### What changes were proposed in this pull request?
During Spark History Server startup, there are two things happening simultaneously that call into `java.nio.file.FileSystems.getDefault()` and we sometime hit [JDK-8194653](https://bugs.openjdk.java.net/browse/JDK-8194653).
1) start jetty server
2) start ApplicationHistoryProvider (which reads files from HDFS)
We should do these two things sequentially instead of in parallel.
We introduce a start() method in ApplicationHistoryProvider (and its subclass FsHistoryProvider), and we do initialize inside the start() method instead of the constructor.
In HistoryServer, we explicitly call provider.start() after we call bind() which starts the Jetty server.
### Why are the changes needed?
It is a bug that occasionally starting Spark History Server results in process hang due to deadlock among threads.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
I stress tested this PR with a bash script to stop and start Spark History Server more than 1000 times, it worked fine. Previously I can only do the stop/start loop less than 10 times before I hit the deadlock issue.
Closes#25705 from shanyu/shanyu-29003.
Authored-by: Shanyu Zhao <shzhao@microsoft.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
We already have found and fixed the correctness issue before when RDD output is INDETERMINATE. One missing part is sampling-based RDD. This kind of RDDs is order sensitive to its input. A sampling-based RDD with unordered input, should be INDETERMINATE.
### Why are the changes needed?
A sampling-based RDD with unordered input is just like MapPartitionsRDD with isOrderSensitive parameter as true. The RDD output can be different after a rerun.
It is a problem in ML applications.
In ML, sample is used to prepare training data. ML algorithm fits the model based on the sampled data. If rerun tasks of sample produce different output during model fitting, ML results will be unreliable and also buggy.
Each sample is random output, but once you sampled, the output should be determinate.
### Does this PR introduce any user-facing change?
Previously, a sampling-based RDD can possibly come with different output after a rerun.
After this patch, sampling-based RDD is INDETERMINATE. For an INDETERMINATE map stage, currently Spark scheduler will re-try all the tasks of the failed stage.
### How was this patch tested?
Added test.
Closes#25751 from viirya/sample-order-sensitive.
Authored-by: Liang-Chi Hsieh <liangchi@uber.com>
Signed-off-by: Liang-Chi Hsieh <liangchi@uber.com>
JIRA :https://issues.apache.org/jira/browse/SPARK-29050
'a hdfs' change into 'an hdfs'
'an unique' change into 'a unique'
'an url' change into 'a url'
'a error' change into 'an error'
Closes#25756 from dengziming/feature_fix_typos.
Authored-by: dengziming <dengziming@growingio.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
This patch ensures accessing recorded values in listener is always after letting listeners fully process all events. To ensure this, this patch adds new class to hide these values and access with methods which will ensure above condition. Without this guard, two threads are running concurrently - 1) listeners process thread 2) test main thread - and race condition would occur.
That's why we also see very odd thing, error message saying condition is met but test failed:
```
- Barrier task failures from the same stage attempt don't trigger multiple stage retries *** FAILED ***
ArrayBuffer(0) did not equal List(0) (DAGSchedulerSuite.scala:2656)
```
which means verification failed, and condition is met just before constructing error message.
The guard is properly placed in many spots, but missed in some places. This patch enforces that it can't be missed.
### Why are the changes needed?
UT fails intermittently and this patch will address the flakyness.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Modified UT.
Also made the flaky tests artificially failing via applying 50ms of sleep on each onXXX method.
I found 3 methods being failed. (They've marked as X. Just ignore ! as they failed on waiting listener in given timeout and these tests don't deal with these recorded values - it uses other timeout value 1000ms than 10000ms for this listener so affected via side-effect.)
When I applied same in this patch all tests marked as X passed.
Closes#25706 from HeartSaVioR/SPARK-26989.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
### What changes were proposed in this pull request?
PR #22112 fixed the todo added by PR #20393(SPARK-23207). We can remove it now.
### Why are the changes needed?
In order not to confuse developers.
### Does this PR introduce any user-facing change?
no
### How was this patch tested?
no need to test
Closes#25755 from LinhongLiu/remove-todo.
Authored-by: Liu,Linhong <liulinhong@baidu.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
Uses the APIs introduced in SPARK-28209 in the UnsafeShuffleWriter.
## How was this patch tested?
Since this is just a refactor, existing unit tests should cover the relevant code paths. Micro-benchmarks from the original fork where this code was built show no degradation in performance.
Closes#25304 from mccheah/shuffle-writer-refactor-unsafe-writer.
Lead-authored-by: mcheah <mcheah@palantir.com>
Co-authored-by: mccheah <mcheah@palantir.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
Running spark on yarn, I got
```
java.lang.ClassCastException: org.apache.hadoop.ipc.CallerContext$Builder cannot be cast to scala.runtime.Nothing$
```
Utils.classForName return Class[Nothing], I think it should be defind as Class[_] to resolve this issue
## How was this patch tested?
not need
Closes#25389 from hddong/SPARK-28657-fix-currentContext-Instance-failed.
Lead-authored-by: hongdd <jn_hdd@163.com>
Co-authored-by: hongdongdong <hongdongdong@cmss.chinamobile.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
### What changes were proposed in this pull request?
- Remove SQLContext.createExternalTable and Catalog.createExternalTable, deprecated in favor of createTable since 2.2.0, plus tests of deprecated methods
- Remove HiveContext, deprecated in 2.0.0, in favor of `SparkSession.builder.enableHiveSupport`
- Remove deprecated KinesisUtils.createStream methods, plus tests of deprecated methods, deprecate in 2.2.0
- Remove deprecated MLlib (not Spark ML) linear method support, mostly utility constructors and 'train' methods, and associated docs. This includes methods in LinearRegression, LogisticRegression, Lasso, RidgeRegression. These have been deprecated since 2.0.0
- Remove deprecated Pyspark MLlib linear method support, including LogisticRegressionWithSGD, LinearRegressionWithSGD, LassoWithSGD
- Remove 'runs' argument in KMeans.train() method, which has been a no-op since 2.0.0
- Remove deprecated ChiSqSelector isSorted protected method
- Remove deprecated 'yarn-cluster' and 'yarn-client' master argument in favor of 'yarn' and deploy mode 'cluster', etc
Notes:
- I was not able to remove deprecated DataFrameReader.json(RDD) in favor of DataFrameReader.json(Dataset); the former was deprecated in 2.2.0, but, it is still needed to support Pyspark's .json() method, which can't use a Dataset.
- Looks like SQLContext.createExternalTable was not actually deprecated in Pyspark, but, almost certainly was meant to be? Catalog.createExternalTable was.
- I afterwards noted that the toDegrees, toRadians functions were almost removed fully in SPARK-25908, but Felix suggested keeping just the R version as they hadn't been technically deprecated. I'd like to revisit that. Do we really want the inconsistency? I'm not against reverting it again, but then that implies leaving SQLContext.createExternalTable just in Pyspark too, which seems weird.
- I *kept* LogisticRegressionWithSGD, LinearRegressionWithSGD, LassoWithSGD, RidgeRegressionWithSGD in Pyspark, though deprecated, as it is hard to remove them (still used by StreamingLogisticRegressionWithSGD?) and they are not fully removed in Scala. Maybe should not have been deprecated.
### Why are the changes needed?
Deprecated items are easiest to remove in a major release, so we should do so as much as possible for Spark 3. This does not target items deprecated 'recently' as of Spark 2.3, which is still 18 months old.
### Does this PR introduce any user-facing change?
Yes, in that deprecated items are removed from some public APIs.
### How was this patch tested?
Existing tests.
Closes#25684 from srowen/SPARK-28980.
Lead-authored-by: Sean Owen <sean.owen@databricks.com>
Co-authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
### What changes were proposed in this pull request?
If a Spark task is killed due to intentional job kills, automated killing of redundant speculative tasks, etc, ClosedByInterruptException occurs if task has unfinished I/O operation with AbstractInterruptibleChannel. A single cancelled task can result in hundreds of stack trace of ClosedByInterruptException being logged.
In this PR, stack trace of ClosedByInterruptException won't be logged like Executor.run do for InterruptedException.
### Why are the changes needed?
Large numbers of spurious exceptions is confusing to users when they are inspecting Spark logs to diagnose other issues.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
N/A
Closes#25674 from colinmjj/spark-28340.
Authored-by: colinma <colinma@tencent.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
### What changes were proposed in this pull request?
In spark-shell local mode, in the task page, host name is coming as localhost
This PR changes it to show machine IP, as shown in the "spark.driver.host" in the environment page
### Why are the changes needed?
To show the proper IP in the task page host column
### Does this PR introduce any user-facing change?
It updates the SPARK UI->Task page->Host Column
### How was this patch tested?
verfied in spark UI
Closes#25645 from shivusondur/localhost1.
Authored-by: shivusondur <shivusondur@gmail.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
### What changes were proposed in this pull request?
This patch fixes the bug which throws ConcurrentModificationException when job with 0 partition is submitted via DAGScheduler.
### Why are the changes needed?
Without this patch, structured streaming query throws ConcurrentModificationException, like below stack trace:
```
19/09/04 09:48:49 ERROR AsyncEventQueue: Listener EventLoggingListener threw an exception
java.util.ConcurrentModificationException
at java.util.Hashtable$Enumerator.next(Hashtable.java:1387)
at scala.collection.convert.Wrappers$JPropertiesWrapper$$anon$6.next(Wrappers.scala:424)
at scala.collection.convert.Wrappers$JPropertiesWrapper$$anon$6.next(Wrappers.scala:420)
at scala.collection.Iterator.foreach(Iterator.scala:941)
at scala.collection.Iterator.foreach$(Iterator.scala:941)
at scala.collection.AbstractIterator.foreach(Iterator.scala:1429)
at scala.collection.IterableLike.foreach(IterableLike.scala:74)
at scala.collection.IterableLike.foreach$(IterableLike.scala:73)
at scala.collection.AbstractIterable.foreach(Iterable.scala:56)
at scala.collection.TraversableLike.map(TraversableLike.scala:237)
at scala.collection.TraversableLike.map$(TraversableLike.scala:230)
at scala.collection.AbstractTraversable.map(Traversable.scala:108)
at org.apache.spark.util.JsonProtocol$.mapToJson(JsonProtocol.scala:514)
at org.apache.spark.util.JsonProtocol$.$anonfun$propertiesToJson$1(JsonProtocol.scala:520)
at scala.Option.map(Option.scala:163)
at org.apache.spark.util.JsonProtocol$.propertiesToJson(JsonProtocol.scala:519)
at org.apache.spark.util.JsonProtocol$.jobStartToJson(JsonProtocol.scala:155)
at org.apache.spark.util.JsonProtocol$.sparkEventToJson(JsonProtocol.scala:79)
at org.apache.spark.scheduler.EventLoggingListener.logEvent(EventLoggingListener.scala:149)
at org.apache.spark.scheduler.EventLoggingListener.onJobStart(EventLoggingListener.scala:217)
at org.apache.spark.scheduler.SparkListenerBus.doPostEvent(SparkListenerBus.scala:37)
at org.apache.spark.scheduler.SparkListenerBus.doPostEvent$(SparkListenerBus.scala:28)
at org.apache.spark.scheduler.AsyncEventQueue.doPostEvent(AsyncEventQueue.scala:37)
at org.apache.spark.scheduler.AsyncEventQueue.doPostEvent(AsyncEventQueue.scala:37)
at org.apache.spark.util.ListenerBus.postToAll(ListenerBus.scala:99)
at org.apache.spark.util.ListenerBus.postToAll$(ListenerBus.scala:84)
at org.apache.spark.scheduler.AsyncEventQueue.super$postToAll(AsyncEventQueue.scala:102)
at org.apache.spark.scheduler.AsyncEventQueue.$anonfun$dispatch$1(AsyncEventQueue.scala:102)
at scala.runtime.java8.JFunction0$mcJ$sp.apply(JFunction0$mcJ$sp.java:23)
at scala.util.DynamicVariable.withValue(DynamicVariable.scala:62)
at org.apache.spark.scheduler.AsyncEventQueue.org$apache$spark$scheduler$AsyncEventQueue$$dispatch(AsyncEventQueue.scala:97)
at org.apache.spark.scheduler.AsyncEventQueue$$anon$2.$anonfun$run$1(AsyncEventQueue.scala:93)
at org.apache.spark.util.Utils$.tryOrStopSparkContext(Utils.scala:1319)
at org.apache.spark.scheduler.AsyncEventQueue$$anon$2.run(AsyncEventQueue.scala:93)
```
Please refer https://issues.apache.org/jira/browse/SPARK-28967 for detailed reproducer.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Newly added UT. Also manually tested via running simple structured streaming query in spark-shell.
Closes#25672 from HeartSaVioR/SPARK-28967.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan@gmail.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
### What changes were proposed in this pull request?
`ReplayListenerSuite` depends on a listener class to listen for replayed events. This class was implemented by extending `EventLoggingListener`. `EventLoggingListener` does not log executor metrics update events, but uses them to update internal state; on a stage completion event, it then logs stage executor metrics events using this internal state. As executor metrics update events do not get written to the event log, they do not get replayed. The internal state of the replay listener can therefore be different from the original listener, leading to different stage completion events being logged.
We reimplement the replay listener to simply buffer each and every event it receives. This makes it a simpler yet better tool for verifying the events that get sent through the ReplayListenerBus.
### Why are the changes needed?
As explained above. Tests sometimes fail due to events being received by the `EventLoggingListener` that do not get logged (and thus do not get replayed) but influence other events that get logged.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
Existing unit tests.
Closes#25673 from wypoon/SPARK-28770.
Authored-by: Wing Yew Poon <wypoon@cloudera.com>
Signed-off-by: Imran Rashid <irashid@cloudera.com>
### What changes were proposed in this pull request?
Use `KeyLock` added in #25612 to simplify `MapOutputTracker.getStatuses`. It also has some improvement after the refactoring:
- `InterruptedException` is no longer sallowed.
- When a shuffle block is fetched, we don't need to wake up unrelated sleeping threads.
### Why are the changes needed?
`MapOutputTracker.getStatuses` is pretty hard to maintain right now because it has a special lock mechanism which we needs to pay attention to whenever updating this method. As we can use `KeyLock` to hide the complexity of locking behind a dedicated lock class, it's better to refactor it to make it easy to understand and maintain.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Existing tests.
Closes#25680 from zsxwing/getStatuses.
Authored-by: Shixiong Zhu <zsxwing@gmail.com>
Signed-off-by: Shixiong Zhu <zsxwing@gmail.com>
### What changes were proposed in this pull request?
Replaces some incorrect usage of `new Configuration()` as it will load default configs defined in Hadoop
### Why are the changes needed?
Unexpected config could be accessed instead of the expected config, see SPARK-28203 for example
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
Existed tests.
Closes#25616 from advancedxy/remove_invalid_configuration.
Authored-by: Xianjin YE <advancedxy@gmail.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
### What changes were proposed in this pull request?
This patch implements dynamic partition pruning by adding a dynamic-partition-pruning filter if there is a partitioned table and a filter on the dimension table. The filter is then planned using a heuristic approach:
1. As a broadcast relation if it is a broadcast hash join. The broadcast relation will then be transformed into a reused broadcast exchange by the `ReuseExchange` rule; or
2. As a subquery duplicate if the estimated benefit of partition table scan being saved is greater than the estimated cost of the extra scan of the duplicated subquery; otherwise
3. As a bypassed condition (`true`).
### Why are the changes needed?
This is an important performance feature.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
Added UT
- Testing DPP by enabling / disabling the reuse broadcast results feature and / or the subquery duplication feature.
- Testing DPP with reused broadcast results.
- Testing the key iterators on different HashedRelation types.
- Testing the packing and unpacking of the broadcast keys in a LongType.
Closes#25600 from maryannxue/dpp.
Authored-by: maryannxue <maryannxue@apache.org>
Signed-off-by: Xiao Li <gatorsmile@gmail.com>
### What changes were proposed in this pull request?
This patch fixes the bugs in test code itself, FsHistoryProviderSuite.
1. When creating log file via `newLogFile`, codec is ignored, leading to wrong file name. (No one tends to create test for test code, as well as the bug doesn't affect existing tests indeed, so not easy to catch.)
2. When writing events to log file via `writeFile`, metadata (in case of new format) gets written to file regardless of its codec, and the content is overwritten by another stream, hence no information for Spark version is available. It affects existing test, hence we have wrong expected value to workaround the bug.
This patch also removes redundant parameter `isNewFormat` in `writeFile`, as according to review comment, Spark no longer supports old format.
### Why are the changes needed?
Explained in above section why they're bugs, though they only reside in test-code. (Please note that the bug didn't come from non-test side of code.)
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Modified existing UTs, as well as read event log file in console to see metadata is not overwritten by other contents.
Closes#25629 from HeartSaVioR/MINOR-FIX-FsHistoryProviderSuite.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
### What changes were proposed in this pull request?
We are adding other resource type support to the executors and Spark. We should show the resource information for each executor on the UI Executors page.
This also adds a toggle button to show the resources column. It is off by default.
### Why are the changes needed?
to show user what resources the executors have. Like Gpus, fpgas, etc
### Does this PR introduce any user-facing change?
Yes introduces UI and rest api changes to show the resources
### How was this patch tested?
Unit tests and manual UI tests on yarn and standalone modes.
Closes#25613 from tgravescs/SPARK-27489-gpu-ui-latest.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Gengliang Wang <gengliang.wang@databricks.com>
### What changes were proposed in this pull request?
This PR provides a new lock mechanism `KeyLock` to lock with a given key. Also use this new lock in `TorrentBroadcast` to avoid blocking tasks from fetching different broadcast values.
### Why are the changes needed?
`TorrentBroadcast.readObject` uses a global lock so only one task can be fetching the blocks at the same time. This is not optimal if we are running multiple stages concurrently because they should be able to independently fetch their own blocks.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
Existing tests.
Closes#25612 from zsxwing/SPARK-3137.
Authored-by: Shixiong Zhu <zsxwing@gmail.com>
Signed-off-by: Shixiong Zhu <zsxwing@gmail.com>
### What changes were proposed in this pull request?
The Experimental and Evolving annotations are both (like Unstable) used to express that a an API may change. However there are many things in the code that have been marked that way since even Spark 1.x. Per the dev thread, anything introduced at or before Spark 2.3.0 is pretty much 'stable' in that it would not change without a deprecation cycle. Therefore I'd like to remove most of these annotations. And, remove the `:: Experimental ::` scaladoc tag too. And likewise for Python, R.
