minor features which I left in
the YARN page.
I also re-organized the page's topics, since they didn't make a lot of
sense. You had kerberos features described inside paragraphs talking
about UI access control, and other oddities. It should be easier now
to find information about specific Spark security features. I also
enabled TOCs for both the Security and YARN pages, since that makes it
easier to see what is covered.
I removed most of the comments from the SecurityManager javadoc since
they just replicated information in the security doc, with different
levels of out-of-dateness.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#20742 from vanzin/SPARK-23572.
## What changes were proposed in this pull request?
The example jar file is now in ./examples/jars directory of Spark distribution.
Author: Arseniy Tashoyan <tashoyan@users.noreply.github.com>
Closes#20349 from tashoyan/patch-1.
## What changes were proposed in this pull request?
Fix spelling in quick-start doc.
## How was this patch tested?
Doc only.
Author: Shashwat Anand <me@shashwat.me>
Closes#20336 from ashashwat/SPARK-23165.
What changes were proposed in this pull request?
This PR contains documentation on the usage of Kubernetes scheduler in Spark 2.3, and a shell script to make it easier to build docker images required to use the integration. The changes detailed here are covered by https://github.com/apache/spark/pull/19717 and https://github.com/apache/spark/pull/19468 which have merged already.
How was this patch tested?
The script has been in use for releases on our fork. Rest is documentation.
cc rxin mateiz (shepherd)
k8s-big-data SIG members & contributors: foxish ash211 mccheah liyinan926 erikerlandson ssuchter varunkatta kimoonkim tnachen ifilonenko
reviewers: vanzin felixcheung jiangxb1987 mridulm
TODO:
- [x] Add dockerfiles directory to built distribution. (https://github.com/apache/spark/pull/20007)
- [x] Change references to docker to instead say "container" (https://github.com/apache/spark/pull/19995)
- [x] Update configuration table.
- [x] Modify spark.kubernetes.allocation.batch.delay to take time instead of int (#20032)
Author: foxish <ramanathana@google.com>
Closes#19946 from foxish/update-k8s-docs.
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.
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?
In the current Spark, when submitting application on YARN with remote resources `./bin/spark-shell --jars http://central.maven.org/maven2/com/github/swagger-akka-http/swagger-akka-http_2.11/0.10.1/swagger-akka-http_2.11-0.10.1.jar --master yarn-client -v`, Spark will be failed with:
```
java.io.IOException: No FileSystem for scheme: http
at org.apache.hadoop.fs.FileSystem.getFileSystemClass(FileSystem.java:2586)
at org.apache.hadoop.fs.FileSystem.createFileSystem(FileSystem.java:2593)
at org.apache.hadoop.fs.FileSystem.access$200(FileSystem.java:91)
at org.apache.hadoop.fs.FileSystem$Cache.getInternal(FileSystem.java:2632)
at org.apache.hadoop.fs.FileSystem$Cache.get(FileSystem.java:2614)
at org.apache.hadoop.fs.FileSystem.get(FileSystem.java:370)
at org.apache.hadoop.fs.Path.getFileSystem(Path.java:296)
at org.apache.spark.deploy.yarn.Client.copyFileToRemote(Client.scala:354)
at org.apache.spark.deploy.yarn.Client.org$apache$spark$deploy$yarn$Client$$distribute$1(Client.scala:478)
at org.apache.spark.deploy.yarn.Client$$anonfun$prepareLocalResources$11$$anonfun$apply$6.apply(Client.scala:600)
at org.apache.spark.deploy.yarn.Client$$anonfun$prepareLocalResources$11$$anonfun$apply$6.apply(Client.scala:599)
at scala.collection.mutable.ArraySeq.foreach(ArraySeq.scala:74)
at org.apache.spark.deploy.yarn.Client$$anonfun$prepareLocalResources$11.apply(Client.scala:599)
at org.apache.spark.deploy.yarn.Client$$anonfun$prepareLocalResources$11.apply(Client.scala:598)
at scala.collection.immutable.List.foreach(List.scala:381)
at org.apache.spark.deploy.yarn.Client.prepareLocalResources(Client.scala:598)
at org.apache.spark.deploy.yarn.Client.createContainerLaunchContext(Client.scala:848)
at org.apache.spark.deploy.yarn.Client.submitApplication(Client.scala:173)
```
This is because `YARN#client` assumes resources are on the Hadoop compatible FS. To fix this problem, here propose to download remote http(s) resources to local and add this local downloaded resources to dist cache. This solution has one downside: remote resources are downloaded and uploaded again, but it only restricted to only remote http(s) resources, also the overhead is not so big. The advantages of this solution is that it is simple and the code changes restricts to only `SparkSubmit`.
