### What changes were proposed in this pull request?
Only calculate `executorKillTime` in `TaskSetManager.executorDecommission()` when speculation is enabled.
### Why are the changes needed?
Avoid unnecessary operations to save time/memory.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Pass existed tests.
Closes#29464 from Ngone51/followup-SPARK-21040.
Authored-by: yi.wu <yi.wu@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
The DecommissionWorkerSuite started becoming flaky and it revealed a real regression. Recently closed#29211 necessitates remembering the decommissioning shortly beyond the removal of the executor.
In addition to fixing this issue, ensure that DecommissionWorkerSuite continues to pass when executors haven't had a chance to exit eagery. That is the old behavior before #29211 also still works.
Added some more tests to TaskSchedulerImpl to ensure that the decommissioning information is indeed purged after a timeout.
Hardened the test DecommissionWorkerSuite to make it wait for successful job completion.
### Why are the changes needed?
First, let me describe the intended behavior of decommissioning: If a fetch failure happens where the source executor was decommissioned, we want to treat that as an eager signal to clear all shuffle state associated with that executor. In addition if we know that the host was decommissioned, we want to forget about all map statuses from all other executors on that decommissioned host. This is what the test "decommission workers ensure that fetch failures lead to rerun" is trying to test. This invariant is important to ensure that decommissioning a host does not lead to multiple fetch failures that might fail the job. This fetch failure can happen before the executor is truly marked "lost" because of heartbeat delays.
- However, #29211 eagerly exits the executors when they are done decommissioning. This removal of the executor was racing with the fetch failure. By the time the fetch failure is triggered the executor is already removed and thus has forgotten its decommissioning information. (I tested this by delaying the decommissioning). The fix is to keep the decommissioning information around for some time after removal with some extra logic to finally purge it after a timeout.
- In addition the executor loss can also bump up `shuffleFileLostEpoch` (added in #28848). This happens because when the executor is lost, it forgets the shuffle state about just that executor and increments the `shuffleFileLostEpoch`. This incrementing precludes the clearing of state of the entire host when the fetch failure happens because the failed task is still reusing the old epoch. The fix here is also simple: Ignore the `shuffleFileLostEpoch` when the shuffle status is being cleared due to a fetch failure resulting from host decommission.
I am strategically making both of these fixes be very local to decommissioning to avoid other regressions. Especially the version stuff is tricky (it hasn't been fundamentally changed since it was first introduced in 2013).
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Manually ran DecommissionWorkerSuite several times using a script and ensured it all passed.
### (Internal) Configs added
I added two configs, one of which is sort of meant for testing only:
- `spark.test.executor.decommission.initial.sleep.millis`: Initial delay by the decommissioner shutdown thread. Default is same as before of 1 second. This is used for testing only. This one is kept "hidden" (ie not added as a constant to avoid config bloat)
- `spark.executor.decommission.removed.infoCacheTTL`: Number of seconds to keep the removed executors decom entries around. It defaults to 5 minutes. It should be around the average time it takes for all of the shuffle data to be fetched from the mapper to the reducer, but I think that can take a while since the reducers also do a multistep sort.
Closes#29422 from agrawaldevesh/decom_fixes.
Authored-by: Devesh Agrawal <devesh.agrawal@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
Add support for full outer join inside shuffled hash join. Currently if the query is a full outer join, we only use sort merge join as the physical operator. However it can be CPU and IO intensive in case input table is large for sort merge join. Shuffled hash join on the other hand saves the sort CPU and IO compared to sort merge join, especially when table is large.
This PR implements the full outer join as followed:
* Process rows from stream side by looking up hash relation, and mark the matched rows from build side by:
* for joining with unique key, a `BitSet` is used to record matched rows from build side (`key index` to represent each row)
* for joining with non-unique key, a `HashSet[Long]` is used to record matched rows from build side (`key index` + `value index` to represent each row).
`key index` is defined as the index into key addressing array `longArray` in `BytesToBytesMap`.
`value index` is defined as the iterator index of values for same key.
* Process rows from build side by iterating hash relation, and filter out rows from build side being looked up already (done in `ShuffledHashJoinExec.fullOuterJoin`)
For context, this PR was originally implemented as followed (up to commit e3322766d4):
1. Construct hash relation from build side, with extra boolean value at the end of row to track look up information (done in `ShuffledHashJoinExec.buildHashedRelation` and `UnsafeHashedRelation.apply`).
2. Process rows from stream side by looking up hash relation, and mark the matched rows from build side be looked up (done in `ShuffledHashJoinExec.fullOuterJoin`).
3. Process rows from build side by iterating hash relation, and filter out rows from build side being looked up already (done in `ShuffledHashJoinExec.fullOuterJoin`).
See discussion of pros and cons between these two approaches [here](https://github.com/apache/spark/pull/29342#issuecomment-672275450), [here](https://github.com/apache/spark/pull/29342#issuecomment-672288194) and [here](https://github.com/apache/spark/pull/29342#issuecomment-672640531).
TODO: codegen for full outer shuffled hash join can be implemented in another followup PR.
### Why are the changes needed?
As implementation in this PR, full outer shuffled hash join will have overhead to iterate build side twice (once for building hash map, and another for outputting non-matching rows), and iterate stream side once. However, full outer sort merge join needs to iterate both sides twice, and sort the large table can be more CPU and IO intensive. So full outer shuffled hash join can be more efficient than sort merge join when stream side is much more larger than build side.
For example query below, full outer SHJ saved 30% wall clock time compared to full outer SMJ.