The changes below can be summarized as:
- Generally, anything introduced at or before Spark 2.3.0 has been unmarked as neither Evolving nor Experimental
- Obviously experimental items like DSv2, Barrier mode, ExperimentalMethods are untouched
- I _did_ unmark a few MLlib classes introduced in 2.4, as I am quite confident they're not going to change (e.g. KolmogorovSmirnovTest, PowerIterationClustering)
It's a big change to review, so I'd suggest scanning the list of _files_ changed to see if any area seems like it should remain partly experimental and examine those.
### Why are the changes needed?
Many of these annotations are incorrect; the APIs are de facto stable. Leaving them also makes legitimate usages of the annotations less meaningful.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
Existing tests.
Closes#25558 from srowen/SPARK-28855.
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?
Use the shuffle writer APIs introduced in SPARK-28209 in the sort shuffle writer.
## How was this patch tested?
Existing unit tests were changed to use the plugin instead, and they used the local disk version to ensure that there were no regressions.
Closes#25342 from mccheah/shuffle-writer-refactor-sort-shuffle-writer.
Lead-authored-by: mcheah <mcheah@palantir.com>
Co-authored-by: mccheah <mcheah@palantir.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
### What changes were proposed in this pull request?
When starting python processes, set `OMP_NUM_THREADS` to the number of cores allocated to an executor or driver if `OMP_NUM_THREADS` is not already set. Each python process will use the same `OMP_NUM_THREADS` setting, even if workers are not shared.
This avoids creating an OpenMP thread pool for parallel processing with a number of threads equal to the number of cores on the executor and [significantly reduces memory consumption](https://github.com/numpy/numpy/issues/10455). Instead, this threadpool should use the number of cores allocated to the executor, if available. If a setting for number of cores is not available, this doesn't change any behavior. OpenMP is used by numpy and pandas.
### Why are the changes needed?
To reduce memory consumption for PySpark jobs.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
Validated this reduces python worker memory consumption by more than 1GB on our cluster.
Closes#25545 from rdblue/SPARK-28843-set-omp-num-cores.
Authored-by: Ryan Blue <blue@apache.org>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
The shuffle writer API introduced in SPARK-28209 has a flaw that leads to a memory usage regression - we ended up tracking the partition lengths in two places. Here, we modify the API slightly to avoid redundant tracking. The implementation of the shuffle writer plugin is now responsible for tracking the lengths of partitions, and propagating this back up to the higher shuffle writer as part of the commitAllPartitions API.
## How was this patch tested?
Existing unit tests.
Closes#25341 from mccheah/dont-redundantly-store-part-lengths.
Authored-by: mcheah <mcheah@palantir.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
In my application spark streaming is restarted programmatically by stopping StreamingContext without stopping of SparkContext and creating/starting a new one. I use it for automatic detection of Kafka topic/partition changes and automatic failover in case of non fatal exceptions.
However i notice that after multiple restarts driver fails with OOM. During investigation of heap dump i figured out that StreamingContext object isn't cleared by GC after stopping.
<img width="1901" alt="Screen Shot 2019-08-14 at 12 23 33" src="https://user-images.githubusercontent.com/13151161/63010149-83f4c200-be8e-11e9-9f48-12b6e97839f4.png">
There are several places which holds reference to it :
1. StreamingTab registers StreamingJobProgressListener which holds reference to Streaming Context directly to LiveListenerBus shared queue via ssc.sc.addSparkListener(listener) method invocation. However this listener isn't unregistered at stop method.
2. json handlers (/streaming/json and /streaming/batch/json) aren't unregistered in SparkUI, while they hold reference to StreamingJobProgressListener. Basically the same issue affects all the pages, i assume that renderJsonHandler should be added to pageToHandlers cache on attachPage method invocation in order to unregistered it as well on detachPage.
3. SparkUi holds reference to StreamingJobProgressListener in the corresponding local variable which isn't cleared after stopping of StreamingContext.
## How was this patch tested?
Added tests to existing test suites.
After i applied these changes via reflection in my app OOM on driver side gone.
Closes#25439 from choojoyq/SPARK-28709-fix-streaming-context-leak-on-stop.
Authored-by: Nikita Gorbachevsky <nikitag@playtika.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
### What changes were proposed in this pull request?
Resolves [SPARK-28778: Shuffle jobs fail due to incorrect advertised address when running in a virtual network on Mesos](https://issues.apache.org/jira/browse/SPARK-28778).
This patch fixes a bug which occurs when shuffle jobs are launched by Mesos in a virtual network. Mesos scheduler sets executor `--hostname` parameter to `0.0.0.0` in the case when `spark.mesos.network.name` is provided. This makes executors use `0.0.0.0` as their advertised address and, in the presence of shuffle, executors fail to fetch shuffle blocks from each other using `0.0.0.0` as the origin. When a virtual network is used the hostname or IP address is not known upfront and assigned to a container at its start time so the executor process needs to advertise the correct dynamically assigned address to be reachable by other executors.
Changes:
- added a fallback to `Utils.localHostName()` in Spark Executors when `--hostname` is not provided
- removed setting executor address to `0.0.0.0` from Mesos scheduler
- refactored the code related to building executor command in Mesos scheduler
- added network configuration support to Docker containerizer
- added unit tests
### Why are the changes needed?
The bug described above prevents Mesos users from running any jobs which involve shuffle due to the inability of executors to fetch shuffle blocks because of incorrect advertised address when virtual network is used.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
- added unit test to `MesosCoarseGrainedSchedulerBackendSuite` which verifies the absence of `--hostname` parameter when `spark.mesos.network.name` is provided and its presence otherwise
- added unit test to `MesosSchedulerBackendUtilSuite` which verifies that `MesosSchedulerBackendUtil.buildContainerInfo` sets network-related properties for Docker containerizer
- unit tests from this repo launched with profiles: `./build/mvn test -Pmesos -Pnetlib-lgpl -Psparkr -Phive -Phive-thriftserver`, build log attached: [mvn.test.log](https://github.com/apache/spark/files/3516891/mvn.test.log)
- integration tests from [DCOS Spark repo](https://github.com/mesosphere/spark-build), more specifically - [test_spark_cni.py](https://github.com/mesosphere/spark-build/blob/master/tests/test_spark_cni.py) which runs a specific [shuffle job](https://github.com/mesosphere/spark-build/blob/master/tests/jobs/scala/src/main/scala/ShuffleApp.scala) and verifies its successful completion, Mesos task network configuration, and IP addresses for both Mesos and Docker containerizers
Closes#25500 from akirillov/DCOS-45840-fix-advertised-ip-in-virtual-networks.
Authored-by: Anton Kirillov <akirillov@mesosophere.io>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
Improved warning message in Barrier Execution Mode when required slots > maximum slots.
The new message contains information about required slots, maximum slots and how many times retry failed.
### Why are the changes needed?
Providing to users with the number of required slots, maximum slots and how many times retry failed might help users to decide what they should do.
For example, continuing to wait for retry succeeded or killing jobs.
### Does this PR introduce any user-facing change?
Yes.
If `spark.scheduler.barrier.maxConcurrentTaskCheck.maxFailures=3`, we get following warning message.
Before applying this change:
```
19/08/18 15:18:09 WARN DAGScheduler: The job 2 requires to run a barrier stage that requires more slots than the total number of slots in the cluster currently.
19/08/18 15:18:24 WARN DAGScheduler: The job 2 requires to run a barrier stage that requires more slots than the total number of slots in the cluster currently.
19/08/18 15:18:39 WARN DAGScheduler: The job 2 requires to run a barrier stage that requires more slots than the total number of slots in the cluster currently.
19/08/18 15:18:54 WARN DAGScheduler: The job 2 requires to run a barrier stage that requires more slots than the total number of slots in the cluster currently.
org.apache.spark.scheduler.BarrierJobSlotsNumberCheckFailed: [SPARK-24819]: Barrier execution mode does not allow run a barrier stage that requires more slots than the total number of slots in the cluster currently. Please init a new cluster with more CPU cores or repartition the input RDD(s) to reduce the number of slots required to run this barrier stage.
at org.apache.spark.scheduler.DAGScheduler.checkBarrierStageWithNumSlots(DAGScheduler.scala:439)
at org.apache.spark.scheduler.DAGScheduler.createResultStage(DAGScheduler.scala:453)
at org.apache.spark.scheduler.DAGScheduler.handleJobSubmitted(DAGScheduler.scala:983)
at org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.doOnReceive(DAGScheduler.scala:2140)
at org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.onReceive(DAGScheduler.scala:2132)
at org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.onReceive(DAGScheduler.scala:2121)
at org.apache.spark.util.EventLoop$$anon$1.run(EventLoop.scala:49)
at org.apache.spark.scheduler.DAGScheduler.runJob(DAGScheduler.scala:749)
at org.apache.spark.SparkContext.runJob(SparkContext.scala:2080)
at org.apache.spark.SparkContext.runJob(SparkContext.scala:2101)
at org.apache.spark.SparkContext.runJob(SparkContext.scala:2120)
at org.apache.spark.SparkContext.runJob(SparkContext.scala:2145)
at org.apache.spark.rdd.RDD.$anonfun$collect$1(RDD.scala:961)
at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:151)
at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:112)
at org.apache.spark.rdd.RDD.withScope(RDD.scala:366)
at org.apache.spark.rdd.RDD.collect(RDD.scala:960)
... 47 elided
```
After applying this change:
```
19/08/18 16:52:23 WARN DAGScheduler: The job 0 requires to run a barrier stage that requires 3 slots than the total number of slots(2) in the cluster currently.
19/08/18 16:52:38 WARN DAGScheduler: The job 0 requires to run a barrier stage that requires 3 slots than the total number of slots(2) in the cluster currently (Retry 1/3 failed).
19/08/18 16:52:53 WARN DAGScheduler: The job 0 requires to run a barrier stage that requires 3 slots than the total number of slots(2) in the cluster currently (Retry 2/3 failed).
19/08/18 16:53:08 WARN DAGScheduler: The job 0 requires to run a barrier stage that requires 3 slots than the total number of slots(2) in the cluster currently (Retry 3/3 failed).
org.apache.spark.scheduler.BarrierJobSlotsNumberCheckFailed: [SPARK-24819]: Barrier execution mode does not allow run a barrier stage that requires more slots than the total number of slots in the cluster currently. Please init a new cluster with more CPU cores or repartition the input RDD(s) to reduce the number of slots required to run this barrier stage.
at org.apache.spark.scheduler.DAGScheduler.checkBarrierStageWithNumSlots(DAGScheduler.scala:439)
at org.apache.spark.scheduler.DAGScheduler.createResultStage(DAGScheduler.scala:453)
at org.apache.spark.scheduler.DAGScheduler.handleJobSubmitted(DAGScheduler.scala:983)
at org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.doOnReceive(DAGScheduler.scala:2140)
at org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.onReceive(DAGScheduler.scala:2132)
at org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.onReceive(DAGScheduler.scala:2121)
at org.apache.spark.util.EventLoop$$anon$1.run(EventLoop.scala:49)
at org.apache.spark.scheduler.DAGScheduler.runJob(DAGScheduler.scala:749)
at org.apache.spark.SparkContext.runJob(SparkContext.scala:2080)
at org.apache.spark.SparkContext.runJob(SparkContext.scala:2101)
at org.apache.spark.SparkContext.runJob(SparkContext.scala:2120)
at org.apache.spark.SparkContext.runJob(SparkContext.scala:2145)
at org.apache.spark.rdd.RDD.$anonfun$collect$1(RDD.scala:961)
at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:151)
at org.apache.spark.rdd.RDDOperationScope$.withScope(RDDOperationScope.scala:112)
at org.apache.spark.rdd.RDD.withScope(RDD.scala:366)
at org.apache.spark.rdd.RDD.collect(RDD.scala:960)
... 47 elided
```
### How was this patch tested?
I tested manually using Spark Shell with following configuration and script. And then, checked log message.
```
$ bin/spark-shell --master local[2] --conf spark.scheduler.barrier.maxConcurrentTasksCheck.maxFailures=3
scala> sc.parallelize(1 to 100, sc.defaultParallelism+1).barrier.mapPartitions(identity(_)).collect
```
Closes#25487 from sarutak/barrier-exec-mode-warning-message.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
### What changes were proposed in this pull request?
Spark uses Netty 4 directly, but also includes Netty 3 only because transitive dependencies do. The dependencies (Hadoop HDFS, Zookeeper, Avro) don't seem to need this dependency as used in Spark. I think we can forcibly remove it to slim down the dependencies.
Previous attempts were blocked by its usage in Flume, but that dependency has gone away.
https://github.com/apache/spark/pull/15436
### Why are the changes needed?
Mostly to reduce the transitive dependency size and complexity a little bit and avoid triggering spurious security alerts on Netty 3.x usage.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Existing tests
Closes#25544 from srowen/SPARK-17875.
Authored-by: Sean Owen <sean.owen@databricks.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
<!--
Thanks for sending a pull request! Here are some tips for you:
1. If this is your first time, please read our contributor guidelines: https://spark.apache.org/contributing.html
2. Ensure you have added or run the appropriate tests for your PR: https://spark.apache.org/developer-tools.html
3. If the PR is unfinished, add '[WIP]' in your PR title, e.g., '[WIP][SPARK-XXXX] Your PR title ...'.
4. Be sure to keep the PR description updated to reflect all changes.
5. Please write your PR title to summarize what this PR proposes.
6. If possible, provide a concise example to reproduce the issue for a faster review.
-->
### What changes were proposed in this pull request?
<!--
Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster reviews in your PR. See the examples below.
1. If you refactor some codes with changing classes, showing the class hierarchy will help reviewers.
2. If you fix some SQL features, you can provide some references of other DBMSes.
3. If there is design documentation, please add the link.
4. If there is a discussion in the mailing list, please add the link.
-->
Dealing with interrupted exception in BarrierTaskContext.barrier()
### Why are the changes needed?
<!--
Please clarify why the changes are needed. For instance,
1. If you propose a new API, clarify the use case for a new API.
2. If you fix a bug, you can clarify why it is a bug.
-->
Interrupted exception will happen in the case sparkContext local property "spark.job.interruptOnCancel" set true.
### Does this PR introduce any user-facing change?
<!--
If yes, please clarify the previous behavior and the change this PR proposes - provide the console output, description and/or an example to show the behavior difference if possible.
If no, write 'No'.
-->
No.
### How was this patch tested?
<!--
If tests were added, say they were added here. Please make sure to add some test cases that check the changes thoroughly including negative and positive cases if possible.
If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future.
If tests were not added, please describe why they were not added and/or why it was difficult to add.
-->
UT.
Closes#25519 from WeichenXu123/barrier_fl.
Authored-by: WeichenXu <weichen.xu@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## 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?
Fix canceling a spark job using barrier mode but barrier tasks do not exit.
Currently, when spark tasks are killed, `BarrierTaskContext.barrier()` cannot be killed (it will blocking on RPC request), cause the task blocking and cannot exit.
In my PR I implement an interface for RPC which support `abort` in class `RpcEndpointRef`
```
def askAbortable[T: ClassTag](
message: Any,
timeout: RpcTimeout): AbortableRpcFuture[T]
```
The returned `AbortableRpcFuture` instance include an `abort` method so that we can abort the RPC before it timeout.
## How was this patch tested?
Unit test added.
Manually test:
### Test code
launch spark-shell via `spark-shell --master local[4]`
and run following code:
```
sc.setLogLevel("INFO")
import org.apache.spark.BarrierTaskContext
val n = 4
def taskf(iter: Iterator[Int]): Iterator[Int] = {
val context = BarrierTaskContext.get()
val x = iter.next()
if (x % 2 == 0) {
// sleep 6000000 seconds with task killed checking
for (i <- 0 until 6000000) {
Thread.sleep(1000)
if (context.isInterrupted()) {
throw new org.apache.spark.TaskKilledException()
}
}
}
context.barrier()
return Iterator.empty
}
// launch spark job, including 4 tasks, tasks 1/3 will enter `barrier()`, and tasks 0/2 will enter `sleep`
sc.parallelize((0 to n), n).barrier().mapPartitions(taskf).collect()
```
And then press Ctrl+C to exit the running job.
### Before
press Ctrl+C to exit the running job, then open spark UI we can see 2 tasks (task 1/3) are not killed. They are blocking.
### After
press Ctrl+C to exit the running job, we can see in spark UI all tasks killed successfully.
Please review https://spark.apache.org/contributing.html before opening a pull request.
Closes#25235 from WeichenXu123/sc_14848.
Authored-by: WeichenXu <weichen.xu@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
Change the logic of collecting the indeterminate stage, we should look at stages from mapStage, not failedStage during handle FetchFailed.
### Why are the changes needed?
In the fetch failed error handle logic, the original logic of collecting indeterminate stage from the fetch failed stage. And in the scenario of the fetch failed happened in the first task of this stage, this logic will cause the indeterminate stage to resubmit partially. Eventually, we are capable of getting correctness bug.
### Does this PR introduce any user-facing change?
It makes the corner case of indeterminate stage abort as expected.
### How was this patch tested?
New UT in DAGSchedulerSuite.
Run below integrated test with `local-cluster[5, 2, 5120]`, and set `spark.sql.execution.sortBeforeRepartition`=false, it will abort the indeterminate stage as expected:
```
import scala.sys.process._
import org.apache.spark.TaskContext
val res = spark.range(0, 10000 * 10000, 1).map{ x => (x % 1000, x)}
// kill an executor in the stage that performs repartition(239)
val df = res.repartition(113).map{ x => (x._1 + 1, x._2)}.repartition(239).map { x =>
if (TaskContext.get.attemptNumber == 0 && TaskContext.get.partitionId < 1 && TaskContext.get.stageAttemptNumber == 0) {
throw new Exception("pkill -f -n java".!!)
}
x
}
val r2 = df.distinct.count()
```
Closes#25498 from xuanyuanking/SPARK-28699-followup.
Authored-by: Yuanjian Li <xyliyuanjian@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
Update Jersey to 2.27+, ideally 2.29, for possible JDK 11 fixes.
## How was this patch tested?
Existing tests.
Closes#25455 from srowen/SPARK-28737.
Authored-by: Sean Owen <sean.owen@databricks.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
In pr #[24533](https://github.com/apache/spark/pull/24533/files) , it prevent retry to a removed Executor.
In my test, I can't catch exceptions from
` new OneForOneBlockFetcher(client, appId, execId, blockIds, listener,
transportConf, tempFileManager).start()`
And I check the code carefully， method **start()** will handle exception of IOException in it's retry logical, won't throw it out. until it meet maxRetry times or meet exception that is not IOException.
And if we meet the situation that when we fetch block , the executor is dead, when we rerun
`RetryingBlockFetcher.BlockFetchStarter.createAndStart()`
we may failed when we create a transport client to dead executor. it will throw a IOException.