## How was this patch tested?
Unit test added, also verified in local cluster.
Author: jerryshao <sshao@hortonworks.com>
Closes#19130 from jerryshao/SPARK-21917.
## 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.
## 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.
Allow an application to use the History Server URL as the tracking
URL in the YARN RM, so there's still a link to the web UI somewhere
in YARN even if the driver's UI is disabled. This is useful, for
example, if an admin wants to disable the driver UI by default for
applications, since it's harder to secure it (since it involves non
trivial ssl certificate and auth management that admins may not want
to expose to user apps).
This needs to be opt-in, because of the way the YARN proxy works, so
a new configuration was added to enable the option.
The YARN RM will proxy requests to live AMs instead of redirecting
the client, so pages in the SHS UI will not render correctly since
they'll reference invalid paths in the RM UI. The proxy base support
in the SHS cannot be used since that would prevent direct access to
the SHS.
So, to solve this problem, for the feature to work end-to-end, a new
YARN-specific filter was added that detects whether the requests come
from the proxy and redirects the client appropriatly. The SHS admin has
to add this filter manually if they want the feature to work.
Tested with new unit test, and by running with the documented configuration
set in a test cluster. Also verified the driver UI is used when it's
enabled.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#16946 from vanzin/SPARK-19554.
## What changes were proposed in this pull request?
Due to the newly added API in Hadoop 2.6.4+, Spark builds against Hadoop 2.6.0~2.6.3 will meet compile error. So here still reverting back to use reflection to handle this issue.
## How was this patch tested?
Manual verification.
Author: jerryshao <sshao@hortonworks.com>
Closes#16884 from jerryshao/SPARK-19545.
## What changes were proposed in this pull request?
- Remove support for Hadoop 2.5 and earlier
- Remove reflection and code constructs only needed to support multiple versions at once
- Update docs to reflect newer versions
- Remove older versions' builds and profiles.
## How was this patch tested?
Existing tests
Author: Sean Owen <sowen@cloudera.com>
Closes#16810 from srowen/SPARK-19464.
## What changes were proposed in this pull request?
`spark.yarn.access.namenodes` configuration cannot actually reflects the usage of it, inside the code it is the Hadoop filesystems we get tokens, not NNs. So here propose to update the name of this configuration, also change the related code and doc.
## How was this patch tested?
Local verification.
Author: jerryshao <sshao@hortonworks.com>
Closes#16560 from jerryshao/SPARK-19179.
Currently Spark can only get token renewal interval from security HDFS (hdfs://), if Spark runs with other security file systems like webHDFS (webhdfs://), wasb (wasb://), ADLS, it will ignore these tokens and not get token renewal intervals from these tokens. These will make Spark unable to work with these security clusters. So instead of only checking HDFS token, we should generalize to support different DelegationTokenIdentifier.
## How was this patch tested?
Manually verified in security cluster.
Author: jerryshao <sshao@hortonworks.com>
Closes#16432 from jerryshao/SPARK-19021.
## What changes were proposed in this pull request?
The configuration `spark.yarn.security.tokens.{service}.enabled` is deprecated. Now we should use `spark.yarn.security.credentials.{service}.enabled`. Some places in the doc is not updated yet.