```
def shuffleHashJoin(): Unit = {
val N: Long = 4 << 22
withSQLConf(
SQLConf.SHUFFLE_PARTITIONS.key -> "2",
SQLConf.AUTO_BROADCASTJOIN_THRESHOLD.key -> "20000000") {
codegenBenchmark("shuffle hash join", N) {
val df1 = spark.range(N).selectExpr(s"cast(id as string) as k1")
val df2 = spark.range(N / 10).selectExpr(s"cast(id * 10 as string) as k2")
val df = df1.join(df2, col("k1") === col("k2"), "full_outer")
df.noop()
}
}
}
```
```
Running benchmark: shuffle hash join
Running case: shuffle hash join off
Stopped after 2 iterations, 16602 ms
Running case: shuffle hash join on
Stopped after 5 iterations, 31911 ms
Java HotSpot(TM) 64-Bit Server VM 1.8.0_181-b13 on Mac OS X 10.15.4
Intel(R) Core(TM) i9-9980HK CPU 2.40GHz
shuffle hash join: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
shuffle hash join off 7900 8301 567 2.1 470.9 1.0X
shuffle hash join on 6250 6382 95 2.7 372.5 1.3X
```
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Added unit test in `JoinSuite.scala`, `AbstractBytesToBytesMapSuite.java` and `HashedRelationSuite.scala`.
Closes#29342 from c21/full-outer-shj.
Authored-by: Cheng Su <chengsu@fb.com>
Signed-off-by: Takeshi Yamamuro <yamamuro@apache.org>
### What changes were proposed in this pull request?
This PR changes Executor to load jars and files added by --jars and --files on Executor initialization.
To avoid downloading those jars/files twice, they are assosiated with `startTime` as their uploaded timestamp.
### Why are the changes needed?
ExecutorPlugin can't work with Standalone Cluster and Kubernetes
when a jar which contains plugins and files used by the plugins are added by --jars and --files option with spark-submit.
This is because jars and files added by --jars and --files are not loaded on Executor initialization.
I confirmed it works with YARN because jars/files are distributed as distributed cache.
### Does this PR introduce _any_ user-facing change?
Yes. jars/files added by --jars and --files are downloaded on each executor on initialization.
### How was this patch tested?
Added a new testcase.
Closes#28939 from sarutak/fix-plugin-issue.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Mridul Muralidharan <mridul<at>gmail.com>
### What changes were proposed in this pull request?
If graceful decommissioning is enabled, Spark's dynamic scaling uses this instead of directly killing executors.
### Why are the changes needed?
When scaling down Spark we should avoid triggering recomputes as much as possible.
### Does this PR introduce _any_ user-facing change?
Hopefully their jobs run faster or at the same speed. It also enables experimental shuffle service free dynamic scaling when graceful decommissioning is enabled (using the same code as the shuffle tracking dynamic scaling).
### How was this patch tested?
For now I've extended the ExecutorAllocationManagerSuite for both core & streaming.
Closes#29367 from holdenk/SPARK-31198-use-graceful-decommissioning-as-part-of-dynamic-scaling.
Lead-authored-by: Holden Karau <hkarau@apple.com>
Co-authored-by: Holden Karau <holden@pigscanfly.ca>
Signed-off-by: Holden Karau <hkarau@apple.com>
### What changes were proposed in this pull request?
Set more dispatcher threads for the flaky test.
### Why are the changes needed?
When running test on Github Actions machine, the available processors in JVM is only 2, while on Jenkins it's 32. For this specific test, 2 available processors, which also decides the number of threads in Dispatcher, are not enough to consume the messages. In the worst situation, `MockExecutorLaunchFailWorker` would occupy these 2 threads for handling messages `LaunchDriver`, `LaunchExecutor` at the same time but leave no thread for the driver to handle the message `RegisteredApplication`. At the end, it results in a deadlock situation and causes the test failure.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
We can check whether the test is still flaky in Github Actions after this fix.
Closes#29408 from Ngone51/spark-32250.
Authored-by: yi.wu <yi.wu@databricks.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
<!--
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.
7. If you want to add a new configuration, please read the guideline first for naming configurations in
'core/src/main/scala/org/apache/spark/internal/config/ConfigEntry.scala'.
-->
### What changes were proposed in this pull request?
Clear Ivy resolution files as part of finally block if not failures while artifacts resolution can leave the resolution files around.
Use tempIvyPath for SparkSubmitUtils.buildIvySettings in tests. This why the test
"SPARK-10878: test resolution files cleaned after resolving artifact" did not capture these issues.
### Why are the changes needed?
This is a bug
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Existing unit tests
Closes#29411 from venkata91/SPARK-32596.
Authored-by: Venkata krishnan Sowrirajan <vsowrirajan@linkedin.com>
Signed-off-by: Mridul Muralidharan <mridul<at>gmail.com>
### What changes were proposed in this pull request?
This PR removes the total task result size check for shuffle map stage tasks, as these tasks return map status and metrics, which will not be cached on the driver and thus will not crash the driver.
### Why are the changes needed?
Checking total task result size for shuffle map stage tasks would lead to erroring normal jobs which create a big number of tasks even if the job eventually does not return a large dataset.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Added UT.
Closes#29276 from maryannxue/spark-32470.
Authored-by: Maryann Xue <maryann.xue@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
This PR aims to add `StorageLevel.DISK_ONLY_3` as a built-in `StorageLevel`.
### Why are the changes needed?
In a YARN cluster, HDFS uaually provides storages with replication factor 3. So, we can save the result to HDFS to get `StorageLevel.DISK_ONLY_3` technically. However, disaggregate clusters or clusters without storage services are rising. Previously, in that situation, the users were able to use similar `MEMORY_AND_DISK_2` or a user-created `StorageLevel`. This PR aims to support those use cases officially for better UX.