We should catch this IOException.
### Why are the changes needed?
Old solution not comprehensive. Didn't cover more case.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Existed Unit Test
Closes#25469 from AngersZhuuuu/SPARK-27637-FLLOW-UP.
Authored-by: angerszhu <angers.zhu@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
This patch adds the binding classes to enable spark to switch dataframe output to using the S3A zero-rename committers shipping in Hadoop 3.1+. It adds a source tree into the hadoop-cloud-storage module which only compiles with the hadoop-3.2 profile, and contains a binding for normal output and a specific bridge class for Parquet (as the parquet output format requires a subclass of `ParquetOutputCommitter`.
Commit algorithms are a critical topic. There's no formal proof of correctness, but the algorithms are documented an analysed in [A Zero Rename Committer](https://github.com/steveloughran/zero-rename-committer/releases). This also reviews the classic v1 and v2 algorithms, IBM's swift committer and the one from EMRFS which they admit was based on the concepts implemented here.
Test-wise
* There's a public set of scala test suites [on github](https://github.com/hortonworks-spark/cloud-integration)
* We have run integration tests against Spark on Yarn clusters.
* This code has been shipping for ~12 months in HDP-3.x.
Closes#24970 from steveloughran/cloud/SPARK-23977-s3a-committer.
Authored-by: Steve Loughran <stevel@cloudera.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
1. PythonHadoopUtil.mapToConf generates a Configuration with loadDefaults disabled
2. merging hadoop conf in several places of PythonRDD is consistent.
## How was this patch tested?
Added a new test and existed tests
Closes#25002 from advancedxy/SPARK-28203.
Authored-by: Xianjin YE <advancedxy@gmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
This change implements a few changes to the k8s pod allocator so
that it behaves a little better when dynamic allocation is on.
(i) Allow the application to ramp up immediately when there's a
change in the target number of executors. Without this change,
scaling would only trigger when a change happened in the state of
the cluster, e.g. an executor going down, or when the periodical
snapshot was taken (default every 30s).
(ii) Get rid of pending pod requests, both acknowledged (i.e. Spark
knows that a pod is pending resource allocation) and unacknowledged
(i.e. Spark has requested the pod but the API server hasn't created it
yet), when they're not needed anymore. This avoids starting those
executors to just remove them after the idle timeout, wasting resources
in the meantime.
(iii) Re-work some of the code to avoid unnecessary logging. While not
bad without dynamic allocation, the existing logging was very chatty
when dynamic allocation was on. With the changes, all the useful
information is still there, but only when interesting changes happen.
(iv) Gracefully shut down executors when they become idle. Just deleting
the pod causes a lot of ugly logs to show up, so it's better to ask pods
to exit nicely. That also allows Spark to respect the "don't delete
pods" option when dynamic allocation is on.
Tested on a small k8s cluster running different TPC-DS workloads.
Closes#25236 from vanzin/SPARK-28487.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
In the current UI, we cannot identify which RDDs are barrier. Visualizing it will make easy to debug.
Following images are shown after this change.
The boxes in pale green mean barrier (We might need to discuss which color is proper).
## How was this patch tested?
Tested manually.
The images above are shown by following operations.
```
val rdd1 = sc.parallelize(1 to 10)
val rdd2 = sc.parallelize(1 to 10)
val rdd3 = rdd1.zip(rdd2).barrier.mapPartitions(identity(_))
val rdd4 = rdd3.map(identity(_))
val rdd5 = rdd4.reduceByKey(_+_)
rdd5.collect
```
Closes#25296 from sarutak/barrierexec-dagviz.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Xingbo Jiang <xingbo.jiang@databricks.com>
## What changes were proposed in this pull request?
By SPARK-17019, `On Heap Memory` and `Off Heap Memory` are introduced as optional metrics.
But they are not displayed because they are made `display: none` in css and there are no way to appear them.
I know #22595 also try to resolve this issue but that will use `additional-metrics.js`.
Initially, `additional-metrics.js` is created for `StagePage` but `StagePage` currently uses `stagepage.js` for its additional metrics to be toggle because `DataTable (one of jQuery plugins)` was introduced and we needed another mechanism to add/remove columns for additional metrics.
Now that `ExecutorsPage` also uses `DataTable` so it might be better to introduce same mechanism as `StagePage` for additional metrics.
And then, we can remove `additional-metrics.js` which is no longer used from anywhere.
## How was this patch tested?
After this change is applied, I confirmed `ExecutorsPage` and `StagePage` are properly rendered and all checkboxes for additional metrics work.
Closes#25374 from sarutak/remove-additional-metrics.js.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
Currently, on requesting summary metrics, cached data are returned if the current number of "SUCCESS" tasks is the same as the value in cached data.
However, the number of "SUCCESS" tasks is wrong when there are running tasks. In `AppStatusStore`, the KVStore is `ElementTrackingStore`, instead of `InMemoryStore`. The value count is always the number of "SUCCESS" tasks + "RUNNING" tasks.
Thus, even when the running tasks are finished, the out-of-update cached data is returned.
This PR is to fix the code in getting the number of "SUCCESS" tasks.
## How was this patch tested?
Test manually, run
```
sc.parallelize(1 to 160, 40).map(i => Thread.sleep(i*100)).collect()
```
and keep refreshing the stage page , we can see the task summary metrics is wrong.
### Before fix:
### After fix:
Closes#25369 from gengliangwang/fixStagePage.
Authored-by: Gengliang Wang <gengliang.wang@databricks.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
* Add log in `NewHadoopRDD`
* Remove some words in logs which related to specific user API.
## How was this patch tested?
Manual.
Please review https://spark.apache.org/contributing.html before opening a pull request.
Closes#25391 from WeichenXu123/log_sf.
Authored-by: WeichenXu <weichen.xu@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
SPARK-24817 and SPARK-24819 introduced new 3 non-internal properties for barrier-execution mode but they are not documented.
So I've added a section into configuration.md for barrier-mode execution.
## How was this patch tested?
Built using jekyll and confirm the layout by browser.
Closes#25370 from sarutak/barrier-exec-mode-conf-doc.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
In StagePage, only the first optional column (Scheduler Delay, in this case) appears even though "Select All" checkbox is checked.
The cause is that wrong method is used to manipulate multiple columns. columns should have been used but column was used.
I've fixed this issue by replacing the `column` with `columns`.
## How was this patch tested?
Confirmed behavior of the check-box.
Closes#25397 from sarutak/fix-stagepage.js.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
This PR fixed typos in comments and replace the explicit type with '<>' for Java 8+.
## How was this patch tested?
Manually tested.
Closes#25338 from younggyuchun/younggyu.
Authored-by: younggyu chun <younggyuchun@gmail.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
In high workload environments, ContextCleaner seems to have excessive logging at INFO level which do not give much information. In one Particular case we see that ``INFO ContextCleaner: Cleaned accumulator`` message is 25-30% of the generated logs. We can log this information for cleanup in DEBUG level instead.
## How was this patch tested?
This do not modify any functionality. This is just changing cleanup log levels to DEBUG for ContextCleaner
Closes#25396 from ajithme/logss.
Authored-by: Ajith <ajith2489@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
In this PR, we implements a complete process of GPU-aware resources scheduling
in Standalone. The whole process looks like: Worker sets up isolated resources
when it starts up and registers to master along with its resources. And, Master
picks up usable workers according to driver/executor's resource requirements to
launch driver/executor on them. Then, Worker launches the driver/executor after
preparing resources file, which is created under driver/executor's working directory,
with specified resource addresses(told by master). When driver/executor finished,
their resources could be recycled to worker. Finally, if a worker stops, it
should always release its resources firstly.
For the case of Workers and Drivers in **client** mode run on the same host, we introduce
a config option named `spark.resources.coordinate.enable`(default true) to indicate
whether Spark should coordinate resources for user. If `spark.resources.coordinate.enable=false`, user should be responsible for configuring different resources for Workers and Drivers when use resourcesFile or discovery script. If true, Spark would help user to assign different resources for Workers and Drivers.
The solution for Spark to coordinate resources among Workers and Drivers is:
Generally, use a shared file named *____allocated_resources____.json* to sync allocated
resources info among Workers and Drivers on the same host.
After a Worker or Driver found all resources using the configured resourcesFile and/or
discovery script during launching, it should filter out available resources by excluding resources already allocated in *____allocated_resources____.json* and acquire resources from available resources according to its own requirement. After that, it should write its allocated resources along with its process id (pid) into *____allocated_resources____.json*. Pid (proposed by tgravescs) here used to check whether the allocated resources are still valid in case of Worker or Driver crashes and doesn't release resources properly. And when a Worker or Driver finished, normally, it would always clean up its own allocated resources in *____allocated_resources____.json*.
Note that we'll always get a file lock before any access to file *____allocated_resources____.json*
and release the lock finally.
Futhermore, we appended resources info in `WorkerSchedulerStateResponse` to work
around master change behaviour in HA mode.
## How was this patch tested?
Added unit tests in WorkerSuite, MasterSuite, SparkContextSuite.
Manually tested with client/cluster mode (e.g. multiple workers) in a single node Standalone.
Closes#25047 from Ngone51/SPARK-27371.
Authored-by: wuyi <ngone_5451@163.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
## What changes were proposed in this pull request?
This patch tries to keep consistency whenever UTF-8 charset is needed, as using `StandardCharsets.UTF_8` instead of using "UTF-8". If the String type is needed, `StandardCharsets.UTF_8.name()` is used.
This change also brings the benefit of getting rid of `UnsupportedEncodingException`, as we're providing `Charset` instead of `String` whenever possible.
This also changes some private Catalyst helper methods to operate on encodings as `Charset` objects rather than strings.
## How was this patch tested?
Existing unit tests.
Closes#25335 from HeartSaVioR/SPARK-28601.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
Currently, PythonBroadcast may delete its data file while a python worker still needs it. This happens because PythonBroadcast overrides the `finalize()` method to delete its data file. So, when GC happens and no references on broadcast variable, it may trigger `finalize()` to delete
data file. That's also means, data under python Broadcast variable couldn't be deleted when `unpersist()`/`destroy()` called but relys on GC.
In this PR, we removed the `finalize()` method, and map the PythonBroadcast data file to a BroadcastBlock(which has the same broadcast id with the broadcast variable who wrapped this PythonBroadcast) when PythonBroadcast is deserializing. As a result, the data file could be deleted just like other pieces of the Broadcast variable when `unpersist()`/`destroy()` called and do not rely on GC any more.
## How was this patch tested?
Added a Python test, and tested manually(verified create/delete the broadcast block).
Closes#25262 from Ngone51/SPARK-28486.
Authored-by: wuyi <ngone_5451@163.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
After SPARK-27677, the shuffle client not only handles the shuffle block but also responsible for local persist RDD blocks. For better code scalability and precise semantics(as the [discussion](https://github.com/apache/spark/pull/24892#discussion_r300173331)), here we did several changes:
- Rename ShuffleClient to BlockStoreClient.
- Correspondingly rename the ExternalShuffleClient to ExternalBlockStoreClient, also change the server-side class from ExternalShuffleBlockHandler to ExternalBlockHandler.
- Move MesosExternalBlockStoreClient to Mesos package.
Note, we still keep the name of BlockTransferService, because the `Service` contains both client and server, also the name of BlockTransferService is not referencing shuffle client only.
## How was this patch tested?
Existing UT.
Closes#25327 from xuanyuanking/SPARK-28593.
Lead-authored-by: Yuanjian Li <xyliyuanjian@gmail.com>
Co-authored-by: Yuanjian Li <yuanjian.li@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
Add configuration spark.scheduler.listenerbus.eventqueue.${name}.capacity to allow configuration of different event queue size.
## How was this patch tested?
Unit test in core/src/test/scala/org/apache/spark/scheduler/SparkListenerSuite.scala
Closes#25307 from yunzoud/SPARK-28574.
Authored-by: yunzoud <yun.zou@databricks.com>
Signed-off-by: Shixiong Zhu <zsxwing@gmail.com>
## What changes were proposed in this pull request?
Today all registered metric sources are reported to GraphiteSink with no filtering mechanism, although the codahale project does support it.
GraphiteReporter (ScheduledReporter) from the codahale project requires you implement and supply the MetricFilter interface (there is only a single implementation by default in the codahale project, MetricFilter.ALL).
Propose to add an additional regex config to match and filter metrics to the GraphiteSink
## How was this patch tested?
Included a GraphiteSinkSuite that tests:
1. Absence of regex filter (existing default behavior maintained)
2. Presence of `regex=<regexexpr>` correctly filters metric keys
Closes#25232 from nkarpov/graphite_regex.
Authored-by: Nick Karpov <nick@nickkarpov.com>
Signed-off-by: jerryshao <jerryshao@tencent.com>
There's a small, probably very hard to hit thread-safety issue in the blacklist
abort timers in the task scheduler, where they access a non-thread-safe map without
locks.
In the tests, the code was also calling methods on the TaskSetManager without
holding the proper locks, which could cause threads to call non-thread-safe
TSM methods concurrently.
Closes#25317 from vanzin/SPARK-28584.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
Prior to this change, in an executor, on each heartbeat, memory metrics are polled and sent in the heartbeat. The heartbeat interval is 10s by default. With this change, in an executor, memory metrics can optionally be polled in a separate poller at a shorter interval.
For each executor, we use a map of (stageId, stageAttemptId) to (count of running tasks, executor metric peaks) to track what stages are active as well as the per-stage memory metric peaks. When polling the executor memory metrics, we attribute the memory to the active stage(s), and update the peaks. In a heartbeat, we send the per-stage peaks (for stages active at that time), and then reset the peaks. The semantics would be that the per-stage peaks sent in each heartbeat are the peaks since the last heartbeat.
We also keep a map of taskId to memory metric peaks. This tracks the metric peaks during the lifetime of the task. The polling thread updates this as well. At end of a task, we send the peak metric values in the task result. In case of task failure, we send the peak metric values in the `TaskFailedReason`.
We continue to do the stage-level aggregation in the EventLoggingListener.
For the driver, we still only poll on heartbeats. What the driver sends will be the current values of the metrics in the driver at the time of the heartbeat. This is semantically the same as before.
## How was this patch tested?
Unit tests. Manually tested applications on an actual system and checked the event logs; the metrics appear in the SparkListenerTaskEnd and SparkListenerStageExecutorMetrics events.
Closes#23767 from wypoon/wypoon_SPARK-26329.
Authored-by: Wing Yew Poon <wypoon@cloudera.com>
Signed-off-by: Imran Rashid <irashid@cloudera.com>
## What changes were proposed in this pull request?
Logging in driver when loading single large unsplittable file via `sc.textFile` or csv/json datasouce.
Current condition triggering logging is
* only generate one partition
* file is unsplittable, possible reason is:
- compressed by unsplittable compression algo such as gzip.
- multiLine mode in csv/json datasource
- wholeText mode in text datasource
* file size exceed the config threshold `spark.io.warning.largeFileThreshold` (default value is 1GB)
## How was this patch tested?
Manually test.
Generate one gzip file exceeding 1GB,
```
base64 -b 50 /dev/urandom | head -c 2000000000 > file1.txt
cat file1.txt | gzip > file1.gz
```
then launch spark-shell,
run
```
sc.textFile("file:///path/to/file1.gz").count()
```
Will print log like:
```
WARN HadoopRDD: Loading one large unsplittable file file:/.../f1.gz with only one partition, because the file is compressed by unsplittable compression codec
```
run
```
sc.textFile("file:///path/to/file1.txt").count()
```
Will print log like:
```
WARN HadoopRDD: Loading one large file file:/.../f1.gz with only one partition, we can increase partition numbers by the `minPartitions` argument in method `sc.textFile
```
run
```
spark.read.csv("file:///path/to/file1.gz").count
```
Will print log like:
```
WARN CSVScan: Loading one large unsplittable file file:/.../f1.gz with only one partition, the reason is: the file is compressed by unsplittable compression codec
```
run
```
spark.read.option("multiLine", true).csv("file:///path/to/file1.gz").count
```
Will print log like:
```
WARN CSVScan: Loading one large unsplittable file file:/.../f1.gz with only one partition, the reason is: the csv datasource is set multiLine mode
```
JSON and Text datasource also tested with similar cases.
Please review https://spark.apache.org/contributing.html before opening a pull request.
Closes#25134 from WeichenXu123/log_gz.
Authored-by: WeichenXu <weichen.xu@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
The issue is that the test tried to stop an existing scheduler and replace it with
a new one set up for the test. That can cause issues because both were sharing the
same RpcEnv underneath, and unregistering RpcEndpoints is actually asynchronous
(see comment in Dispatcher.unregisterRpcEndpoint). So that could lead to races where
the new scheduler tried to register before the old one was fully unregistered.
The updated test avoids the issue by using a separate RpcEnv / scheduler instance
altogether, and also avoids a misleading NPE in the test logs.
Closes#25318 from vanzin/SPARK-24352.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
When we set ```spark.history.ui.maxApplications``` to a small value, we can't get some apps from the page search.
If the url is spliced (http://localhost:18080/history/local-xxx), it can be accessed if the app has no attempt.
But in the case of multiple attempted apps, such a url cannot be accessed, and the page displays Not Found.
## How was this patch tested?
Add UT
Closes#25301 from cxzl25/hs_app_last_attempt_id.
Authored-by: sychen <sychen@ctrip.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
This PR proposes to use `AtomicReference` so that parent and child threads can access to the same file block holder.
Python UDF expressions are turned to a plan and then it launches a separate thread to consume the input iterator. In the separate child thread, the iterator sets `InputFileBlockHolder.set` before the parent does which the parent thread is unable to read later.
1. In this separate child thread, if it happens to call `InputFileBlockHolder.set` first without initialization of the parent's thread local (which is done when the `ThreadLocal.get()` is first called), the child thread seems calling its own `initialValue` to initialize.
2. After that, the parent calls its own `initialValue` to initializes at the first call of `ThreadLocal.get()`.
3. Both now have two different references. Updating at child isn't reflected to parent.
This PR fixes it via initializing parent's thread local with `AtomicReference` for file status so that they can be used in each task, and children thread's update is reflected.
I also tried to explain this a bit more at https://github.com/apache/spark/pull/24958#discussion_r297203041.
## How was this patch tested?
Manually tested and unittest was added.
Closes#24958 from HyukjinKwon/SPARK-28153.
Authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
As part of the shuffle storage API proposed in SPARK-25299, this introduces an API for persisting shuffle data in arbitrary storage systems.
This patch introduces several concepts:
* `ShuffleDataIO`, which is the root of the entire plugin tree that will be proposed over the course of the shuffle API project.