## How was this patch tested?
N/A. Just doc change.
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes#16444 from viirya/minor-credential-provider-doc.
Looking at the distributions provided on spark.apache.org, I see that the Spark YARN shuffle jar is under `yarn/` and not `lib/`.
This change is so minor I'm not sure it needs a JIRA. But let me know if so and I'll create one.
Author: Nicholas Chammas <nicholas.chammas@gmail.com>
Closes#16130 from nchammas/yarn-doc-fix.
## What changes were proposed in this pull request?
Remove `spark.driver.memory`, `spark.executor.memory`, `spark.driver.cores`, and `spark.executor.cores` from `running-on-yarn.md` as they are not Yarn-specific, and they are also defined in`configuration.md`.
## How was this patch tested?
Build passed & Manually check.
Author: Weiqing Yang <yangweiqing001@gmail.com>
Closes#15869 from weiqingy/yarnDoc.
## What changes were proposed in this pull request?
Suggest users to increase `NodeManager's` heap size if `External Shuffle Service` is enabled as
`NM` can spend a lot of time doing GC resulting in shuffle operations being a bottleneck due to `Shuffle Read blocked time` bumped up.
Also because of GC `NodeManager` can use an enormous amount of CPU and cluster performance will suffer.
I have seen NodeManager using 5-13G RAM and up to 2700% CPU with `spark_shuffle` service on.
## How was this patch tested?
#### Added step 5:
Author: Artur Sukhenko <artur.sukhenko@gmail.com>
Closes#15906 from Devian-ua/nmHeapSize.
## What changes were proposed in this pull request?
Add a configurable token manager for Spark on running on yarn.
### Current Problems ###
1. Supported token provider is hard-coded, currently only hdfs, hbase and hive are supported and it is impossible for user to add new token provider without code changes.
2. Also this problem exits in timely token renewer and updater.
### Changes In This Proposal ###
In this proposal, to address the problems mentioned above and make the current code more cleaner and easier to understand, mainly has 3 changes:
1. Abstract a `ServiceTokenProvider` as well as `ServiceTokenRenewable` interface for token provider. Each service wants to communicate with Spark through token way needs to implement this interface.
2. Provide a `ConfigurableTokenManager` to manage all the register token providers, also token renewer and updater. Also this class offers the API for other modules to obtain tokens, get renewal interval and so on.
3. Implement 3 built-in token providers `HDFSTokenProvider`, `HiveTokenProvider` and `HBaseTokenProvider` to keep the same semantics as supported today. Whether to load in these built-in token providers is controlled by configuration "spark.yarn.security.tokens.${service}.enabled", by default for all the built-in token providers are loaded.
### Behavior Changes ###
For the end user there's no behavior change, we still use the same configuration `spark.yarn.security.tokens.${service}.enabled` to decide which token provider is enabled (hbase or hive).
For user implemented token provider (assume the name of token provider is "test") needs to add into this class should have two configurations:
1. `spark.yarn.security.tokens.test.enabled` to true
2. `spark.yarn.security.tokens.test.class` to the full qualified class name.
So we still keep the same semantics as current code while add one new configuration.
### Current Status ###
- [x] token provider interface and management framework.
- [x] implement built-in token providers (hdfs, hbase, hive).
- [x] Coverage of unit test.
- [x] Integrated test with security cluster.
## How was this patch tested?
Unit test and integrated test.
Please suggest and review, any comment is greatly appreciated.
Author: jerryshao <sshao@hortonworks.com>
Closes#14065 from jerryshao/SPARK-16342.
This prevents the NM from starting when something is wrong, which would
lead to later errors which are confusing and harder to debug.
Added a unit test to verify startup fails if something is wrong.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#14162 from vanzin/SPARK-16505.
## What changes were proposed in this pull request?
Yarn supports rolling log aggregation since 2.6, previously log will only be aggregated to HDFS after application is finished, it is quite painful for long running applications like Spark Streaming, thriftserver. Also out of disk problem will be occurred when log file is too large. So here propose to add support of rolling log aggregation for Spark on yarn.