### Does this PR introduce _any_ user-facing change?
Yes. This provides a new built-in option.
### How was this patch tested?
Pass the GitHub Action or Jenkins with the revised test cases.
Closes#29331 from dongjoon-hyun/SPARK-32517.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
As reported by HyukjinKwon, BlockManagerDecommissionIntegrationSuite test is apparently still flaky (even after https://github.com/apache/spark/pull/29226): https://github.com/apache/spark/pull/29226#issuecomment-670286829.
The new flakyness is because the executors are not launching in the 6 seconds time out I had given them when run under github checks.
Bumped up the timeouts.
### Why are the changes needed?
To make this test not flaky so that it can give us high signal if decommissioning regresses.
### Does this PR introduce _any_ user-facing change?
No, unit test only check.
### How was this patch tested?
No new tests. Just github and jenkins.
Closes#29388 from agrawaldevesh/more_bm_harden.
Authored-by: Devesh Agrawal <devesh.agrawal@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR adds a try catch wrapping the History server scan logic to log and swallow the exception per entry.
### Why are the changes needed?
As discussed in #29350 , one entry failure shouldn't affect others.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Manually tested.
Closes#29374 from yanxiaole/SPARK-32557.
Authored-by: Yan Xiaole <xiaole.yan@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR proposes to change the behavior of DataTable for stage-page and executors-page not to save the previous search text.
### Why are the changes needed?
DataTable is used in stage-page and executors-page for pagination and filter tasks/executors by search text.
In the current implementation, search text is saved so if we visit stage-page for a job, the previous search text is filled in the textbox and the task table is filtered.
I'm sometimes surprised by this behavior as the stage-page lists no tasks because tasks are filtered by the previous search text.
I think, it's not useful.
### Does this PR introduce _any_ user-facing change?
Yes. Search text is no longer saved.
### How was this patch tested?
New testcase with the following command.
```
$ build/sbt -Dguava.version=27.0-jre -Dtest.default.exclude.tags= -Dspark.test.webdriver.chrome.driver=/path/to/chromedriver "testOnly org.apache.spark.ui.ChromeUISeleniumSuite -- -z Search"
```
Closes#29265 from sarutak/fix-search-box.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR fixes some typos in <code>core/src/main/scala/org/apache/spark/ExecutorAllocationManager.scala</code> file.
### Why are the changes needed?
<code>spark.dynamicAllocation.sustainedSchedulerBacklogTimeout</code> (N) is used only after the <code>spark.dynamicAllocation.schedulerBacklogTimeout</code> (M) is exceeded.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
No test needed.
Closes#29351 from JoeyValentine/master.
Authored-by: JoeyValentine <rlaalsdn0506@naver.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR lets the logger log timestamp based on local time zone during test.
`SparkFunSuite` fixes the default time zone to America/Los_Angeles so the timestamp logged in unit-tests.log is also based on the fixed time zone.
### Why are the changes needed?
It's confusable for developers whose time zone is not America/Los_Angeles.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Run existing tests and confirmed uint-tests.log.
If your local time zone is America/Los_Angeles, you can test by setting the environment variable `TZ` like as follows.
```
$ TZ=Asia/Tokyo build/sbt "testOnly org.apache.spark.executor.ExecutorSuite"
$ tail core/target/unit-tests.log
```
Closes#29356 from sarutak/fix-unit-test-log-timezone.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Takeshi Yamamuro <yamamuro@apache.org>
### What changes were proposed in this pull request?
Disallow `FileSystem.get(Configuration conf)` in Scala style check by default and suggest developers use `FileSystem.get(URI uri, Configuration conf)` or `Path.getFileSystem()` instead.
### Why are the changes needed?
The method `FileSystem.get(Configuration conf)` will return a default FileSystem instance if the conf `fs.file.impl` is not set. This can cause file not found exception on reading a target path of non-default file system, e.g. S3. It is hard to discover such a mistake via unit tests.
If we disallow it in Scala style check by default and suggest developers use `FileSystem.get(URI uri, Configuration conf)` or `Path.getFileSystem(Configuration conf)`, we can reduce potential regression and PR review effort.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Manually run scala style check and test.
Closes#29357 from gengliangwang/newStyleRule.
Authored-by: Gengliang Wang <gengliang.wang@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
1. Make `CoarseGrainedSchedulerBackend.maxNumConcurrentTasks()` considers all kinds of resources when calculating the max concurrent tasks
2. Refactor `calculateAvailableSlots()` to make it be able to be used for both `CoarseGrainedSchedulerBackend` and `TaskSchedulerImpl`
### Why are the changes needed?
Currently, `CoarseGrainedSchedulerBackend.maxNumConcurrentTasks()` only considers the CPU for the max concurrent tasks. This can cause the application to hang when a barrier stage requires extra custom resources but the cluster doesn't have enough corresponding resources. Because, without the checking for other custom resources in `maxNumConcurrentTasks`, the barrier stage can be submitted to the `TaskSchedulerImpl`. But the `TaskSchedulerImpl` won't launch tasks for the barrier stage due to the insufficient task slots calculated by `TaskSchedulerImpl.calculateAvailableSlots` (which does check all kinds of resources).
The application hang issue can be reproduced by the added unit test.
### Does this PR introduce _any_ user-facing change?
Yes. In case of a barrier stage requires more custom resources than the cluster has, the application can get hang before this PR but can fail due to insufficient resources at the end after this PR.
### How was this patch tested?
Added a unit test.
Closes#29332 from Ngone51/fix-slots.
Authored-by: yi.wu <yi.wu@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
Exit the executor when it has been asked to decommission and there is nothing left for it to do.