* `ShuffleExecutorComponents` - the subset of plugins for managing shuffle-related components for each executor. This will in turn instantiate shuffle readers and writers.
* `ShuffleMapOutputWriter` interface - instantiated once per map task. This provides child `ShufflePartitionWriter` instances for persisting the bytes for each partition in the map task.
The default implementation of these plugins exactly mirror what was done by the existing shuffle writing code - namely, writing the data to local disk and writing an index file. We leverage the APIs in the `BypassMergeSortShuffleWriter` only. Follow-up PRs will use the APIs in `SortShuffleWriter` and `UnsafeShuffleWriter`, but are left as future work to minimize the review surface area.
## How was this patch tested?
New unit tests were added. Micro-benchmarks indicate there's no slowdown in the affected code paths.
Closes#25007 from mccheah/spark-shuffle-writer-refactor.
Lead-authored-by: mcheah <mcheah@palantir.com>
Co-authored-by: mccheah <mcheah@palantir.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
… more efficiently
This PR improves the performance of scheduling speculative tasks to be O(1) instead of O(numSpeculativeTasks), using the same approach used for scheduling regular tasks. The performance of this method is particularly important because a lock is held on the TaskSchedulerImpl which is a bottleneck for all scheduling operations. We ran a Join query on a large dataset with speculation enabled and out of 100000 tasks for the ShuffleMapStage, the maximum number of speculatable tasks that was noted was close to 7900-8000 at a point. That is when we start seeing the bottleneck on the scheduler lock.
In particular, this works by storing a separate stack of tasks by executor, node, and rack locality preferences. Then when trying to schedule a speculative task, rather than scanning all speculative tasks to find ones which match the given executor (or node, or rack) preference, we can jump to a quick check of tasks matching the resource offer. This technique was already used for regular tasks -- this change refactors the code to allow sharing with regular and speculative task execution.
## What changes were proposed in this pull request?
Have split the main queue "speculatableTasks" into 5 separate queues based on locality preference similar to how normal tasks are enqueued. Thus, the "dequeueSpeculativeTask" method will avoid performing locality checks for each task at runtime and simply return the preferable task to be executed.
## How was this patch tested?
We ran a spark job that performed a join on a 10 TB dataset to test the code change.
Original Code:
<img width="1433" alt="screen shot 2019-01-28 at 5 07 22 pm" src="https://user-images.githubusercontent.com/22228190/51873321-572df280-2322-11e9-9149-0aae08d5edc6.png">
Optimized Code:
<img width="1435" alt="screen shot 2019-01-28 at 5 08 19 pm" src="https://user-images.githubusercontent.com/22228190/51873343-6745d200-2322-11e9-947b-2cfd0f06bcab.png">
As you can see, the run time of the ShuffleMapStage came down from 40 min to 6 min approximately, thus, reducing the overall running time of the spark job by a significant amount.
Another example for the same job:
Original Code:
<img width="1440" alt="screen shot 2019-01-28 at 5 11 30 pm" src="https://user-images.githubusercontent.com/22228190/51873355-70cf3a00-2322-11e9-9c3a-af035449a306.png">
Optimized Code:
<img width="1440" alt="screen shot 2019-01-28 at 5 12 16 pm" src="https://user-images.githubusercontent.com/22228190/51873367-7dec2900-2322-11e9-8d07-1b1b49285f71.png">
Closes#23677 from pgandhi999/SPARK-26755.
Lead-authored-by: pgandhi <pgandhi@verizonmedia.com>
Co-authored-by: pgandhi <pgandhi@oath.com>
Signed-off-by: Imran Rashid <irashid@cloudera.com>
These are what I found during working on #22282.
- Remove unused value: `UnsafeArraySuite#defaultTz`
- Remove redundant new modifier to the case class, `KafkaSourceRDDPartition`
- Remove unused variables from `RDD.scala`
- Remove trailing space from `structured-streaming-kafka-integration.md`
- Remove redundant parameter from `ArrowConvertersSuite`: `nullable` is `true` by default.
- Remove leading empty line: `UnsafeRow`
- Remove trailing empty line: `KafkaTestUtils`
- Remove unthrown exception type: `UnsafeMapData`
- Replace unused declarations: `expressions`
- Remove duplicated default parameter: `AnalysisErrorSuite`
- `ObjectExpressionsSuite`: remove duplicated parameters, conversions and unused variable
Closes#25251 from dongjinleekr/cleanup/201907.
Authored-by: Lee Dongjin <dongjin@apache.org>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
This test tries to detect correct behavior in racy code, where the event
thread is racing with the executor thread that's trying to kill the running
task.
If the event that signals the stage end arrives first, any delay in the
delivery of the message to kill the task causes the code to rapidly process
elements, and may cause the test to assert. Adding a 10ms delay in
LocalSchedulerBackend before the task kill makes the test run through
~1000 elements. A longer delay can easily cause the 10000 elements to
be processed.
Instead, by adding a small delay (10ms) in the test code that processes
elements, there's a much lower probability that the kill event will not
arrive before the end; that leaves a window of 100s for the event
to be delivered to the executor. And because each element only sleeps for
10ms, the test is not really slowed down at all.
Closes#25270 from vanzin/SPARK-28535.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
The motivation for these additional metrics is to help in troubleshooting and monitoring task execution workload when running on a cluster. Currently available metrics include executor threadpool metrics for task completed and for active tasks. The addition of threadpool taskStarted metric will allow for example to collect info on the (approximate) number of failed tasks by computing the difference thread started – (active threads + completed tasks and/or successfully finished tasks).
The proposed metric finishedTasks is also intended for this type of troubleshooting. The difference between finshedTasks and threadpool.completeTasks, is that the latter is a (dropwizard library) gauge taken from the threadpool, while the former is a (dropwizard) counter computed in the [[Executor]] class, when a task successfully finishes, together with several other task metrics counters.
Note, there are similarities with some of the metrics introduced in SPARK-24398, however there are key differences, coming from the fact that this PR concerns the executor source, therefore providing metric values per executor + metric values do not require to pass through the listerner bus in this case.
## How was this patch tested?
Manually tested on a YARN cluster
Closes#22290 from LucaCanali/AddMetricExecutorStartedTasks.
Lead-authored-by: Luca Canali <luca.canali@cern.ch>
Co-authored-by: LucaCanali <luca.canali@cern.ch>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
This would cause the timeout time to be negative, so executors would be
timed out immediately (instead of never).
I also tweaked a couple of log messages that could get pretty long when
lots of executors were active.
Added unit test (which failed without the fix).
Closes#25208 from vanzin/SPARK-28455.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
The Spark UI's stages table misrenders the input/output metrics columns when some tasks are missing input metrics. See the screenshot below for an example of the problem:
This is because those columns' are defined as
```scala
{if (hasInput(stage)) {
metricInfo(task) { m =>
...
<td>....</td>
}
}
```
where `metricInfo` renders the node returned by the closure in case metrics are defined or returns `Nil` in case metrics are not defined. If metrics are undefined then we'll fail to render the empty `<td></td>` tag, causing columns to become misaligned as shown in the screenshot.
To fix this, this patch changes this to
```scala
{if (hasInput(stage)) {
<td>{
metricInfo(task) { m =>
...
Unparsed(...)
}
}</td>
}
```
which is an idiom that's already in use for the shuffle read / write columns.
## How was this patch tested?
It isn't. I'm arguing for correctness because the modifications are consistent with rendering methods that work correctly for other columns.
Closes#25183 from JoshRosen/joshrosen/fix-task-table-with-partial-io-metrics.
Authored-by: Josh Rosen <rosenville@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
This change adds a new option that enables dynamic allocation without
the need for a shuffle service. This mode works by tracking which stages
generate shuffle files, and keeping executors that generate data for those
shuffles alive while the jobs that use them are active.
A separate timeout is also added for shuffle data; so that executors that
hold shuffle data can use a separate timeout before being removed because
of being idle. This allows the shuffle data to be kept around in case it
is needed by some new job, or allow users to be more aggressive in timing
out executors that don't have shuffle data in active use.
The code also hooks up to the context cleaner so that shuffles that are
garbage collected are detected, and the respective executors not held
unnecessarily.
Testing done with added unit tests, and also with TPC-DS workloads on
YARN without a shuffle service.
Closes#24817 from vanzin/SPARK-27963.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
ISSUE : https://issues.apache.org/jira/browse/SPARK-28106
When we use add jar in SQL, it will have three step:
- add jar to HiveClient's classloader
- HiveClientImpl.runHiveSQL("ADD JAR" + PATH)
- SessionStateBuilder.addJar
The second step seems has no impact to the whole process. Since event it failed, we still can execute.
The first step will add jar path to HiveClient's ClassLoader, then we can use the jar in HiveClientImpl
The Third Step will add this jar path to SparkContext. But expect local file path, it will call RpcServer's FileServer to add this to Env, the is you pass wrong path. it will cause error, but if you pass HDFS path or VIEWFS path, it won't check it and just add it to jar Path Map.
Then when next TaskSetManager send out Task, this path will be brought by TaskDescription. Then Executor will call updateDependencies, this method will check all jar path and file path in TaskDescription. Then error happends like below:
## How was this patch tested?
Exist Unit Test
Environment Test
Closes#24909 from AngersZhuuuu/SPARK-28106.
Lead-authored-by: Angers <angers.zhu@gamil.com>
Co-authored-by: 朱夷 <zhuyi01@corp.netease.com>
Signed-off-by: jerryshao <jerryshao@tencent.com>
When fetching delegation tokens for a proxy user, don't try to log in,
since it will fail.
Closes#25141 from vanzin/SPARK-28150.2.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
This upgraded to a newer version of Pyrolite. Most updates [1] in the newer version are for dotnot. For java, it includes a bug fix to Unpickler regarding cleaning up Unpickler memo, and support of protocol 5.
After upgrading, we can remove the fix at SPARK-27629 for the bug in Unpickler.
[1] https://github.com/irmen/Pyrolite/compare/pyrolite-4.23...master
## How was this patch tested?
Manually tested on Python 3.6 in local on existing tests.
Closes#25143 from viirya/upgrade-pyrolite.
Authored-by: Liang-Chi Hsieh <viirya@gmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
The `_prepare_for_python_RDD` method currently broadcasts a pickled command if its length is greater than the hardcoded value `1 << 20` (1M). This change sets this value as a Spark conf instead.
## How was this patch tested?
Unit tests, manual tests.
Closes#25123 from jessecai/SPARK-28355.
Authored-by: Jesse Cai <jesse.cai@databricks.com>
Signed-off-by: gatorsmile <gatorsmile@gmail.com>
## What changes were proposed in this pull request?
`SizeBasedRollingPolicy.shouldRollover` returns false when the size is equal to `rolloverSizeBytes`.
```scala
/** Should rollover if the next set of bytes is going to exceed the size limit */
def shouldRollover(bytesToBeWritten: Long): Boolean = {
logDebug(s"$bytesToBeWritten + $bytesWrittenSinceRollover > $rolloverSizeBytes")
bytesToBeWritten + bytesWrittenSinceRollover > rolloverSizeBytes
}
```
- https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/107553/testReport/org.apache.spark.util/FileAppenderSuite/rolling_file_appender___size_based_rolling__compressed_/
```
org.scalatest.exceptions.TestFailedException: 1000 was not less than 1000
```
## How was this patch tested?
Pass the Jenkins with the updated test.
Closes#25125 from dongjoon-hyun/SPARK-28357.
Authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
This PR adds two new config properties: `spark.driver.defaultJavaOptions` and `spark.executor.defaultJavaOptions`. These are intended to be set by administrators in a file of defaults for options like JVM garbage collection algorithm. Users will still set `extraJavaOptions` properties, and both sets of JVM options will be added to start a JVM (default options are prepended to extra options).
## How was this patch tested?
Existing + additional unit tests.
```
cd docs/
SKIP_API=1 jekyll build
```
Manual webpage check.
Closes#24804 from gaborgsomogyi/SPARK-23472.
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?
Add python api support and JavaSparkContext support for resources(). I needed the JavaSparkContext support for it to properly translate into python with the py4j stuff.
## How was this patch tested?
Unit tests added and manually tested in local cluster mode and on yarn.
Closes#25087 from tgravescs/SPARK-28234-python.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
`SslContextFactory` is deprecated at Jetty 9.4 and we are using `9.4.18.v20190429`. This PR aims to replace it with `SslContextFactory.Server`.
- https://www.eclipse.org/jetty/javadoc/9.4.19.v20190610/org/eclipse/jetty/util/ssl/SslContextFactory.html
- https://www.eclipse.org/jetty/javadoc/9.3.24.v20180605/org/eclipse/jetty/util/ssl/SslContextFactory.html
```
[WARNING] /Users/dhyun/APACHE/spark/core/src/main/scala/org/apache/spark/SSLOptions.scala:71:
constructor SslContextFactory in class SslContextFactory is deprecated:
see corresponding Javadoc for more information.
[WARNING] val sslContextFactory = new SslContextFactory()
[WARNING] ^
```
## How was this patch tested?
Pass the Jenkins with the existing tests.
Closes#25067 from dongjoon-hyun/SPARK-28290.
Authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
Up to now, Apache Spark maintains the given event log directory by **time** policy, `spark.history.fs.cleaner.maxAge`. However, there are two issues.
1. Some file system has a limitation on the maximum number of files in a single directory. For example, HDFS `dfs.namenode.fs-limits.max-directory-items` is 1024 * 1024 by default.
https://hadoop.apache.org/docs/r3.2.0/hadoop-project-dist/hadoop-hdfs/hdfs-default.xml
2. Spark is sometimes unable to to clean up some old log files due to permission issues (mainly, security policy).
To handle both (1) and (2), this PR aims to support an additional policy configuration for the maximum number of files in the event log directory, `spark.history.fs.cleaner.maxNum`. Spark will try to keep the number of files in the event log directory according to this policy.
## How was this patch tested?
Pass the Jenkins with a newly added test case.
Closes#25072 from dongjoon-hyun/SPARK-28294.
Authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
This PR reverts the partial bug fix in `ShuffleBlockFetcherIterator` which was introduced by https://github.com/apache/spark/pull/23638 .
The reasons:
1. It's a potential bug. After fixing `PipelinedRDD` in #23638 , the original problem was resolved.
2. The fix is incomplete according to [the discussion](https://github.com/apache/spark/pull/23638#discussion_r251869084)
We should fix the potential bug completely later.
## How was this patch tested?
existing tests
Closes#25049 from cloud-fan/revert.
Authored-by: Wenchen Fan <wenchen@databricks.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
`requestHeaderSize` is added in https://github.com/apache/spark/pull/23090 and applies to Spark + History server UI as well. Without debug log it's hard to find out on which side what configuration is used.
In this PR I've added a log message which prints out the value.
## How was this patch tested?
Manually checked log files.
Closes#25045 from gaborgsomogyi/SPARK-26118.
Authored-by: Gabor Somogyi <gabor.g.somogyi@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
When SPARK_HOME of env is set and contains a specific `spark-defaults,conf`, `org.apache.spark.util.loadDefaultSparkProperties` method may noise `system props`. So when runs `core/test` module, it is possible to fail to run `SparkConfSuite` .
It's easy to repair by setting `loadDefaults` in `SparkConf` to be false.
```
[info] - deprecated configs *** FAILED *** (79 milliseconds)
[info] 7 did not equal 4 (SparkConfSuite.scala:266)
[info] org.scalatest.exceptions.TestFailedException:
[info] at org.scalatest.Assertions.newAssertionFailedException(Assertions.scala:528)
[info] at org.scalatest.Assertions.newAssertionFailedException$(Assertions.scala:527)
[info] at org.scalatest.FunSuite.newAssertionFailedException(FunSuite.scala:1560)
[info] at org.scalatest.Assertions$AssertionsHelper.macroAssert(Assertions.scala:501)
[info] at org.apache.spark.SparkConfSuite.$anonfun$new$26(SparkConfSuite.scala:266)
[info] at org.scalatest.OutcomeOf.outcomeOf(OutcomeOf.scala:85)
[info] at org.scalatest.OutcomeOf.outcomeOf$(OutcomeOf.scala:83)
[info] at org.scalatest.OutcomeOf$.outcomeOf(OutcomeOf.scala:104)
[info] at org.scalatest.Transformer.apply(Transformer.scala:22)
[info] at org.scalatest.Transformer.apply(Transformer.scala:20)
[info] at org.scalatest.FunSuiteLike$$anon$1.apply(FunSuiteLike.scala:186)
[info] at org.apache.spark.SparkFunSuite.withFixture(SparkFunSuite.scala:149)
[info] at org.scalatest.FunSuiteLike.invokeWithFixture$1(FunSuiteLike.scala:184)
[info] at org.scalatest.FunSuiteLike.$anonfun$runTest$1(FunSuiteLike.scala:196)
[info] at org.scalatest.SuperEngine.runTestImpl(Engine.scala:289)
```
Closes#24998 from LiShuMing/SPARK-28202.
Authored-by: ShuMingLi <ming.moriarty@gmail.com>
Signed-off-by: jerryshao <jerryshao@tencent.com>
## What changes were proposed in this pull request?
This PR proposes to add `mapPartitionsInPandas` API to DataFrame by using existing `SCALAR_ITER` as below:
1. Filtering via setting the column
```python
from pyspark.sql.functions import pandas_udf, PandasUDFType
df = spark.createDataFrame([(1, 21), (2, 30)], ("id", "age"))
pandas_udf(df.schema, PandasUDFType.SCALAR_ITER)
def filter_func(iterator):
for pdf in iterator:
yield pdf[pdf.id == 1]
df.mapPartitionsInPandas(filter_func).show()
```
```
+---+---+
| id|age|
+---+---+
| 1| 21|
+---+---+
```
2. `DataFrame.loc`
```python
from pyspark.sql.functions import pandas_udf, PandasUDFType
import pandas as pd
df = spark.createDataFrame([['aa'], ['bb'], ['cc'], ['aa'], ['aa'], ['aa']], ["value"])
pandas_udf(df.schema, PandasUDFType.SCALAR_ITER)
def filter_func(iterator):
for pdf in iterator:
yield pdf.loc[pdf.value.str.contains('^a'), :]
df.mapPartitionsInPandas(filter_func).show()
```
```
+-----+
|value|
+-----+
| aa|
| aa|
| aa|
| aa|
+-----+
```
3. `pandas.melt`
```python
from pyspark.sql.functions import pandas_udf, PandasUDFType
import pandas as pd
df = spark.createDataFrame(
pd.DataFrame({'A': {0: 'a', 1: 'b', 2: 'c'},
'B': {0: 1, 1: 3, 2: 5},
'C': {0: 2, 1: 4, 2: 6}}))
pandas_udf("A string, variable string, value long", PandasUDFType.SCALAR_ITER)
def filter_func(iterator):
for pdf in iterator:
import pandas as pd
yield pd.melt(pdf, id_vars=['A'], value_vars=['B', 'C'])
df.mapPartitionsInPandas(filter_func).show()
```
```
+---+--------+-----+
| A|variable|value|
+---+--------+-----+
| a| B| 1|
| a| C| 2|
| b| B| 3|
| b| C| 4|
| c| B| 5|
| c| C| 6|
+---+--------+-----+
```
The current limitation of `SCALAR_ITER` is that it doesn't allow different length of result, which is pretty critical in practice - for instance, we cannot simply filter by using Pandas APIs but we merely just map N to N. This PR allows map N to M like flatMap.