One limitation for this is that log4j should be set to change to file appender, now in Spark itself uses console appender by default, in which file will not be created again once removed after aggregation. But I think lots of production users should have changed their log4j configuration instead of default on, so this is not a big problem.
## How was this patch tested?
Manually verified with Hadoop 2.7.1.
Author: jerryshao <sshao@hortonworks.com>
Closes#13712 from jerryshao/SPARK-15990.
## What changes were proposed in this pull request?
This changes the behavior of --num-executors and spark.executor.instances when using dynamic allocation. Instead of turning dynamic allocation off, it uses the value for the initial number of executors.
This changes was discussed on [SPARK-13723](https://issues.apache.org/jira/browse/SPARK-13723). I highly recommend using it while we can change the behavior for 2.0.0. In practice, the 1.x behavior causes unexpected behavior for users (it is not clear that it disables dynamic allocation) and wastes cluster resources because users rarely notice the log message.
## How was this patch tested?
This patch updates tests and adds a test for Utils.getDynamicAllocationInitialExecutors.
Author: Ryan Blue <blue@apache.org>
Closes#13338 from rdblue/SPARK-13723-num-executors-with-dynamic-allocation.
## What changes were proposed in this pull request?
Remove several obsolete env variables not supported for Spark on YARN now, also updates the docs to include several changes with 2.0.
## How was this patch tested?
N/A
CC vanzin tgravescs
Author: jerryshao <sshao@hortonworks.com>
Closes#13296 from jerryshao/yarn-doc.
This patch provides detail on what to do for keytabless Oozie launches of spark apps, and adds some debug-level diagnostics of what credentials have been submitted
Author: Steve Loughran <stevel@hortonworks.com>
Author: Steve Loughran <stevel@apache.org>
Closes#11033 from steveloughran/stevel/feature/SPARK-13148-oozie.
This work is based on twinkle-sachdeva 's proposal. In parallel to such mechanism for AM failures, here add similar mechanism for executor failure tracking, this is useful for long running Spark service to mitigate the executor failure problems.
Please help to review, tgravescs sryza and vanzin
Author: jerryshao <sshao@hortonworks.com>
Closes#10241 from jerryshao/SPARK-6735.
## What changes were proposed in this pull request?
The configuration docs are updated to reflect the changes introduced with [SPARK-12384](https://issues.apache.org/jira/browse/SPARK-12384). This allows the user to specify initial heap memory settings through the extraJavaOptions for executor, driver and am.
## How was this patch tested?
The changes are tested in [SPARK-12384](https://issues.apache.org/jira/browse/SPARK-12384). This is just documenting the changes made.
Author: Dhruve Ashar <dhruveashar@gmail.com>
Closes#12333 from dhruve/doc/SPARK-14572.
## What changes were proposed in this pull request?
Made the SPARK YARN STAGING DIR as configurable with the configuration as 'spark.yarn.staging-dir'.
## How was this patch tested?
I have verified it manually by running applications on yarn, If the 'spark.yarn.staging-dir' is configured then the value used as staging directory otherwise uses the default value i.e. file system’s home directory for the user.
Author: Devaraj K <devaraj@apache.org>
Closes#12082 from devaraj-kavali/SPARK-13063.
## What changes were proposed in this pull request?
Currently in Spark on YARN, configurations can be passed through SparkConf, env and command arguments, some parts are duplicated, like client argument and SparkConf. So here propose to simplify the command arguments.
## How was this patch tested?
This patch is tested manually with unit test.
CC vanzin tgravescs , please help to suggest this proposal. The original purpose of this JIRA is to remove `ClientArguments`, through refactoring some arguments like `--class`, `--arg` are not so easy to replace, so here I remove the most part of command line arguments, only keep the minimal set.
Author: jerryshao <sshao@hortonworks.com>
Closes#11603 from jerryshao/SPARK-12343.