This is a rebase of https://github.com/apache/spark/pull/28817
### Why are the changes needed?
If we want to use decommissioning in Spark's own scale down we should terminate the executor once finished.
Furthermore, in graceful shutdown it makes sense to release resources we no longer need if we've been asked to shutdown by the cluster manager instead of always holding the resources as long as possible.
### Does this PR introduce _any_ user-facing change?
The decommissioned executors will exit and the end of decommissioning. This is sort of a user facing change, however decommissioning hasn't been in any releases yet.
### How was this patch tested?
I changed the unit test to not send the executor exit message and still wait on the executor exited message.
Closes#29211 from holdenk/SPARK-31197-exit-execs-redone.
Authored-by: Holden Karau <hkarau@apple.com>
Signed-off-by: Holden Karau <hkarau@apple.com>
# What changes were proposed in this pull request?
This PR adds a `FileNotFoundException` try catch block while adding a new entry to history server application listing to skip the non-existing path.
### Why are the changes needed?
If there are a large number (>100k) of applications log dir, listing the log dir will take a few seconds. After getting the path list some applications might have finished already, and the filename will change from `foo.inprogress` to `foo`.
It leads to a problem when adding an entry to the listing, querying file status like `fileSizeForLastIndex` will throw out a `FileNotFoundException` exception if the application was finished. And the exception will abort current loop, in a busy cluster, it will make history server couldn't list and load any application log.
```
20/08/03 15:17:23 ERROR FsHistoryProvider: Exception in checking for event log updates
java.io.FileNotFoundException: File does not exist: hdfs://xx/logs/spark/application_11111111111111.lz4.inprogress
at org.apache.hadoop.hdfs.DistributedFileSystem$29.doCall(DistributedFileSystem.java:1527)
at org.apache.hadoop.hdfs.DistributedFileSystem$29.doCall(DistributedFileSystem.java:1520)
at org.apache.hadoop.fs.FileSystemLinkResolver.resolve(FileSystemLinkResolver.java:81)
at org.apache.hadoop.hdfs.DistributedFileSystem.getFileStatus(DistributedFileSystem.java:1520)
at org.apache.spark.deploy.history.SingleFileEventLogFileReader.status$lzycompute(EventLogFileReaders.scala:170)
```
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
1. setup another script keeps changing the filename of applications under history log dir
2. launch the history server
3. check whether the `File does not exist` error log was gone.
Closes#29350 from yanxiaole/SPARK-32529.
Authored-by: Yan Xiaole <xiaole.yan@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This is a follow-up of #29278.
This PR changes the config name to switch allow/disallow `SparkContext` in executors as per the comment https://github.com/apache/spark/pull/29278#pullrequestreview-460256338.
### Why are the changes needed?
The config name `spark.executor.allowSparkContext` is more reasonable.
### Does this PR introduce _any_ user-facing change?
Yes, the config name is changed.
### How was this patch tested?
Updated tests.
Closes#29340 from ueshin/issues/SPARK-32160/change_config_name.
Authored-by: Takuya UESHIN <ueshin@databricks.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
This PR modified the parser code to handle invalid usages of a SET/RESET command.
For example;
```
SET spark.sql.ansi.enabled true
```
The above SQL command does not change the configuration value and it just tries to display the value of the configuration
`spark.sql.ansi.enabled true`. This PR disallows using special characters including spaces in the configuration name and reports a user-friendly error instead. In the error message, it tells users a workaround to use quotes or a string literal if they still needs to specify a configuration with them.
Before this PR:
```
scala> sql("SET spark.sql.ansi.enabled true").show(1, -1)
+---------------------------+-----------+
|key |value |
+---------------------------+-----------+
|spark.sql.ansi.enabled true|<undefined>|
+---------------------------+-----------+
```
After this PR:
```
scala> sql("SET spark.sql.ansi.enabled true")
org.apache.spark.sql.catalyst.parser.ParseException:
Expected format is 'SET', 'SET key', or 'SET key=value'. If you want to include special characters in key, please use quotes, e.g., SET `ke y`=value.(line 1, pos 0)
== SQL ==
SET spark.sql.ansi.enabled true
^^^
```
### Why are the changes needed?
For better user-friendly errors.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Added tests in `SparkSqlParserSuite`.
Closes#29146 from maropu/SPARK-32257.
Lead-authored-by: Takeshi Yamamuro <yamamuro@apache.org>
Co-authored-by: Wenchen Fan <wenchen@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
When https is enabled for Spark UI, an HTTP request will be redirected as an encoded HTTPS URL: https://github.com/apache/spark/pull/10238/files#diff-f79a5ead735b3d0b34b6b94486918e1cR312
When we create the redirect url, we will call getRequestURI and getQueryString. Both two methods may return an encoded string. However, we pass them directly to the following URI constructor
```
URI(String scheme, String authority, String path, String query, String fragment)
```
As this URI constructor assumes both path and query parameters are decoded strings, it will encode them again. This makes the redirect URL encoded twice.
This problem is on stage page with HTTPS enabled. The URL of "/taskTable" contains query parameter `order%5B0%5D%5Bcolumn%5D`. After encoded it becomes `order%255B0%255D%255Bcolumn%255D` and it will be decoded as `order%5B0%5D%5Bcolumn%5D` instead of `order[0][dir]`. When the parameter `order[0][dir]` is missing, there will be an excetpion from:
https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/status/api/v1/StagesResource.scala#L176
and the stage page fail to load.
To fix the problem, we can try decoding the query parameters before encoding it. This is to make sure we encode the URL
### Why are the changes needed?
Fix a UI issue when HTTPS is enabled
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
A new Unit test + manually test on a cluster
Closes#29271 from gengliangwang/urlEncode.