This API mimics the way of `mapPartitions` but keeps API shape of `SCALAR_ITER` by allowing different results.
### How does this PR implement?
This PR adds mimics both `dapply` with Arrow optimization and Grouped Map Pandas UDF. At Python execution side, it reuses existing `SCALAR_ITER` code path.
Therefore, externally, we don't introduce any new type of Pandas UDF but internally we use another evaluation type code `205` (`SQL_MAP_PANDAS_ITER_UDF`).
This approach is similar with Pandas' Windows function implementation with Grouped Aggregation Pandas UDF functions - internally we have `203` (`SQL_WINDOW_AGG_PANDAS_UDF`) but externally we just share the same `GROUPED_AGG`.
## How was this patch tested?
Manually tested and unittests were added.
Closes#24997 from HyukjinKwon/scalar-udf-iter.
Authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
This ensures that tokens are always created with an empty UGI, which
allows multiple contexts to be (sequentially) started from the same JVM.
Tested with code attached to the bug, and also usual kerberos tests.
Closes#24955 from vanzin/SPARK-28150.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
To support GPU-aware scheduling in Standalone (cluster mode), Worker should have ability to setup resources(e.g. GPU/FPGA) when it starts up.
Similar as driver/executor do, Worker has two ways(resourceFile & resourceDiscoveryScript) to setup resources when it starts up. User could use `SPARK_WORKER_OPTS` to apply resource configs on Worker in the form of "-Dx=y". For example,
```
SPARK_WORKER_OPTS="-Dspark.worker.resource.gpu.amount=2 \
-Dspark.worker.resource.fpga.amount=1 \
-Dspark.worker.resource.fpga.discoveryScript=/Users/wuyi/tmp/getFPGAResources.sh \
-Dspark.worker.resourcesFile=/Users/wuyi/tmp/worker-resource-file"
```
## How was this patch tested?
Tested manually in Standalone locally:
- Worker could start up normally when no resources are configured
- Worker should fail to start up when exception threw during setup resources(e.g. unknown directory, parse fail)
- Worker could setup resources from resource file
- Worker could setup resources from discovery scripts
- Worker should setup resources from resource file & discovery scripts when both are configure.
Closes#24841 from Ngone51/dev-worker-resources-setup.
Authored-by: wuyi <ngone_5451@163.com>
Signed-off-by: Xingbo Jiang <xingbo.jiang@databricks.com>
## What changes were proposed in this pull request?
Currently with `toLocalIterator()` and `toPandas()` with Arrow enabled, if the Spark job being run in the background serving thread errors, it will be caught and sent to Python through the PySpark serializer.
This is not the ideal solution because it is only catch a SparkException, it won't handle an error that occurs in the serializer, and each method has to have it's own special handling to propagate the error.
This PR instead returns the Python Server object along with the serving port and authentication info, so that it allows the Python caller to join with the serving thread. During the call to join, the serving thread Future is completed either successfully or with an exception. In the latter case, the exception will be propagated to Python through the Py4j call.
## How was this patch tested?
Existing tests
Closes#24834 from BryanCutler/pyspark-propagate-server-error-SPARK-27992.
Authored-by: Bryan Cutler <cutlerb@gmail.com>
Signed-off-by: Bryan Cutler <cutlerb@gmail.com>
## What changes were proposed in this pull request?
At Spark 2.4.0/2.3.2/2.2.3, [SPARK-24948](https://issues.apache.org/jira/browse/SPARK-24948) delegated access permission checks to the file system, and maintains a blacklist for all event log files failed once at reading. The blacklisted log files are released back after `CLEAN_INTERVAL_S` seconds.
However, the released files whose sizes don't changes are ignored forever due to `info.fileSize < entry.getLen()` condition (previously [here](3c96937c7b (diff-a7befb99e7bd7e3ab5c46c2568aa5b3eR454)) and now at [shouldReloadLog](https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/deploy/history/FsHistoryProvider.scala#L571)) which returns `false` always when the size is the same with the existing value in `KVStore`. This is recovered only via SHS restart.
This PR aims to remove the existing entry from `KVStore` when it goes to the blacklist.
## How was this patch tested?
Pass the Jenkins with the updated test case.
Closes#24966 from dongjoon-hyun/SPARK-28157.
Authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: DB Tsai <d_tsai@apple.com>
## What changes were proposed in this pull request?
Track tasks separately for each stage attempt (instead of tracking by stage), and do NOT reset the numRunningTasks to 0 on StageCompleted.
In the case of stage retry, the `taskEnd` event from the zombie stage sometimes makes the number of `totalRunningTasks` negative, which will causes the job to get stuck.
Similar problem also exists with `stageIdToTaskIndices` & `stageIdToSpeculativeTaskIndices`.
If it is a failed `taskEnd` event of the zombie stage, this will cause `stageIdToTaskIndices` or `stageIdToSpeculativeTaskIndices` to remove the task index of the active stage, and the number of `totalPendingTasks` will increase unexpectedly.
## How was this patch tested?
unit test properly handle task end events from completed stages
Closes#24497 from cxzl25/fix_stuck_job_follow_up.
Authored-by: sychen <sychen@ctrip.com>
Signed-off-by: Imran Rashid <irashid@cloudera.com>
## What changes were proposed in this pull request?
Before this PR during fetching a disk persisted RDD block the network was always used to get the requested block content even when both the source and fetcher executor was running on the same host.
The idea to access another executor local disk files by directly reading the disk comes from the external shuffle service where the local dirs are stored for each executor (block manager).
To make this possible the following changes are done:
- `RegisterBlockManager` message is extended with the `localDirs` which is stored by the block manager master for each block manager as a new property of the `BlockManagerInfo`
- `GetLocationsAndStatus` is extended with the requester host
- `BlockLocationsAndStatus` (the reply for `GetLocationsAndStatus` message) is extended with the an option of local directories, which is filled with a local directories of a same host executor (if there is any, otherwise None is used). This is where the block content can be read from.
Shuffle blocks are out of scope of this PR: there will be a separate PR opened for that (for another Jira issue).
## How was this patch tested?
With a new unit test in `BlockManagerSuite`. See the the test prefixed by "SPARK-27622: avoid the network when block requested from same host".
Closes#24554 from attilapiros/SPARK-27622.
Authored-by: “attilapiros” <piros.attila.zsolt@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
Update doc of config `spark.driver.resourcesFile`
## How was this patch tested?
N/A
Closes#24954 from jiangxb1987/ResourceAllocation.
Authored-by: Xingbo Jiang <xingbo.jiang@databricks.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
remove the oldest checkpointed file only if next checkpoint exists.
I think this patch needs back-porting.
## How was this patch tested?
existing test
local check in spark-shell with following suite:
```
import org.apache.spark.ml.linalg.Vectors
import org.apache.spark.ml.classification.GBTClassifier
case class Row(features: org.apache.spark.ml.linalg.Vector, label: Int)
sc.setCheckpointDir("/checkpoints")
val trainingData = sc.parallelize(1 to 2426874, 256).map(x => Row(Vectors.dense(x, x + 1, x * 2 % 10), if (x % 5 == 0) 1 else 0)).toDF
val classifier = new GBTClassifier()
.setLabelCol("label")
.setFeaturesCol("features")
.setProbabilityCol("probability")
.setMaxIter(100)
.setMaxDepth(10)
.setCheckpointInterval(2)
classifier.fit(trainingData)
```
Closes#24870 from zhengruifeng/ck_update.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
Disabled negative dns caching for docker images
Improved logging on DNS resolution, convenient for slow k8s clusters
## What changes were proposed in this pull request?
Added retries when building the connection to the driver in K8s.
In some scenarios DNS reslution can take more than the timeout.
Also openjdk-8 by default has negative dns caching enabled, which means even retries may not help depending on the times.
## How was this patch tested?
This patch was tested agains an specific k8s cluster with slow response time in DNS to ensure it woks.
Closes#24702 from jlpedrosa/feature/kuberetries.
Authored-by: Jose Luis Pedrosa <jlpedrosa@gmail.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
Event logs are different from the other data in terms of the lifetime. It would be great to have a new configuration for Spark event log compression like `spark.eventLog.compression.codec` .
This PR adds this new configuration as an optional configuration. So, if `spark.eventLog.compression.codec` is not given, `spark.io.compression.codec` will be used.
## How was this patch tested?
Pass the Jenkins with the newly added test case.
Closes#24921 from dongjoon-hyun/SPARK-28118.
Authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: DB Tsai <d_tsai@apple.com>
## What changes were proposed in this pull request?
The test `SparkContextSuite.test resource scheduling under local-cluster mode` has been flaky, because it expects the size of `sc.statusTracker.getExecutorInfos` be the same as the number of executors, while the returned list contains both the driver and executors.
## How was this patch tested?
Updated existing tests.
Closes#24917 from jiangxb1987/getExecutorInfos.
Authored-by: Xingbo Jiang <xingbo.jiang@databricks.com>
Signed-off-by: Xiangrui Meng <meng@databricks.com>
## What changes were proposed in this pull request?
In a nutshell, it looks like the absence of ML / MLlib classes on the classpath causes code in KryoSerializer to throw and catch ClassNotFoundExceptions whenever instantiating a new serializer in newInstance(). This isn't a performance problem in production (since MLlib is on the classpath there) but it's a huge issue in tests and appears to account for an enormous amount of test time
We can address this problem by reducing the total number of ClassNotFoundExceptions by performing the class existence checks once and storing the results in KryoSerializer instances rather than repeating the checks on each newInstance() call.
## How was this patch tested?
The existing tests.
Authored-by: Josh Rosen <joshrosendatabricks.com>
Closes#24916 from gatorsmile/kryoException.
Lead-authored-by: Josh Rosen <rosenville@gmail.com>
Co-authored-by: gatorsmile <gatorsmile@gmail.com>
Signed-off-by: Josh Rosen <rosenville@gmail.com>
## What changes were proposed in this pull request?
This PR fixes a performance problem in environments where `lz4-java`'s native JNI libraries fail to initialize.
Spark's uses `lz4-java` for LZ4 compression. Under the hood, the `LZ4BlockInputStream` and `LZ4BlockOutputStream` constructors call `LZ4Factory.fastestInstance()`, which attempts to load JNI libraries and falls back on Java implementations in case the JNI library cannot be loaded or initialized.
If the LZ4 JNI libraries are present on the library load path (`Native.isLoaded()`) but cannot be initialized (e.g. due to breakage caused by shading) then an exception will be thrown and caught, triggering fallback to `fastestJavaInstance()` (a non-JNI implementation).
Unfortunately, the LZ4 library does not cache the fact that the JNI library failed during initialization, so every call to `LZ4Factory.fastestInstance()` re-attempts (and fails) to initialize the native code. These initialization attempts are performed in a `static synchronized` method, so exceptions from failures are thrown while holding shared monitors and this causes monitor-contention performance issues. Here's an example stack trace showing the problem:
```java
java.lang.Throwable.fillInStackTrace(Native Method)
java.lang.Throwable.fillInStackTrace(Throwable.java:783) => holding Monitor(java.lang.NoClassDefFoundError441628568})
java.lang.Throwable.<init>(Throwable.java:265)
java.lang.Error.<init>(Error.java:70)
java.lang.LinkageError.<init>(LinkageError.java:55)
java.lang.NoClassDefFoundError.<init>(NoClassDefFoundError.java:59)
shaded.net.jpountz.lz4.LZ4JNICompressor.compress(LZ4JNICompressor.java:36)
shaded.net.jpountz.lz4.LZ4Factory.<init>(LZ4Factory.java:200)
shaded.net.jpountz.lz4.LZ4Factory.instance(LZ4Factory.java:51)
shaded.net.jpountz.lz4.LZ4Factory.nativeInstance(LZ4Factory.java:84) => holding Monitor(java.lang.Class1475983836})
shaded.net.jpountz.lz4.LZ4Factory.fastestInstance(LZ4Factory.java:157)
shaded.net.jpountz.lz4.LZ4BlockOutputStream.<init>(LZ4BlockOutputStream.java:135)
org.apache.spark.io.LZ4CompressionCodec.compressedOutputStream(CompressionCodec.scala:122)
org.apache.spark.serializer.SerializerManager.wrapForCompression(SerializerManager.scala:156)
org.apache.spark.serializer.SerializerManager.wrapStream(SerializerManager.scala:131)
org.apache.spark.storage.DiskBlockObjectWriter.open(DiskBlockObjectWriter.scala:120)
org.apache.spark.storage.DiskBlockObjectWriter.write(DiskBlockObjectWriter.scala:249)
org.apache.spark.shuffle.sort.ShuffleExternalSorter.writeSortedFile(ShuffleExternalSorter.java:211)
[...]
```
To avoid this problem, this PR modifies Spark's `LZ4CompressionCodec` to call `fastestInstance()` itself and cache the result (which is safe because these factories [are thread-safe](https://github.com/lz4/lz4-java/issues/82)).
## How was this patch tested?
Existing unit tests.
Closes#24905 from JoshRosen/lz4-factory-flags.
Lead-authored-by: Josh Rosen <rosenville@gmail.com>
Co-authored-by: Josh Rosen <joshrosen@stripe.com>
Signed-off-by: Josh Rosen <rosenville@gmail.com>
## What changes were proposed in this pull request?
Continue the work from https://github.com/apache/spark/pull/24821. Refactor resource handling code to make the code more readable. Major changes:
* Moved resource-related classes to `spark.resource` from `spark`.
* Added ResourceUtils and helper classes so we don't need to directly deal with Spark conf.
* ResourceID: resource identifier and it provides conf keys
* ResourceRequest/Allocation: abstraction for requested and allocated resources
* Added `TestResourceIDs` to reference commonly used resource IDs in tests like `spark.executor.resource.gpu`.
cc: tgravescs jiangxb1987 Ngone51
## How was this patch tested?
Unit tests for added utils and existing unit tests.
Closes#24856 from mengxr/SPARK-27823.
Lead-authored-by: Xiangrui Meng <meng@databricks.com>
Co-authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Xingbo Jiang <xingbo.jiang@databricks.com>
## What changes were proposed in this pull request?
PythonRunner uses an asynchronous way, which produces elements in WriteThread but consumes elements in another thread, to execute task. When child operator, like take()/first(), does not consume all elements produced by WriteThread, task would finish before WriteThread and releases all locks on blocks. However, WriteThread would continue to produce elements by pulling elements from parent operator until it exhausts all elements. And at the time WriteThread exhausts all elements, it will try to release the corresponding block but hit a AssertionError since task has already released that lock previously.
#24542 previously fix this by catching AssertionError, so that we won't fail our executor.
However, when not using PySpark, issue still exists when user implements a custom RDD or task, which spawn a separate child thread to consume iterator from a cached parent RDD. Below is a demo which could easily reproduce the issue.
```
val rdd0 = sc.parallelize(Range(0, 10), 1).cache()
rdd0.collect()
rdd0.mapPartitions { iter =>
val t = new Thread(new Runnable {
override def run(): Unit = {
while(iter.hasNext) {
println(iter.next())
Thread.sleep(1000)
}
}
})
t.setDaemon(false)
t.start()
Iterator(0)
}.collect()
Thread.sleep(100000)
```
So, if we could prevent the separate thread from releasing lock on block when TaskContext has already completed, we won't hit this issue again.
## How was this patch tested?
Added in new unit test in RDDSuite.
Closes#24699 from Ngone51/SPARK-27666.
Authored-by: wuyi <ngone_5451@163.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
Allow Pandas UDF to take an iterator of pd.Series or an iterator of tuple of pd.Series.
Note the UDF input args will be always one iterator:
* if the udf take only column as input, the iterator's element will be pd.Series (corresponding to the column values batch)
* if the udf take multiple columns as inputs, the iterator's element will be a tuple composed of multiple `pd.Series`s, each one corresponding to the multiple columns as inputs (keep the same order). For example:
```
pandas_udf("int", PandasUDFType.SCALAR_ITER)
def the_udf(iterator):
for col1_batch, col2_batch in iterator:
yield col1_batch + col2_batch
df.select(the_udf("col1", "col2"))
```
The udf above will add col1 and col2.
I haven't add unit tests, but manually tests show it works fine. So it is ready for first pass review.
We can test several typical cases:
```
from pyspark.sql import SparkSession
from pyspark.sql.functions import pandas_udf, PandasUDFType
from pyspark.sql.functions import udf
from pyspark.taskcontext import TaskContext
df = spark.createDataFrame([(1, 20), (3, 40)], ["a", "b"])
pandas_udf("int", PandasUDFType.SCALAR_ITER)
def fi1(it):
pid = TaskContext.get().partitionId()
print("DBG: fi1: do init stuff, partitionId=" + str(pid))
for batch in it:
yield batch + 100
print("DBG: fi1: do close stuff, partitionId=" + str(pid))
pandas_udf("int", PandasUDFType.SCALAR_ITER)
def fi2(it):
pid = TaskContext.get().partitionId()
print("DBG: fi2: do init stuff, partitionId=" + str(pid))
for batch in it:
yield batch + 10000
print("DBG: fi2: do close stuff, partitionId=" + str(pid))
pandas_udf("int", PandasUDFType.SCALAR_ITER)
def fi3(it):
pid = TaskContext.get().partitionId()
print("DBG: fi3: do init stuff, partitionId=" + str(pid))
for x, y in it:
yield x + y * 10 + 100000
print("DBG: fi3: do close stuff, partitionId=" + str(pid))
pandas_udf("int", PandasUDFType.SCALAR)
def fp1(x):
return x + 1000
udf("int")
def fu1(x):
return x + 10
# test select "pandas iter udf/pandas udf/sql udf" expressions at the same time.
# Note this case the `fi1("a"), fi2("b"), fi3("a", "b")` will generate only one plan,
# and `fu1("a")`, `fp1("a")` will generate another two separate plans.
df.select(fi1("a"), fi2("b"), fi3("a", "b"), fu1("a"), fp1("a")).show()
# test chain two pandas iter udf together
# Note this case `fi2(fi1("a"))` will generate only one plan
# Also note the init stuff/close stuff call order will be like:
# (debug output following)
# DBG: fi2: do init stuff, partitionId=0
# DBG: fi1: do init stuff, partitionId=0
# DBG: fi1: do close stuff, partitionId=0
# DBG: fi2: do close stuff, partitionId=0
df.select(fi2(fi1("a"))).show()
# test more complex chain
# Note this case `fi1("a"), fi2("a")` will generate one plan,
# and `fi3(fi1_output, fi2_output)` will generate another plan
df.select(fi3(fi1("a"), fi2("a"))).show()
```
## How was this patch tested?
To be added.