Authored-by: Gengliang Wang <gengliang.wang@databricks.com>
Signed-off-by: Gengliang Wang <gengliang.wang@databricks.com>
### What changes were proposed in this pull request?
Just few log lines fixes which are logging the object name instead of the stage IDs
### Why are the changes needed?
This would make it easier later for debugging.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Just log messages. Existing tests should be enough
Closes#29279 from venkata91/SPARK-31418-follow-up.
Authored-by: Venkata krishnan Sowrirajan <vsowrirajan@linkedin.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
This is a follow-up of #28986.
This PR adds a config to switch allow/disallow to create `SparkContext` in executors.
- `spark.driver.allowSparkContextInExecutors`
### Why are the changes needed?
Some users or libraries actually create `SparkContext` in executors.
We shouldn't break their workloads.
### Does this PR introduce _any_ user-facing change?
Yes, users will be able to create `SparkContext` in executors with the config enabled.
### How was this patch tested?
More tests are added.
Closes#29278 from ueshin/issues/SPARK-32160/add_configs.
Authored-by: Takuya UESHIN <ueshin@databricks.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
This PR removes a test added in SPARK-32175(#29002).
### Why are the changes needed?
That test is flaky. It can be mitigated by increasing the timeout but it would rather be simpler to remove the test.
See also the [discussion](https://github.com/apache/spark/pull/29002#issuecomment-666746857).
### Does this PR introduce _any_ user-facing change?
No.
Closes#29314 from sarutak/remove-flaky-test.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Kousuke Saruta <sarutak@oss.nttdata.com>
### What changes were proposed in this pull request?
This test tries to fix the flakyness of BlockManagerDecommissionIntegrationSuite.
### Description of the problem
Make the block manager decommissioning test be less flaky
An interesting failure happens when migrateDuring = true (and persist or shuffle is true):
- We schedule the job with tasks on executors 0, 1, 2.
- We wait 300 ms and decommission executor 0.
- If the task is not yet done on executor 0, it will now fail because
the block manager won't be able to save the block. This condition is
easy to trigger on a loaded machine where the github checks run.
- The task with retry on a different executor (1 or 2) and its shuffle
blocks will land there.
- No actual block migration happens here because the decommissioned
executor technically failed before it could even produce a block.
To remove the above race, this change replaces the fixed wait for 300 ms to wait for an actual task to succeed. When a task has succeeded, we know its blocks would have been written for sure and thus its executor would certainly be forced to migrate those blocks when it is decommissioned.
The change always decommissions an executor on which a real task finished successfully instead of picking the first executor. Because the system may choose to schedule nothing on the first executor and instead run the two tasks on one executor.
### Why are the changes needed?
I have had bad luck with BlockManagerDecommissionIntegrationSuite and it has failed several times on my PRs. So fixing it.
### Does this PR introduce _any_ user-facing change?
No, unit test only change.
### How was this patch tested?
Github checks. Ran this test 100 times, 10 at a time in parallel in a script.
Closes#29226 from agrawaldevesh/block-manager-decom-flaky.
Authored-by: Devesh Agrawal <devesh.agrawal@gmail.com>
Signed-off-by: Holden Karau <hkarau@apple.com>
### What changes were proposed in this pull request?
This PR reduces the prospect of a job loss during decommissioning. It
fixes two holes in the current decommissioning framework:
- (a) Loss of decommissioned executors is not treated as a job failure:
We know that the decommissioned executor would be dying soon, so its death is
clearly not caused by the application.
- (b) Shuffle files on the decommissioned host are cleared when the
first fetch failure is detected from a decommissioned host: This is a
bit tricky in terms of when to clear the shuffle state ? Ideally you
want to clear it the millisecond before the shuffle service on the node
dies (or the executor dies when there is no external shuffle service) --
too soon and it could lead to some wastage and too late would lead to
fetch failures.
The approach here is to do this clearing when the very first fetch
failure is observed on the decommissioned block manager, without waiting for
other blocks to also signal a failure.
### Why are the changes needed?
Without them decommissioning a lot of executors at a time leads to job failures.
### Code overview
The task scheduler tracks the executors that were decommissioned along with their
`ExecutorDecommissionInfo`. This information is used by: (a) For handling a `ExecutorProcessLost` error, or (b) by the `DAGScheduler` when handling a fetch failure.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Added a new unit test `DecommissionWorkerSuite` to test the new behavior by exercising the Master-Worker decommissioning. I chose to add a new test since the setup logic was quite different from the existing `WorkerDecommissionSuite`. I am open to changing the name of the newly added test suite :-)
### Questions for reviewers
- Should I add a feature flag to guard these two behaviors ? They seem safe to me that they should only get triggered by decommissioning, but you never know :-).
Closes#29014 from agrawaldevesh/decom_harden.
Authored-by: Devesh Agrawal <devesh.agrawal@gmail.com>
Signed-off-by: Holden Karau <hkarau@apple.com>
### What changes were proposed in this pull request?
So far, we fixed many stuffs in `core` module. This PR fixes the remaining UT failures in Scala 2.13.
- `OneApplicationResource.environmentInfo` will return a deterministic result for `sparkProperties`, `hadoopProperties`, `systemProperties`, and `classpathEntries`.
- `SubmitRestProtocolSuite` has Scala 2.13 answer in addition to the existing Scala 2.12 answer, and uses the expected answer based on the Scala runtime version.
### Why are the changes needed?
To support Scala 2.13.
### Does this PR introduce _any_ user-facing change?
Yes, `environmentInfo` is changed, but this fixes the indeterministic behavior.
### How was this patch tested?
- Scala 2.12: Pass the Jenkins or GitHub Action
- Scala 2.13: Do the following.
```
$ dev/change-scala-version.sh 2.13
$ build/mvn test -pl core --am -Pscala-2.13
```
**BEFORE**
```
Tests: succeeded 2612, failed 3, canceled 1, ignored 8, pending 0
*** 3 TESTS FAILED ***
```
**AFTER**
```
Tests: succeeded 2615, failed 0, canceled 1, ignored 8, pending 0
All tests passed.
```
Closes#29298 from dongjoon-hyun/SPARK-32489.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR aims to remove `java.ws.rs.NotFoundException` from two problematic `import` statements. All the other use cases are correct.
### Why are the changes needed?
In `StagesResource` and `OneApplicationResource`, there exist two `NotFoundException`s.
- javax.ws.rs.NotFoundException
- org.apache.spark.status.api.v1.NotFoundException
To use `org.apache.spark.status.api.v1.NotFoundException` correctly, we should not import `java.ws.rs.NotFoundException`. This causes UT failures in Scala 2.13 environment.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
- Scala 2.12: Pass the GitHub Action or Jenkins.
- Scala 2.13: Do the following manually.
```
$ dev/change-scala-version.sh 2.13
$ build/mvn test -pl core --am -Pscala-2.13 -Dtest=none -DwildcardSuites=org.apache.spark.deploy.history.HistoryServerSuite
```
**BEFORE**
```
*** 4 TESTS FAILED ***
```
**AFTER**
```
*** 1 TEST FAILED ***
```
Closes#29293 from dongjoon-hyun/SPARK-32487.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR aims to make `ResourceAllocator.availableAddrs` deterministic.
### Why are the changes needed?
Currently, this function returns indeterministically due to the underlying `HashMap`. So, the test case itself is creating a list `[0, 1, 2]` initially, but ends up with comparing `[2, 1, 0]`.
Not only this happens in the 3.0.0, but also this causes UT failures on Scala 2.13 environment.
### Does this PR introduce _any_ user-facing change?
Yes, but this fixes the in-deterministic behavior.
### How was this patch tested?
- Scala 2.12: This should pass the UT with the modified test case.
- Scala 2.13: This can be tested like the following (at least `JsonProtocolSuite`)
```
$ dev/change-scala-version.sh 2.13
$ build/mvn test -pl core --am -Pscala-2.13 -Dtest=none -DwildcardSuites=org.apache.spark.deploy.JsonProtocolSuite
```
**BEFORE**
```
*** 2 TESTS FAILED ***
```
**AFTER**
```
All tests passed.
```
Closes#29281 from dongjoon-hyun/SPARK-32476.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR aims to make `JsonProtocol.accumulablesToJson` deterministic.
### Why are the changes needed?
Currently, `JsonProtocol.accumulablesToJson` is indeterministic. So, `JsonProtocolSuite` itself is also using mixed test cases in terms of `"Accumulables": [ ... ]`.
Not only this is indeterministic, but also this causes a UT failure in `JsonProtocolSuite` in Scala 2.13.
### Does this PR introduce _any_ user-facing change?
Yes. However, this is a fix on indeterministic behavior.
### How was this patch tested?
- Scala 2.12: Pass the GitHub Action or Jenkins.
- Scala 2.13: Do the following.
```
$ dev/change-scala-version.sh 2.13
$ build/mvn test -pl core --am -Pscala-2.13 -Dtest=none -DwildcardSuites=org.apache.spark.util.JsonProtocolSuite
```
**BEFORE**
```
*** 1 TEST FAILED ***
```
**AFTER**
```
All tests passed.
```
Closes#29282 from dongjoon-hyun/SPARK-32477.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR changes the order between initialization for ExecutorPlugin and starting heartbeat thread in Executor.
### Why are the changes needed?
In the current master, heartbeat thread in a executor starts after plugin initialization so if the initialization takes long time, heartbeat is not sent to driver and the executor will be removed from cluster.
### Does this PR introduce _any_ user-facing change?
Yes. Plugins for executors will be allowed to take long time for initialization.
### How was this patch tested?
New testcase.
Closes#29002 from sarutak/fix-heartbeat-issue.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
`spark.kryo.registrator` in 3.0 has a regression problem. From [SPARK-12080](https://issues.apache.org/jira/browse/SPARK-12080), it supports multiple user registrators by
```scala
private val userRegistrators = conf.get("spark.kryo.registrator", "")
.split(',').map(_.trim)
.filter(!_.isEmpty)
```
But it donsn't work in 3.0. Fix it by `toSequence` in `Kryo.scala`
### Why are the changes needed?
In previous Spark version (2.x), it supported multiple user registrators by
```scala
private val userRegistrators = conf.get("spark.kryo.registrator", "")
.split(',').map(_.trim)
.filter(!_.isEmpty)
```
But it doesn't work in 3.0. It's should be a regression.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Existed unit tests.
Closes#29123 from LantaoJin/SPARK-32283.
Authored-by: LantaoJin <jinlantao@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
This PR aims to use `command -v` in non-Window operating systems instead of executing the given command.
### Why are the changes needed?
1. `command` is POSIX-compatible
- **POSIX.1-2017**: https://pubs.opengroup.org/onlinepubs/9699919799/utilities/command.html
2. `command` is faster and safer than the direct execution
- `command` doesn't invoke another process.
```scala
scala> sys.process.Process("ls").run().exitValue()
LICENSE
NOTICE
bin
doc
lib
man
res1: Int = 0
```
3. The existing way behaves inconsistently.
- `rm` cannot be checked.
**AS-IS**
```scala
scala> sys.process.Process("rm").run().exitValue()
usage: rm [-f | -i] [-dPRrvW] file ...
unlink file
res0: Int = 64
```
**TO-BE**
```
Welcome to Scala 2.13.3 (OpenJDK 64-Bit Server VM, Java 1.8.0_262).