Please review http://spark.apache.org/contributing.html before opening a pull request.
Closes#24643 from WeichenXu123/pandas_udf_iter.
Lead-authored-by: WeichenXu <weichen.xu@databricks.com>
Co-authored-by: Xiangrui Meng <meng@databricks.com>
Signed-off-by: Xiangrui Meng <meng@databricks.com>
## What changes were proposed in this pull request?
- Remove unused variables.
- Remove unused imports.
- Change var to val in few places.
## How was this patch tested?
Unit tests.
Closes#24857 from imback82/unused_variable.
Authored-by: Terry Kim <yuminkim@gmail.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
We're using an old-ish jQuery, 1.12.4, and should probably update for Spark 3 to keep up in general, but also to keep up with CVEs. In fact, we know of at least one resolved in only 3.4.0+ (https://nvd.nist.gov/vuln/detail/CVE-2019-11358). They may not affect Spark, but, if the update isn't painful, maybe worthwhile in order to make future 3.x updates easier.
jQuery 1 -> 2 doesn't sound like a breaking change, as 2.0 is supposed to maintain compatibility with 1.9+ (https://blog.jquery.com/2013/04/18/jquery-2-0-released/)
2 -> 3 has breaking changes: https://jquery.com/upgrade-guide/3.0/. It's hard to evaluate each one, but the most likely area for problems is in ajax(). However, our usage of jQuery (and plugins) is pretty simple.
Update jquery to 3.4.1; update jquery blockUI and mustache to latest
## How was this patch tested?
Manual testing of docs build (except R docs), worker/master UI, spark application UI.
Note: this really doesn't guarantee it works, as our tests can't test javascript, and this is merely anecdotal testing, although I clicked about every link I could find. There's a risk this breaks a minor part of the UI; it does seem to work fine in the main.
Closes#24843 from srowen/SPARK-28004.
Authored-by: Sean Owen <sean.owen@databricks.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
(Continuation of https://github.com/apache/spark/pull/19118 ; see for details)
## How was this patch tested?
Existing tests.
Closes#24863 from srowen/SPARK-21882.2.
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?
Handle the case when ParsedStatement subclass has a Map field but not of type Map[String, String].
In ParsedStatement.productIterator, `case mapArg: Map[_, _]` can match any Map type due to type erasure, thus causing `asInstanceOf[Map[String, String]]` to throw ClassCastException.
The following test reproduces the issue:
```
case class TestStatement(p: Map[String, Int]) extends ParsedStatement {
override def output: Seq[Attribute] = Nil
override def children: Seq[LogicalPlan] = Nil
}
TestStatement(Map("abc" -> 1)).toString
```
Changing the code to `case mapArg: Map[String, String]` will not help due to type erasure. As a matter of fact, compiler gives this warning:
```
Warning:(41, 18) non-variable type argument String in type pattern
scala.collection.immutable.Map[String,String] (the underlying of Map[String,String])
is unchecked since it is eliminated by erasure
case mapArg: Map[String, String] =>
```
## How was this patch tested?
Add 2 unit tests.
Closes#24800 from jzhuge/SPARK-27947.
Authored-by: John Zhuge <jzhuge@apache.org>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
Hadoop `Configuration` has an internal `properties` map which is lazily initialized. Initialization of this field, done in the private `Configuration.getProps()` method, is rather expensive because it ends up parsing XML configuration files. When cloning a `Configuration`, this `properties` field is cloned if it has been initialized.
In some cases it's possible that `sc.hadoopConfiguration` never ends up computing this `properties` field, leading to performance problems when this configuration is cloned in `SessionState.newHadoopConf()` because each cloned `Configuration` needs to re-parse configuration XML files from disk.
To avoid this problem, we can call `Configuration.size()` to trigger a call to `getProps()`, ensuring that this expensive computation is cached and re-used when cloning configurations.
I discovered this problem while performance profiling the Spark ThriftServer while running a SQL fuzzing workload.
## How was this patch tested?
Examined YourKit profiles before and after my change.
Closes#24714 from JoshRosen/fuzzing-perf-improvements.
Authored-by: Josh Rosen <rosenville@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
Change the resource config spark.{executor/driver}.resource.{resourceName}.count to .amount to allow future usage of containing both a count and a unit. Right now we only support counts - # of gpus for instance, but in the future we may want to support units for things like memory - 25G. I think making the user only have to specify a single config .amount is better then making them specify 2 separate configs of a .count and then a .unit. Change it now since its a user facing config.
Amount also matches how the spark on yarn configs are setup.
## How was this patch tested?
Unit tests and manually verified on yarn and local cluster mode
Closes#24810 from tgravescs/SPARK-27760-amount.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
This change refactors the portions of the ExecutorAllocationManager class that
track executor state into a new class, to achieve a few goals:
- make the code easier to understand
- better separate concerns (task backlog vs. executor state)
- less synchronization between event and allocation threads
- less coupling between the allocation code and executor state tracking
The executor tracking code was moved to a new class (ExecutorMonitor) that
encapsulates all the logic of tracking what happens to executors and when
they can be timed out. The logic to actually remove the executors remains
in the EAM, since it still requires information that is not tracked by the
new executor monitor code.
In the executor monitor itself, of interest, specifically, is a change in
how cached blocks are tracked; instead of polling the block manager, the
monitor now uses events to track which executors have cached blocks, and
is able to detect also unpersist events and adjust the time when the executor
should be removed accordingly. (That's the bug mentioned in the PR title.)
Because of the refactoring, a few tests in the old EAM test suite were removed,
since they're now covered by the newly added test suite. The EAM suite was
also changed a little bit to not instantiate a SparkContext every time. This
allowed some cleanup, and the tests also run faster.
Tested with new and updated unit tests, and with multiple TPC-DS workloads
running with dynamic allocation on; also some manual tests for the caching
behavior.
Closes#24704 from vanzin/SPARK-20286.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Imran Rashid <irashid@cloudera.com>
## What changes were proposed in this pull request?
Use objects in `ResourceName` to represent resource names.
## How was this patch tested?
Existing tests.
Closes#24799 from jiangxb1987/ResourceName.
Authored-by: Xingbo Jiang <xingbo.jiang@databricks.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
This PR adds support to schedule tasks with extra resource requirements (eg. GPUs) on executors with available resources. It also introduce a new method `TaskContext.resources()` so tasks can access available resource addresses allocated to them.
## How was this patch tested?
* Added new end-to-end test cases in `SparkContextSuite`;
* Added new test case in `CoarseGrainedSchedulerBackendSuite`;
* Added new test case in `CoarseGrainedExecutorBackendSuite`;
* Added new test case in `TaskSchedulerImplSuite`;
* Added new test case in `TaskSetManagerSuite`;
* Updated existing tests.
Closes#24374 from jiangxb1987/gpu.
Authored-by: Xingbo Jiang <xingbo.jiang@databricks.com>
Signed-off-by: Xiangrui Meng <meng@databricks.com>
## What changes were proposed in this pull request?
This PR targets to deduplicate hardcoded `py4j-0.10.8.1-src.zip` in order to make py4j upgrade easier.
## How was this patch tested?
N/A
Closes#24770 from HyukjinKwon/minor-py4j-dedup.
Authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
This PR targets to add an integrated test base for various UDF test cases so that Scalar UDF, Python UDF and Scalar Pandas UDFs can be tested in SBT & Maven tests.
### Problem
One of the problems we face is that: `ExtractPythonUDFs` (for Python UDF and Scalar Pandas UDF) has unevaluable expressions that always has to be wrapped with special plans. This special rule seems producing many issues, for instance, SPARK-27803, SPARK-26147, SPARK-26864, SPARK-26293, SPARK-25314 and SPARK-24721.
### Why do we have less test cases dedicated for SQL and plans with Python UDFs?
We have virtually no such SQL (or plan) dedicated tests in PySpark to catch such issues because:
- A developer should know all the analyzer, the optimizer, SQL, PySpark, Py4J and version differences in Python to write such good test cases
- To test plans, we should access to plans in JVM via Py4J which is tricky, messy and duplicates Scala test cases
- Usually we just add end-to-end test cases in PySpark therefore there are not so many dedicated examples to refer to write in PySpark
It is also a non-trivial overhead to switch test base and method (IMHO).
### How does this PR fix?
This PR adds Python UDF and Scalar Pandas UDF into our `*.sql` file based test base in runtime of SBT / Maven test cases. It generates Python-pickled instance (consisting of return type and Python native function) that is used in Python or Scalar Pandas UDF and directly brings into JVM.
After that, (we don't interact via Py4J) run the tests directly in JVM - we can just register and run Python UDF and Scalar Pandas UDF in JVM.
Currently, I only integrated this change into SQL file based testing. This is how works with test files under `udf` directory:
After the test files under 'inputs/udf' directory are detected, it creates three test cases:
- Scala UDF test case with a Scalar UDF registered named 'udf'.
- Python UDF test case with a Python UDF registered named 'udf' iff Python executable and pyspark are available.
- Scalar Pandas UDF test case with a Scalar Pandas UDF registered named 'udf' iff Python executable, pandas, pyspark and pyarrow are available.
Therefore, UDF test cases should have single input and output files but executed by three different types of UDFs.
For instance,
```sql
CREATE TEMPORARY VIEW ta AS
SELECT udf(a) AS a, udf('a') AS tag FROM t1
UNION ALL
SELECT udf(a) AS a, udf('b') AS tag FROM t2;
CREATE TEMPORARY VIEW tb AS
SELECT udf(a) AS a, udf('a') AS tag FROM t3
UNION ALL
SELECT udf(a) AS a, udf('b') AS tag FROM t4;
SELECT tb.* FROM ta INNER JOIN tb ON ta.a = tb.a AND ta.tag = tb.tag;
```
will be ran 3 times with Scalar UDF, Python UDF and Scalar Pandas UDF each.
### Appendix
Plus, this PR adds `IntegratedUDFTestUtils` which enables to test and execute Python UDF and Scalar Pandas UDFs as below:
To register Python UDF in SQL:
```scala
IntegratedUDFTestUtils.registerTestUDF(TestPythonUDF(name = "udf"), spark)
```
To register Scalar Pandas UDF in SQL:
```scala
IntegratedUDFTestUtils.registerTestUDF(TestScalarPandasUDF(name = "udf"), spark)
```
To use it in Scala API:
```scala
spark.select(expr("udf(1)").show()
```
To use it in SQL:
```scala
sql("SELECT udf(1)").show()
```
This util could be used in the future for better coverage with Scala API combinations as well.
## How was this patch tested?
Tested via the command below:
```bash
build/sbt "sql/test-only *SQLQueryTestSuite -- -z udf/udf-inner-join.sql"
```
```
[info] SQLQueryTestSuite:
[info] - udf/udf-inner-join.sql - Scala UDF (5 seconds, 47 milliseconds)
[info] - udf/udf-inner-join.sql - Python UDF (4 seconds, 335 milliseconds)
[info] - udf/udf-inner-join.sql - Scalar Pandas UDF (5 seconds, 423 milliseconds)
```
[python] unavailable:
```
[info] SQLQueryTestSuite:
[info] - udf/udf-inner-join.sql - Scala UDF (4 seconds, 577 milliseconds)
[info] - udf/udf-inner-join.sql - Python UDF is skipped because [pyton] and/or pyspark were not available. !!! IGNORED !!!
[info] - udf/udf-inner-join.sql - Scalar Pandas UDF is skipped because pyspark,pandas and/or pyarrow were not available in [pyton]. !!! IGNORED !!!
```
pyspark unavailable:
```
[info] SQLQueryTestSuite:
[info] - udf/udf-inner-join.sql - Scala UDF (4 seconds, 991 milliseconds)
[info] - udf/udf-inner-join.sql - Python UDF is skipped because [python] and/or pyspark were not available. !!! IGNORED !!!
[info] - udf/udf-inner-join.sql - Scalar Pandas UDF is skipped because pyspark,pandas and/or pyarrow were not available in [python]. !!! IGNORED !!!
```
pandas and/or pyarrow unavailable:
```
[info] SQLQueryTestSuite:
[info] - udf/udf-inner-join.sql - Scala UDF (4 seconds, 713 milliseconds)
[info] - udf/udf-inner-join.sql - Python UDF (3 seconds, 89 milliseconds)
[info] - udf/udf-inner-join.sql - Scalar Pandas UDF is skipped because pandas and/or pyarrow were not available in [python]. !!! IGNORED !!!
```
Closes#24752 from HyukjinKwon/udf-tests.
Authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
I found the docs of `spark.driver.memoryOverhead` and `spark.executor.memoryOverhead` exists a little ambiguity.
For example, the origin docs of `spark.driver.memoryOverhead` start with `The amount of off-heap memory to be allocated per driver in cluster mode`.
But `MemoryManager` also managed a memory area named off-heap used to allocate memory in tungsten mode.
So I think the description of `spark.driver.memoryOverhead` always make confused.
`spark.executor.memoryOverhead` has the same confused with `spark.driver.memoryOverhead`.
## How was this patch tested?
Exists UT.
Closes#24671 from beliefer/improve-docs-of-overhead.
Authored-by: gengjiaan <gengjiaan@360.cn>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
Add ability to map the spark resource configs spark.{executor/driver}.resource.{resourceName} to kubernetes Container builder so that we request resources (gpu,s/fpgas/etc) from kubernetes.
Note that the spark configs will overwrite any resource configs users put into a pod template.
I added a generic vendor config which is only used by kubernetes right now. I intentionally didn't put it into the kubernetes config namespace just to avoid adding more config prefixes.
I will add more documentation for this under jira SPARK-27492. I think it will be easier to do all at once to get cohesive story.
## How was this patch tested?
Unit tests and manually testing on k8s cluster.
Closes#24703 from tgravescs/SPARK-27362.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
## What changes were proposed in this pull request?
Move to json4s version 3.6.6
Add scala-xml 1.2.0
## How was this patch tested?
Pass the Jenkins
Closes#24736 from igreenfield/master.
Authored-by: Izek Greenfield <igreenfield@axiomsl.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
Change the Driver resource discovery argument for standalone mode to be a file rather then separate address configs per resource. This makes it consistent with how the Executor is doing it and makes it more flexible in the future, and it makes for less configs if you have multiple resources.
## How was this patch tested?
Unit tests and basic manually testing to make sure files were parsed properly.
Closes#24730 from tgravescs/SPARK-27835-driver-resourcesFile.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
## What changes were proposed in this pull request?
This pr wrap all `PrintWriter` with `Utils.tryWithResource` to prevent resource leak.
## How was this patch tested?
Existing test
Closes#24739 from wangyum/SPARK-27875.
Authored-by: Yuming Wang <yumwang@ebay.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
`A very little spelling mistake.`
## How was this patch tested?
No
Closes#24710 from LiShuMing/minor-spelling.
Authored-by: ShuMingLi <ming.moriarty@gmail.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
Currently system properties are not redacted. This PR fixes that, so that any credentials passed as System properties are redacted as well.
## How was this patch tested?
Manual test. Run the following and see the UI.
```
bin/spark-shell --conf 'spark.driver.extraJavaOptions=-DMYSECRET=app'
```
Closes#24733 from aaruna/27869.
Authored-by: Aaruna <aaruna.godthi@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
### Standalone HA Background
In Spark Standalone HA mode, we'll have multiple masters running at the same time. But, there's only one master leader, which actively serving scheduling requests. Once this master leader crashes, other masters would compete for the leader and only one master is guaranteed to be elected as new master leader, which would reconstruct the state from the original master leader and continute to serve scheduling requests.
### Related Issues
#2828 firstly introduces the bug of *duplicate Worker registration*, and #3447 fixed it. But there're still corner cases(see SPARK-23191 for details) where #3447 can not cover it:
* CASE 1
(1) Initially, Worker registered with Master A.
(2) After a while, the connection channel between Master A and Worker becomes inactive(e.g. due to network drop), and Worker is notified about that by calling `onDisconnected` from NettyRpcEnv
(3) When Worker invokes `onDisconnected`, then, it will attempt to reconnect to all masters(including Master A)
(4) At the meanwhile, network between Worker and Master A recover, Worker successfully register to Master A again
(5) Master A response with `RegisterWorkerFailed("Duplicate worker ID")`
(6) Worker receives that msg, exit
* CASE 2
(1) Master A lost leadership(sends `RevokedLeadership` to itself). Master B takes over and recovery everything from master A(which would register workers for the first time in Master B) and sends `MasterChanged` to Worker
(2) Before Master A receives `RevokedLeadership`, it receives a late `HeartBeat` from Worker(which had been removed in Master A due to heartbeat timeout previously), so it sends `ReconnectWorker` to worker
(3) Worker receives `MasterChanged` before `ReconnectWorker` , changing masterRef to Master B
(4) Subsequently, Worker receives `ReconnectWorker` from Master A, then it reconnects to all masters
(5) Master B receives register request again from the Worker, response with `RegisterWorkerFailed("Duplicate worker ID")`
(6) Worker receives that msg, exit
In CASE 1, it is difficult for the Worker to know Master A's state. Normally, Worker thinks Master A has already died and is impossible that Master A would response with Worker's re-connect request.
In CASE 2, we can see race condition between `RevokedLeadership` and `HeartBeat`. Actually, Master A has already been revoked leadership while processing `HeartBeat` msg. That's means the state between Master and Zookeeper could be out of sync for a while.
### Solutions
In this PR, instead of exiting Worker process when *duplicate Worker registration* happens, we suggest to log warn about it. This would be fine since Master actually perform no-op when it receives duplicate registration from a Worker. In turn, Worker could continue living with that Master normally without any side effect.
## How was this patch tested?
Tested Manually.
I followed the steps as Neeraj Gupta suggested in JIRA SPARK-23191 to reproduce the case 1.
Before this pr, Worker would be DEAD from UI.
After this pr, Worker just warn the duplicate register behavior (as you can see the second last row in log snippet below), and still be ALIVE from UI.