Type in expressions for evaluation. Or try :help.
scala> sys.process.Process(Seq("sh", "-c", s"command -v ls")).run().exitValue()
/bin/ls
val res1: Int = 0
```
4. The existing logic is already broken in Scala 2.13 environment because it hangs like the following.
```scala
$ bin/scala
Welcome to Scala 2.13.3 (OpenJDK 64-Bit Server VM, Java 1.8.0_262).
Type in expressions for evaluation. Or try :help.
scala> sys.process.Process("cat").run().exitValue() // hang here.
```
### Does this PR introduce _any_ user-facing change?
No. Although this is inside `main` source directory, this is used for testing purpose.
```
$ git grep testCommandAvailable | grep -v 'def testCommandAvailable'
core/src/test/scala/org/apache/spark/rdd/PipedRDDSuite.scala: assume(TestUtils.testCommandAvailable("cat"))
core/src/test/scala/org/apache/spark/rdd/PipedRDDSuite.scala: assume(TestUtils.testCommandAvailable("wc"))
core/src/test/scala/org/apache/spark/rdd/PipedRDDSuite.scala: assume(TestUtils.testCommandAvailable("cat"))
core/src/test/scala/org/apache/spark/rdd/PipedRDDSuite.scala: assume(TestUtils.testCommandAvailable("cat"))
core/src/test/scala/org/apache/spark/rdd/PipedRDDSuite.scala: assume(TestUtils.testCommandAvailable("cat"))
core/src/test/scala/org/apache/spark/rdd/PipedRDDSuite.scala: assume(TestUtils.testCommandAvailable(envCommand))
core/src/test/scala/org/apache/spark/rdd/PipedRDDSuite.scala: assume(!TestUtils.testCommandAvailable("some_nonexistent_command"))
core/src/test/scala/org/apache/spark/rdd/PipedRDDSuite.scala: assume(TestUtils.testCommandAvailable("cat"))
core/src/test/scala/org/apache/spark/rdd/PipedRDDSuite.scala: assume(TestUtils.testCommandAvailable("cat"))
core/src/test/scala/org/apache/spark/rdd/PipedRDDSuite.scala: assume(TestUtils.testCommandAvailable(envCommand))
sql/core/src/test/scala/org/apache/spark/sql/IntegratedUDFTestUtils.scala: private lazy val isPythonAvailable: Boolean = TestUtils.testCommandAvailable(pythonExec)
sql/core/src/test/scala/org/apache/spark/sql/IntegratedUDFTestUtils.scala: if (TestUtils.testCommandAvailable(pythonExec)) {
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveQuerySuite.scala: skip = !TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveQuerySuite.scala: skip = !TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveQuerySuite.scala: skip = !TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveQuerySuite.scala: skip = !TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveQuerySuite.scala: skip = !TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveQuerySuite.scala: skip = !TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveQuerySuite.scala: assume(TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveQuerySuite.scala: skip = !TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveQuerySuite.scala: skip = !TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveScriptTransformationSuite.scala: assume(TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveScriptTransformationSuite.scala: assume(TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveScriptTransformationSuite.scala: assume(TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveScriptTransformationSuite.scala: assume(TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveScriptTransformationSuite.scala: assume(TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveScriptTransformationSuite.scala: assume(TestUtils.testCommandAvailable("python"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveScriptTransformationSuite.scala: assume(TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/HiveScriptTransformationSuite.scala: assume(TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/SQLQuerySuite.scala: assume(TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/SQLQuerySuite.scala: assume(TestUtils.testCommandAvailable("echo | sed"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/SQLQuerySuite.scala: assume(TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/SQLQuerySuite.scala: assume(TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/SQLQuerySuite.scala: assume(TestUtils.testCommandAvailable("/bin/bash"))
sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/SQLQuerySuite.scala: assume(TestUtils.testCommandAvailable("/bin/bash"))
```
### How was this patch tested?
- **Scala 2.12**: Pass the Jenkins with the existing tests and one modified test.
- **Scala 2.13**: Do the following manually. It should pass instead of `hang`.
```
$ dev/change-scala-version.sh 2.13
$ build/mvn test -pl core --am -Pscala-2.13 -Dtest=none -DwildcardSuites=org.apache.spark.rdd.PipedRDDSuite
...
Tests: succeeded 12, failed 0, canceled 0, ignored 0, pending 0
All tests passed.
```
Closes#29241 from dongjoon-hyun/SPARK-32443.
Lead-authored-by: HyukjinKwon <gurwls223@apache.org>
Co-authored-by: Dongjoon Hyun <dongjoon@apache.org>
Co-authored-by: Hyukjin Kwon <gurwls223@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
Support set off heap memory in `ExecutorResourceRequests`
### Why are the changes needed?
Support stage level scheduling
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Added UT in `ResourceProfileSuite` and `DAGSchedulerSuite`
Closes#28972 from warrenzhu25/30794.
Authored-by: Warren Zhu <zhonzh@microsoft.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
This PR aims to increase the memory parameter in `BlockManagerSuite`'s worker decommission test cases.
### Why are the changes needed?
Scala 2.13 generates different Java objects and this affects Spark's `SizeEstimator/SizeTracker/SizeTrackingVector`. This causes UT failures like the following. If we decrease the values, those test cases fails in Scala 2.12, too.
```
$ dev/change-scala-version.sh 2.13
$ build/mvn test -pl core --am -Pscala-2.13 -Dtest=none -DwildcardSuites=org.apache.spark.storage.BlockManagerSuite
...