```
19/05/09 20:58:32 ERROR Worker: Connection to master failed! Waiting for master to reconnect...
19/05/09 20:58:32 INFO Worker: wuyi.local:7077 Disassociated !
19/05/09 20:58:32 INFO Worker: Connecting to master wuyi.local:7077...
19/05/09 20:58:32 ERROR Worker: Connection to master failed! Waiting for master to reconnect...
19/05/09 20:58:32 INFO Worker: Not spawning another attempt to register with the master, since there is an attempt scheduled already.
19/05/09 20:58:37 WARN TransportClientFactory: DNS resolution for wuyi.local/127.0.0.1:7077 took 5005 ms
19/05/09 20:58:37 INFO TransportClientFactory: Found inactive connection to wuyi.local/127.0.0.1:7077, creating a new one.
19/05/09 20:58:37 INFO TransportClientFactory: Successfully created connection to wuyi.local/127.0.0.1:7077 after 3 ms (0 ms spent in bootstraps)
19/05/09 20:58:37 WARN Worker: Duplicate registration at master spark://wuyi.local:7077
19/05/09 20:58:37 INFO Worker: Successfully registered with master spark://wuyi.local:7077
```
Closes#24569 from Ngone51/fix-worker-dup-register-error.
Authored-by: wuyi <ngone_5451@163.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
This PR exposes additional metrics in `status.api.v1.StageData`. These metrics were already computed for `LiveStage`, but they were never exposed to the user. This includes extra metrics about the JVM GC, executor (de)serialization times, shuffle reads and writes, and more.
## How was this patch tested?
Existing tests.
cc hvanhovell
Closes#24011 from tomvanbussel/SPARK-27071.
Authored-by: Tom van Bussel <tom.vanbussel@databricks.com>
Signed-off-by: herman <herman@databricks.com>
## What changes were proposed in this pull request?
As the current approach in OneForOneBlockFetcher, we reuse the OpenBlocks protocol to describe the fetch request for shuffle blocks, and it causes the extension work for shuffle fetching like #19788 and #24110 very awkward.
In this PR, we split the fetch request for shuffle blocks from OpenBlocks which named FetchShuffleBlocks. It's a loose bind with ShuffleBlockId and can easily extend by adding new fields in this protocol.
## How was this patch tested?
Existing and new added UT.
Closes#24565 from xuanyuanking/SPARK-27665.
Authored-by: Yuanjian Li <xyliyuanjian@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
There are no unit test cases for this SortShuffleWriter，so add new test cases;
## How was this patch tested?
new test cases
Closes#24080 from wangjiaochun/UtestForSortShuffleWriter.
Authored-by: 10087686 <wang.jiaochun@zte.com.cn>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
This PR aims to show Spark version at application lists of Spark History UI.
From the following, the first `Version` column is added. When the application has multiple attempts, this will show the first attempt's version number.
**COMPLETED APPLICATION LIST**
**INCOMPLETE APPLICATION LIST**
## How was this patch tested?
Manually launch Spark history server and see the UI. Please use *Private New Window (Safari)* or *New Incognito Window (Chrome)* to avoid browser caching.
```
sbin/start-history-server.sh
```
Closes#24694 from dongjoon-hyun/SPARK-27830.
Authored-by: Dongjoon Hyun <dhyun@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
## What changes were proposed in this pull request?
In the PR the config `spark.shuffle.service.fetch.rdd.enabled` default is changed to **false** to avoid breaking any compatibility with older external shuffle service installations. As external shuffle service is deployed separately and disk persisted RDD block fetching had even introduced new network messages (`RemoveBlocks` and `BlocksRemoved`) and changed the behaviour of the already existing fetching: extended it for RDD blocks.
## How was this patch tested?
With existing unit tests.
Closes#24697 from attilapiros/minor-ext-shuffle-fetch-disabled.
Authored-by: “attilapiros” <piros.attila.zsolt@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
# What changes were proposed in this pull request?
## Problem statement
An executor which has persisted blocks does not consider to be idle and this way ready to be released by dynamic allocation after the regular timeout `spark.dynamicAllocation.executorIdleTimeout` but there is separate configuration `spark.dynamicAllocation.cachedExecutorIdleTimeout` which defaults to `Integer.MAX_VALUE`. This is because releasing the executor also means losing the persisted blocks (as the metadata for individual blocks called `BlockInfo` are kept in memory) and when the RDD is referenced latter on this lost blocks will be recomputed.
On the other hand keeping the executors too long without any task to work on is also a waste of resources (as executors are reserved for the application by the resource manager).
## Solution
This PR focuses on the first part of SPARK-25888: it extends the external shuffle service with the capability to serve RDD blocks which are persisted on the local disk store by the executors. Moreover when this feature is enabled by setting the `spark.shuffle.service.fetch.rdd.enabled` config to true and a block is reported to be persisted on to disk the external shuffle service instance running on the same host as the executor is also registered (along with the reporting block manager) as a possible location for fetching it.
## Some implementation detail
Some explanation about the decisions made during the development:
- the location list to fetch a block was randomized but the groups (same host, same rack, others) order was kept. In this PR the order of groups are kept and external shuffle service added to the end of the each group.
- `BlockManagerInfo` is not introduced for external shuffle service but only a lightweight solution is taken. A hash map from `BlockId` to `BlockStatus` is introduced. A type alias would make the source more readable but I know it is discouraged. On the other hand a new class wrapping this hash map would introduce unnecessary indirection.
- when this feature is on the cleanup triggered during removing of executors (which is handled in `ExternalShuffleBlockResolver`) is modified to keep the disk persisted RDD blocks. This cleanup is triggered in standalone mode when the `spark.storage.cleanupFilesAfterExecutorExit` config is set.
- the unpersisting of an RDD is extended to use the external shuffle service for disk persisted RDD blocks when the original executor which created the blocks are already released. New block transport messages are introduced to support this: `RemoveBlocks` and `BlocksRemoved`.
# How was this patch tested?
## Unit tests
### ExternalShuffleServiceSuite
Here the complete use case is tested by the "SPARK-25888: using external shuffle service fetching disk persisted blocks" with a tiny difference: here the executor is killed manually, this way the test is a bit faster than waiting for the idle timeout.
### ExternalShuffleBlockHandlerSuite
Tests the fetching of the RDD blocks via the external shuffle service.
### BlockManagerInfoSuite
This a new suite. As the `BlockManagerInfo` behaviour depends very much on whether the external shuffle service enabled or not all the tests are executed with and without it.
### BlockManagerSuite
Tests the sorting of the block locations.
## Manually on YARN
Spark App was:
~~~scala
package com.mycompany
import org.apache.spark.rdd.RDD
import org.apache.spark.{SparkContext, SparkConf}
import org.apache.spark.storage.StorageLevel
object TestAppDiskOnlyLevel {
def main(args: Array[String]): Unit = {
val conf = new SparkConf().setAppName("test-app")
println("Attila: START")
val sc = new SparkContext(conf)
val rdd = sc.parallelize(0 until 100, 10)
.map { i =>
println(s"Attila: calculate first rdd i=$i")
Thread.sleep(1000)
i
}
rdd.persist(StorageLevel.DISK_ONLY)
rdd.count()
println("Attila: First RDD is processed, waiting for 60 sec")
Thread.sleep(60 * 1000)
println("Attila: Num executors must be 0 as executorIdleTimeout is way over")
val rdd2 = sc.parallelize(0 until 10, 1)
.map(i => (i, 1))
.persist(StorageLevel.DISK_ONLY)
rdd2.count()
println("Attila: Second RDD with one partition (only one executors must be alive)")
// reduce runs as user code to detect the empty seq (empty blocks)
println("Calling collect on the first RDD: " + rdd.collect().reduce(_ + _))
println("Attila: STOP")
}
}
~~~
I have submitted with the following configuration:
~~~bash
spark-submit --master yarn \
--conf spark.dynamicAllocation.enabled=true \
--conf spark.dynamicAllocation.executorIdleTimeout=30 \
--conf spark.dynamicAllocation.cachedExecutorIdleTimeout=90 \
--class com.mycompany.TestAppDiskOnlyLevel dyn_alloc_demo-core_2.11-0.1.0-SNAPSHOT-jar-with-dependencies.jar
~~~
Checked the result by filtering for the side effect of the task calculations:
~~~bash
[userserver ~]$ yarn logs -applicationId application_1556299359453_0001 | grep "Attila: calculate" | wc -l
WARNING: YARN_OPTS has been replaced by HADOOP_OPTS. Using value of YARN_OPTS.
19/04/26 10:31:59 INFO client.RMProxy: Connecting to ResourceManager at apiros-1.gce.company.com/172.31.115.165:8032
100
~~~
So it is only 100 task execution and not 200 (which would be the case for re-computation).
Moreover from the submit/launcher log we can see executors really stopped in between (see the new total is 0 before the last line):
~~~
[userserver ~]$ grep "Attila: Num executors must be 0" -B 2 spark-submit.log
19/04/26 10:24:27 INFO cluster.YarnScheduler: Executor 9 on apiros-3.gce.company.com killed by driver.
19/04/26 10:24:27 INFO spark.ExecutorAllocationManager: Existing executor 9 has been removed (new total is 0)
Attila: Num executors must be 0 as executorIdleTimeout is way over
~~~
[Full spark submit log](https://github.com/attilapiros/spark/files/3122465/spark-submit.log)
I have done a test also after changing the `DISK_ONLY` storage level to `MEMORY_ONLY` for the first RDD. After this change during the 60sec waiting no executor was removed.
Closes#24499 from attilapiros/SPARK-25888-final.
Authored-by: “attilapiros” <piros.attila.zsolt@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
Added the driver functionality to get the resources.
The user interface is: SparkContext.resources - I called it this to match the TaskContext.resources api proposed in the other PR. Originally it was going to be called SparkContext.getResources but changed to be consistent, if people have strong feelings I can change it.
There are 2 ways the driver can discover what resources it has.
1) user specifies a discoveryScript, this is similar to the executors and is meant for yarn and k8s where they don't tell you what you were allocated but you are running in isolated environment.
2) read the config spark.driver.resource.resourceName.addresses. The config is meant to be used with standalone mode where the Worker will have to assign what GPU addresses the Driver is allowed to use by setting that config.
When the user runs a spark application, if they want the driver to have GPU's they would specify the conf spark.driver.resource.gpu.count=X where x is the number they want. If they are running on yarn or k8s they will also have to specify the discoveryScript as specified above, if they are on standalone mode and cluster is setup properly they wouldn't have to specify anything else. We could potentially get rid of the spark.driver.resources.gpu.addresses config which is really meant to be an internal config for worker to set if the standalone mode Worker wanted to write a discoveryScript out and set that for the user. I'll wait for the jira that implements that to decide if we can remove.
- This PR also has changes to be consistent about using resourceName everywhere.
- change the config names from POSTFIX to SUFFIX to be more consistent with other areas in Spark
- Moved the config checks around a bit since now used by both executor and driver. Note those might overlap a bit with https://github.com/apache/spark/pull/24374 so we will have to figure out which one should go in first.
## How was this patch tested?
Unit tests and manually test the interface.
Closes#24615 from tgravescs/SPARK-27488.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Xiangrui Meng <meng@databricks.com>
## What changes were proposed in this pull request?
Unset InputFileBlockHolder at the end of tasks to stop the file name from leaking over to other tasks in the same thread. This happens in particular in Pyspark because of its complex threading model.
## How was this patch tested?
new pyspark test
Closes#24605 from jose-torres/fix254.
Authored-by: Jose Torres <torres.joseph.f+github@gmail.com>
Signed-off-by: Xingbo Jiang <xingbo.jiang@databricks.com>
## What changes were proposed in this pull request?
- solves the current issue with --packages in cluster mode (there is no ticket for it). Also note of some [issues](https://issues.apache.org/jira/browse/SPARK-22657) of the past here when hadoop libs are used at the spark submit side.
- supports spark.jars, spark.files, app jar.
It works as follows:
Spark submit uploads the deps to the HCFS. Then the driver serves the deps via the Spark file server.
No hcfs uris are propagated.
The related design document is [here](https://docs.google.com/document/d/1peg_qVhLaAl4weo5C51jQicPwLclApBsdR1To2fgc48/edit). the next option to add is the RSS but has to be improved given the discussion in the past about it (Spark 2.3).
## How was this patch tested?
- Run integration test suite.
- Run an example using S3:
```
./bin/spark-submit \
...
--packages com.amazonaws:aws-java-sdk:1.7.4,org.apache.hadoop:hadoop-aws:2.7.6 \
--deploy-mode cluster \
--name spark-pi \
--class org.apache.spark.examples.SparkPi \
--conf spark.executor.memory=1G \
--conf spark.kubernetes.namespace=spark \
--conf spark.kubernetes.authenticate.driver.serviceAccountName=spark-sa \
--conf spark.driver.memory=1G \
--conf spark.executor.instances=2 \
--conf spark.sql.streaming.metricsEnabled=true \
--conf "spark.driver.extraJavaOptions=-Divy.cache.dir=/tmp -Divy.home=/tmp" \
--conf spark.kubernetes.container.image.pullPolicy=Always \
--conf spark.kubernetes.container.image=skonto/spark:k8s-3.0.0 \
--conf spark.kubernetes.file.upload.path=s3a://fdp-stavros-test \
--conf spark.hadoop.fs.s3a.access.key=... \
--conf spark.hadoop.fs.s3a.impl=org.apache.hadoop.fs.s3a.S3AFileSystem \
--conf spark.hadoop.fs.s3a.fast.upload=true \
--conf spark.kubernetes.executor.deleteOnTermination=false \
--conf spark.hadoop.fs.s3a.secret.key=... \
--conf spark.files=client:///...resolv.conf \
file:///my.jar **
```
Added integration tests based on [Ceph nano](https://github.com/ceph/cn). Looks very [active](http://www.sebastien-han.fr/blog/2019/02/24/Ceph-nano-is-getting-better-and-better/).
Unfortunately minio needs hadoop >= 2.8.
Closes#23546 from skonto/support-client-deps.
Authored-by: Stavros Kontopoulos <stavros.kontopoulos@lightbend.com>
Signed-off-by: Erik Erlandson <eerlands@redhat.com>
## What changes were proposed in this pull request?
avoid hardcoded configs in `SparkConf` and `SparkSubmit` and test
## How was this patch tested?
N/A
Closes#24631 from wenxuanguan/minor-fix.
Authored-by: wenxuanguan <choose_home@126.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
The details of the PR are explored in-depth in the sub-tasks of the umbrella jira SPARK-27726.
Briefly:
1. Stop issuing asynchronous requests to cleanup elements in the tracking store when a request is already pending
2. Fix a couple of thread-safety issues (mutable state and mis-ordered updates)
3. Move Summary deletion outside of Stage deletion loop like Tasks already are
4. Reimplement multi-delete in a removeAllKeys call which allows InMemoryStore to implement it in a performant manner.
5. Some generic typing and exception handling cleanup
We see about five orders of magnitude improvement in the deletion code, which for us is the difference between a server that needs restarting daily, and one that is stable over weeks.
Unit tests for the fire-once asynchronous code and the removeAll calls in both LevelDB and InMemoryStore are supplied. It was noted that the testing code for the LevelDB and InMemoryStore is highly repetitive, and should probably be merged, but we did not attempt that in this PR.
A version of this code was run in our production 2.3.3 and we were able to sustain higher throughput without going into GC overload (which was happening on a daily basis some weeks ago).
A version of this code was also put under a purpose-built Performance Suite of tests to verify performance under both types of Store implementations for both before and after code streams and for both total and partial delete cases (this code is not included in this PR).
Closes#24616 from davidnavas/PentaBugFix.
Authored-by: David Navas <davidn@clearstorydata.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
There are several kinds of shuffle client, blockTransferService and externalShuffleClient.
For the externalShuffleClient, there are relative external shuffle service, which guarantees the shuffle block data and regardless the state of executors.
For the blockTransferService, it is used to fetch broadcast block, and fetch the shuffle data when external shuffle service is not enabled.
When fetching data by using blockTransferService, the shuffle client would connect relative executor's blockManager, so if the relative executor is dead, it would never fetch successfully.
When spark.shuffle.service.enabled is true and spark.dynamicAllocation.enabled is true, the executor will be removed while it has been idle for more than idleTimeout.
If a blockTransferService create connection to relative executor successfully, but the relative executor is removed when beginning to fetch broadcast block, it would retry (see RetryingBlockFetcher), which is Ineffective.
If the spark.shuffle.io.retryWait and spark.shuffle.io.maxRetries is big, such as 30s and 10 times, it would waste 5 minutes.
In this PR, we check whether relative executor is alive before retry.
## How was this patch tested?
Unit test.
Closes#24533 from turboFei/SPARK-27637.
Authored-by: hustfeiwang <wangfei3@corp.netease.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
Kafka related Spark parameters has to start with `spark.kafka.` and not with `spark.sql.`. Because of this I've renamed `spark.sql.kafkaConsumerCache.capacity`.
Since Kafka consumer caching is not documented I've added this also.
## How was this patch tested?
Existing + added unit test.
```
cd docs
SKIP_API=1 jekyll build
```
and manual webpage check.
Closes#24590 from gaborgsomogyi/SPARK-27687.
Authored-by: Gabor Somogyi <gabor.g.somogyi@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
This feature allows proxy servers to identify the actual request user
using a request parameter, and performs access control checks against
that user instead of the authenticated user. Impersonation is only
allowed if the authenticated user is configured as an admin.
The request parameter used ("doAs") matches the one currently used by
Knox, but it should be easy to change / customize if different proxy
servers use a different way of identifying the original user.
Tested with updated unit tests and also with a live server behind Knox.
Closes#24582 from vanzin/SPARK-27678.
Authored-by: Marcelo Vanzin <vanzin@cloudera.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
This replaces use of collection classes like `MutableList` and `ArrayStack` with workalikes that are available in 2.12, as they will be removed in 2.13. It also removes use of `.to[Collection]` as its uses was superfluous anyway. Removing `collection.breakOut` will have to wait until 2.13
## How was this patch tested?
Existing tests
Closes#24586 from srowen/SPARK-27682.
Authored-by: Sean Owen <sean.owen@databricks.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
This removes usage of `Traversable`, which is removed in Scala 2.13. This is mostly an internal change, except for the change in the `SparkConf.setAll` method. See additional comments below.
## How was this patch tested?
Existing tests.
Closes#24584 from srowen/SPARK-27680.
Authored-by: Sean Owen <sean.owen@databricks.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
The Pull Request to add datatables to stage page SPARK-21809 got merged. The search functionality in those datatables being a great improvement for searching through a large number of tasks, also performs search over the raw data rather than the formatted data displayed in the tables. It would be great if the search can happen for the formatted data as well.
## What changes were proposed in this pull request?
Added code to enable searching over displayed data in tables e.g. searching on "165.7 MiB" or "0.3 ms" will now return the search results. Also, earlier we were missing search for two columns in the task table "Shuffle Read Bytes" as well as "Shuffle Remote Reads", which I have added here.
## How was this patch tested?
Manual Tests
Closes#24419 from pgandhi999/SPARK-25719.
Authored-by: pgandhi <pgandhi@verizonmedia.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
Add in GPU and generic resource type allocation to the executors.
Note this is part of a bigger feature for gpu-aware scheduling and is just how the executor find the resources. The general flow :
- users ask for a certain set of resources, for instance number of gpus - each cluster manager has a specific way to do this.
- cluster manager allocates a container or set of resources (standalone mode)
- When spark launches the executor in that container, the executor either has to be told what resources it has or it has to auto discover them.
- Executor has to register with Driver and tell the driver the set of resources it has so the scheduler can use that to schedule tasks that requires a certain amount of each of those resources
In this pr I added configs and arguments to the executor to be able discover resources. The argument to the executor is intended to be used by standalone mode or other cluster managers that don't have isolation so that it can assign specific resources to specific executors in case there are multiple executors on a node. The argument is a file contains JSON Array of ResourceInformation objects.
The discovery script is meant to be used in an isolated environment where the executor only sees the resources it should use.
Note that there will be follow on PRs to add other parts like the scheduler part. See the epic high level jira: https://issues.apache.org/jira/browse/SPARK-24615
## How was this patch tested?
Added unit tests and manually tested.
Please review http://spark.apache.org/contributing.html before opening a pull request.
Closes#24406 from tgravescs/gpu-sched-executor-clean.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
## What changes were proposed in this pull request?
For the below three thread configuration items applied to both driver and executor,
spark.rpc.io.serverThreads
spark.rpc.io.clientThreads
spark.rpc.netty.dispatcher.numThreads,
we separate them to driver specifics and executor specifics.
spark.driver.rpc.io.serverThreads < - > spark.executor.rpc.io.serverThreads
spark.driver.rpc.io.clientThreads < - > spark.executor.rpc.io.clientThreads
spark.driver.rpc.netty.dispatcher.numThreads < - > spark.executor.rpc.netty.dispatcher.numThreads
Spark reads these specifics first and fall back to the common configurations.
## How was this patch tested?
We ran the SimpleMap app without shuffle for benchmark purpose to test Spark's scalability in HPC with omini-path NIC which has higher bandwidth than normal ethernet NIC.
Spark's base version is 2.4.0.
Spark ran in the Standalone mode. Driver was in a standalone node.
After the separation, the performance is improved a lot in 256 nodes and 512 nodes. see below test results of SimpleMapTask before and after the enhancement. You can view the tables in the [JIRA](https://issues.apache.org/jira/browse/SPARK-26632) too.
ds: spark.driver.rpc.io.serverThreads
dc: spark.driver.rpc.io.clientThreads
dd: spark.driver.rpc.netty.dispatcher.numThreads
ed: spark.executor.rpc.netty.dispatcher.numThreads
time: Overall Time (s)
old time: Overall Time without Separation (s)
**Before:**
nodes | ds | dc | dd | ed | time
-- |-- | -- | -- | -- | --
128 nodes | 8 | 8 | 8 | 8 | 108
256 nodes | 8 | 8 | 8 | 8 | 196
512 nodes | 8 | 8 | 8 | 8 | 377
**After:**
nodes | ds | dc | dd | ed | time | improvement
-- | -- | -- | -- | -- | -- | --
128 nodes | 15 | 15 | 10 | 30 | 107 | 0.9%
256 nodes | 12 | 15 | 10 | 30 | 159 | 18.8%
512 nodes | 12 | 15 | 10 | 30 | 283 | 24.9%
Closes#23560 from zjf2012/thread_conf_separation.
Authored-by: jiafu.zhang@intel.com <jiafu.zhang@intel.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
Python uses a prefetch approach to read the result from upstream and serve them in another thread, thus it's possible that if the children operator doesn't consume all the data then the Task cleanup may happen before Python side read process finishes, this in turn create a race condition that the block read locks are freed during Task cleanup and then the reader try to release the read lock it holds and find it has been released, in this case we shall hit a AssertionError.
We shall catch the AssertionError in PythonRunner and prevent this kill the Executor.
## How was this patch tested?
Hard to write a unit test case for this case, manually verified with failed job.
Closes#24542 from jiangxb1987/pyError.
Authored-by: Xingbo Jiang <xingbo.jiang@databricks.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
This fixes an error when a PySpark local iterator, for both RDD and DataFrames, goes out of scope and the connection is closed before fully consuming the iterator. The error occurs on the JVM in the serving thread, when Python closes the local socket while the JVM is writing to it. This usually happens when there is enough data to fill the socket read buffer, causing the write call to block.
Additionally, this fixes a problem when an error occurs in the Python worker and the collect job is cancelled with an exception. Previously, the Python driver was never notified of the error so the user could get a partial result (iteration until the error) and the application will continue. With this change, an error in the worker is sent to the Python iterator and is then raised.
The change here introduces a protocol for PySpark local iterators that work as follows:
1) The local socket connection is made when the iterator is created
2) When iterating, Python first sends a request for partition data as a non-zero integer
3) While the JVM local iterator over partitions has next, it triggers a job to collect the next partition
4) The JVM sends a nonzero response to indicate it has the next partition to send
5) The next partition is sent to Python and read by the PySpark deserializer
6) After sending the entire partition, an `END_OF_DATA_SECTION` is sent to Python which stops the deserializer and allows to make another request
7) When the JVM gets a request from Python but has already consumed it's local iterator, it will send a zero response to Python and both will close the socket cleanly
8) If an error occurs in the worker, a negative response is sent to Python followed by the error message. Python will then raise a RuntimeError with the message, stopping iteration.
9) When the PySpark local iterator is garbage-collected, it will read any remaining data from the current partition (this is data that has already been collected) and send a request of zero to tell the JVM to stop collection jobs and close the connection.
Steps 1, 3, 5, 6 are the same as before. Step 8 was completely missing before because errors in the worker were never communicated back to Python. The other steps add synchronization to allow for a clean closing of the socket, with a small trade-off in performance for each partition. This is mainly because the JVM does not start collecting partition data until it receives a request to do so, where before it would eagerly write all data until the socket receive buffer is full.
## How was this patch tested?
Added new unit tests for DataFrame and RDD `toLocalIterator` and tested not fully consuming the iterator. Manual tests with Python 2.7 and 3.6.
Closes#24070 from BryanCutler/pyspark-toLocalIterator-clean-stop-SPARK-23961.
Authored-by: Bryan Cutler <cutlerb@gmail.com>
Signed-off-by: Bryan Cutler <cutlerb@gmail.com>
## What changes were proposed in this pull request?
The actual implementation doesn't support multi-cluster Kafka connection with delegation token. In this PR I've added this functionality.
What this PR contains:
* New way of configuration
* Multiple delegation token obtain/store/use functionality
* Documentation
* The change works on DStreams also
## How was this patch tested?
Existing + additional unit tests.
Additionally tested on cluster.
Test scenario:
* 2 * 4 node clusters
* The 4-4 nodes are in different kerberos realms
* Cross-Realm trust between the 2 realms
* Yarn
* Kafka broker version 2.1.0
* security.protocol = SASL_SSL
* sasl.mechanism = SCRAM-SHA-512
* Artificial exceptions during processing
* Source reads from realm1 sink writes to realm2
Kafka broker settings:
* delegation.token.expiry.time.ms=600000 (10 min)
* delegation.token.max.lifetime.ms=1200000 (20 min)
* delegation.token.expiry.check.interval.ms=300000 (5 min)
Closes#24305 from gaborgsomogyi/SPARK-27294.
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?
Fixed the `spark-<version>-yarn-shuffle.jar` artifact packaging to shade the native netty libraries:
- shade the `META-INF/native/libnetty_*` native libraries when packagin
the yarn shuffle service jar. This is required as netty library loader
derives that based on shaded package name.
- updated the `org/spark_project` shade package prefix to `org/sparkproject`
(i.e. removed underscore) as the former breaks the netty native lib loading.
This was causing the yarn external shuffle service to fail
when spark.shuffle.io.mode=EPOLL
## How was this patch tested?
Manual tests
Closes#24502 from amuraru/SPARK-27610_master.
Authored-by: Adi Muraru <amuraru@adobe.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
This is a followup of https://github.com/apache/spark/pull/24375
When `TaskSetManager` skips a task because its corresponding partition is already completed by other `TaskSetManager`s, we should not consider the duration of the task that is finished by other `TaskSetManager`s to schedule the speculative tasks of this `TaskSetManager`.
## How was this patch tested?
updated test case
Closes#24485 from cloud-fan/minor.
Authored-by: Wenchen Fan <wenchen@databricks.com>
Signed-off-by: Imran Rashid <irashid@cloudera.com>
## What changes were proposed in this pull request?
In SPARK-27612, one correctness issue was reported. When protocol 4 is used to pickle Python objects, we found that unpickled objects were wrong. A temporary fix was proposed by not using highest protocol.
It was found that Opcodes.MEMOIZE was appeared in the opcodes in protocol 4. It is suspect to this issue.
A deeper dive found that Opcodes.MEMOIZE stores objects into internal map of Unpickler object. We use single Unpickler object to unpickle serialized Python bytes. Stored objects intervenes next round of unpickling, if the map is not cleared.
We has two options:
1. Continues to reuse Unpickler, but calls its close after each unpickling.
2. Not to reuse Unpickler and create new Unpickler object in each unpickling.
This patch takes option 1.
## How was this patch tested?
Passing the test added in SPARK-27612 (#24519).
Closes#24521 from viirya/SPARK-27629.
Authored-by: Liang-Chi Hsieh <viirya@gmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
This is a long standing issue which I met before and I've seen other people got trouble with it:
[test cases stuck on "local-cluster mode" of ReplSuite?](http://apache-spark-developers-list.1001551.n3.nabble.com/test-cases-stuck-on-quot-local-cluster-mode-quot-of-ReplSuite-td3086.html)
[Spark tests hang on local machine due to "testGuavaOptional" in JavaAPISuite](http://apache-spark-developers-list.1001551.n3.nabble.com/Spark-tests-hang-on-local-machine-due-to-quot-testGuavaOptional-quot-in-JavaAPISuite-tc10999.html)
When running test under local-cluster mode with wrong SPARK_HOME(spark.test.home), test just get stuck and no response forever. After looking into SPARK_WORKER_DIR, I found there's endless executor directories under it. So, this explains what happens during test getting stuck.
The whole process looks like:
1. Driver submits an app to Master and asks for N executors
2. Master inits executor state with LAUNCHING and sends `LaunchExecutor` to Worker
3. Worker receives `LaunchExecutor`, launches ExecutorRunner asynchronously and sends `ExecutorStateChanged(state=RUNNING)` to Mater immediately
4. Master receives `ExecutorStateChanged(state=RUNNING)` and reset `_retyCount` to 0.
5. ExecutorRunner throws exception during executor launching, sends `ExecutorStateChanged(state=FAILED)` to Worker, Worker forwards the msg to Master
6. Master receives `ExecutorStateChanged(state=FAILED)`. Since Master always reset `_retyCount` when it receives RUNNING msg, so, event if a Worker fails to launch executor for continuous many times, ` _retryCount` would never exceed `maxExecutorRetries`. So, Master continue to launch executor and fall into the dead loop.
The problem exists in step 3. Worker sends `ExecutorStateChanged(state=RUNNING)` to Master immediately while executor is still launching. And, when Master receive that msg, it believes the executor has launched successfully, and reset `_retryCount` subsequently. However, that's not true.
This pr suggests to remove step 3 and requires Worker only send `ExecutorStateChanged(state=RUNNING)` after executor has really launched successfully.
## How was this patch tested?
Tested Manually.
Closes#24408 from Ngone51/fix-dead-loop.
Authored-by: wuyi <ngone_5451@163.com>
Signed-off-by: Xingbo Jiang <xingbo.jiang@databricks.com>
## What changes were proposed in this pull request?
Update taskName in PythonRunner so it keeps align with that in Executor.
## How was this patch tested?
N/A
Closes#24510 from jiangxb1987/pylog.
Authored-by: Xingbo Jiang <xingbo.jiang@databricks.com>
Signed-off-by: Xingbo Jiang <xingbo.jiang@databricks.com>
## What changes were proposed in this pull request?
This patch fixes a bug which YARN file-related configurations are being overwritten when there're some values to assign - e.g. if `--file` is specified as an argument, `spark.yarn.dist.files` is overwritten with the value of argument. After this patch the existing value and new value will be merged before assigning to the value of configuration.
## How was this patch tested?
Added UT, and manually tested with below command:
> ./bin/spark-submit --verbose --files /etc/spark2/conf/spark-defaults.conf.template --master yarn-cluster --class org.apache.spark.examples.SparkPi examples/jars/spark-examples_2.11-2.4.0.jar 10
where the spark conf file has
`spark.yarn.dist.files=file:/etc/spark2/conf/atlas-application.properties.yarn#atlas-application.properties`
```
Spark config:
...
(spark.yarn.dist.files,file:/etc/spark2/conf/atlas-application.properties.yarn#atlas-application.properties,file:///etc/spark2/conf/spark-defaults.conf.template)
...
```
Closes#24465 from HeartSaVioR/SPARK-27575.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan@gmail.com>
Signed-off-by: Marcelo Vanzin <vanzin@cloudera.com>
## What changes were proposed in this pull request?
There is a MemorySink v2 already so v1 can be removed. In this PR I've removed it completely.
What this PR contains:
* V1 memory sink removal
* V2 memory sink renamed to become the only implementation
* Since DSv2 sends exceptions in a chained format (linking them with cause field) I've made python side compliant
* Adapted all the tests
## How was this patch tested?
Existing unit tests.
Closes#24403 from gaborgsomogyi/SPARK-23014.
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 PR avoids usage of reflective calls in Scala. It removes the import that suppresses the warnings and rewrites code in small ways to avoid accessing methods that aren't technically accessible.
## How was this patch tested?
Existing tests.
Closes#24463 from srowen/SPARK-27571.
Authored-by: Sean Owen <sean.owen@databricks.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
I want to get rid of as much use of `scala.language.existentials` as possible for 3.0. It's a complicated language feature that generates warnings unless this value is imported. It might even be on the way out of Scala: https://contributors.scala-lang.org/t/proposal-to-remove-existential-types-from-the-language/2785
For Spark, it comes up mostly where the code plays fast and loose with generic types, not the advanced situations you'll often see referenced where this feature is explained. For example, it comes up in cases where a function returns something like `(String, Class[_])`. Scala doesn't like matching this to any other instance of `(String, Class[_])` because doing so requires inferring the existence of some type that satisfies both. Seems obvious if the generic type is a wildcard, but, not technically something Scala likes to let you get away with.
This is a large PR, and it only gets rid of _most_ instances of `scala.language.existentials`. The change should be all compile-time and shouldn't affect APIs or logic.
Many of the changes simply touch up sloppiness about generic types, making the known correct value explicit in the code.
Some fixes involve being more explicit about the existence of generic types in methods. For instance, `def foo(arg: Class[_])` seems innocent enough but should really be declared `def foo[T](arg: Class[T])` to let Scala select and fix a single type when evaluating calls to `foo`.
For kind of surprising reasons, this comes up in places where code evaluates a tuple of things that involve a generic type, but is OK if the two parts of the tuple are evaluated separately.
One key change was altering `Utils.classForName(...): Class[_]` to the more correct `Utils.classForName[T](...): Class[T]`. This caused a number of small but positive changes to callers that otherwise had to cast the result.
In several tests, `Dataset[_]` was used where `DataFrame` seems to be the clear intent.
Finally, in a few cases in MLlib, the return type `this.type` was used where there are no subclasses of the class that uses it. This really isn't needed and causes issues for Scala reasoning about the return type. These are just changed to be concrete classes as return types.
After this change, we have only a few classes that still import `scala.language.existentials` (because modifying them would require extensive rewrites to fix) and no build warnings.
## How was this patch tested?
Existing tests.
Closes#24431 from srowen/SPARK-27536.
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?
https://issues.apache.org/jira/browse/SPARK-25250 reports a bug that, a task which is failed with `CommitDeniedException` gets retried many times.
This can happen when a stage has 2 task set managers, one is zombie, one is active. A task from the zombie TSM completes, and commits to a central coordinator(assuming it's a file writing task). Then the corresponding task from the active TSM will fail with `CommitDeniedException`. `CommitDeniedException.countTowardsTaskFailures` is false, so the active TSM will keep retrying this task, until the job finishes. This wastes resource a lot.
#21131 firstly implements that a previous successful completed task from zombie `TaskSetManager` could mark the task of the same partition completed in the active `TaskSetManager`. Later #23871 improves the implementation to cover a corner case that, an active `TaskSetManager` hasn't been created when a previous task succeed.
However, #23871 has a bug and was reverted in #24359. With hindsight, #23781 is fragile because we need to sync the states between `DAGScheduler` and `TaskScheduler`, about which partitions are completed.
This PR proposes a new fix:
1. When `DAGScheduler` gets a task success event from an earlier attempt, notify the `TaskSchedulerImpl` about it
2. When `TaskSchedulerImpl` knows a partition is already completed, ask the active `TaskSetManager` to mark the corresponding task as finished, if the task is not finished yet.
This fix covers the corner case, because:
1. If `DAGScheduler` gets the task completion event from zombie TSM before submitting the new stage attempt, then `DAGScheduler` knows that this partition is completed, and it will exclude this partition when creating task set for the new stage attempt. See `DAGScheduler.submitMissingTasks`
2. If `DAGScheduler` gets the task completion event from zombie TSM after submitting the new stage attempt, then the active TSM is already created.
Compared to the previous fix, the message loop becomes longer, so it's likely that, the active task set manager has already retried the task multiple times. But this failure window won't be too big, and we want to avoid the worse case that retries the task many times until the job finishes. So this solution is acceptable.
## How was this patch tested?
a new test case.
Closes#24375 from cloud-fan/fix2.
Authored-by: Wenchen Fan <wenchen@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
This patch makes several test flakiness fixes.
## How was this patch tested?
N/A
Closes#24434 from gatorsmile/fixFlakyTest.
Lead-authored-by: gatorsmile <gatorsmile@gmail.com>
Co-authored-by: Hyukjin Kwon <gurwls223@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