- test decommission block manager should not be part of peers *** FAILED ***
0 did not equal 2 (BlockManagerSuite.scala:1869)
- test decommissionRddCacheBlocks should offload all cached blocks *** FAILED ***
0 did not equal 2 (BlockManagerSuite.scala:1884)
...
Tests: succeeded 81, failed 2, canceled 0, ignored 0, pending 0
*** 2 TESTS FAILED ***
```
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
```
$ dev/change-scala-version.sh 2.13
$ build/mvn test -pl core --am -Pscala-2.13 -Dtest=none -DwildcardSuites=org.apache.spark.storage.BlockManagerSuite
...
Tests: succeeded 83, failed 0, canceled 0, ignored 0, pending 0
All tests passed.
```
Closes#29238 from dongjoon-hyun/SPARK-32440.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
Since Scala 2.13, `HashMap` is changed to become a final in the future and `.withDefault` is recommended. This PR aims to use `HashMap.withDefaultValue` instead of overriding manually in the test case.
- https://www.scala-lang.org/api/current/scala/collection/mutable/HashMap.html
```scala
deprecatedInheritance(message =
"HashMap wil be made final; use .withDefault for the common use case of computing a default value",
since = "2.13.0")
```
### Why are the changes needed?
In Scala 2.13, the existing code causes a failure because the default value function doesn't work correctly.
```
$ dev/change-scala-version.sh 2.13
$ build/mvn test -pl core --am -Pscala-2.13 -Dtest=none -DwildcardSuites=org.apache.spark.rdd.RDDSuite
- aggregate *** FAILED ***
org.apache.spark.SparkException: Job aborted due to stage failure:
Task 0 in stage 61.0 failed 1 times, most recent failure: Lost task 0.0 in stage 61.0 (TID 198, localhost, executor driver):
java.util.NoSuchElementException: key not found: a
```
### Does this PR introduce _any_ user-facing change?
No. This is a test case change.
### How was this patch tested?
1. **Scala 2.12:** Pass the Jenkins or GitHub with the existing tests.
2. **Scala 2.13**: Manually do the following.
```
$ dev/change-scala-version.sh 2.13
$ build/mvn test -pl core --am -Pscala-2.13 -Dtest=none -DwildcardSuites=org.apache.spark.rdd.RDDSuite
...
Tests: succeeded 72, failed 0, canceled 0, ignored 0, pending 0
All tests passed.
```
Closes#29235 from dongjoon-hyun/SPARK-32438.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR aims to initialize `numNonEmptyBlocks` in `HighlyCompressedMapStatus.readExternal`.
In Scala 2.12, this is initialized to `-1` via the following.
```scala
protected def this() = this(null, -1, null, -1, null, -1) // For deserialization only
```
### Why are the changes needed?
In Scala 2.13, this causes several UT failures because `HighlyCompressedMapStatus.readExternal` doesn't initialize this field. The following is one example.
- org.apache.spark.scheduler.MapStatusSuite
```
MapStatusSuite:
- compressSize
- decompressSize
*** RUN ABORTED ***
java.lang.NoSuchFieldError: numNonEmptyBlocks
at org.apache.spark.scheduler.HighlyCompressedMapStatus.<init>(MapStatus.scala:181)
at org.apache.spark.scheduler.HighlyCompressedMapStatus$.apply(MapStatus.scala:281)
at org.apache.spark.scheduler.MapStatus$.apply(MapStatus.scala:73)
at org.apache.spark.scheduler.MapStatusSuite.$anonfun$new$8(MapStatusSuite.scala:64)
at scala.runtime.java8.JFunction1$mcVD$sp.apply(JFunction1$mcVD$sp.scala:18)
at scala.collection.immutable.List.foreach(List.scala:333)
at org.apache.spark.scheduler.MapStatusSuite.$anonfun$new$7(MapStatusSuite.scala:61)
at scala.runtime.java8.JFunction1$mcVJ$sp.apply(JFunction1$mcVJ$sp.scala:18)
at scala.collection.immutable.List.foreach(List.scala:333)
at org.apache.spark.scheduler.MapStatusSuite.$anonfun$new$6(MapStatusSuite.scala:60)
...
```
### Does this PR introduce _any_ user-facing change?
No. This is a private class.
### How was this patch tested?
1. Pass the GitHub Action or Jenkins with the existing tests.
2. Test with Scala-2.13 with `MapStatusSuite`.
```
$ dev/change-scala-version.sh 2.13
$ build/mvn test -pl core --am -Pscala-2.13 -Dtest=none -DwildcardSuites=org.apache.spark.scheduler.MapStatusSuite
...
MapStatusSuite:
- compressSize
- decompressSize
- MapStatus should never report non-empty blocks' sizes as 0
- large tasks should use org.apache.spark.scheduler.HighlyCompressedMapStatus
- HighlyCompressedMapStatus: estimated size should be the average non-empty block size
- SPARK-22540: ensure HighlyCompressedMapStatus calculates correct avgSize
- RoaringBitmap: runOptimize succeeded
- RoaringBitmap: runOptimize failed
- Blocks which are bigger than SHUFFLE_ACCURATE_BLOCK_THRESHOLD should not be underestimated.
- SPARK-21133 HighlyCompressedMapStatus#writeExternal throws NPE
Run completed in 7 seconds, 971 milliseconds.
Total number of tests run: 10
Suites: completed 2, aborted 0
Tests: succeeded 10, failed 0, canceled 0, ignored 0, pending 0
All tests passed.
```
Closes#29231 from dongjoon-hyun/SPARK-32436.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR aims to speed up `MapStatus` deserialization by 5~18% with the latest RoaringBitmap `0.9.0` and new APIs. Note that we focus on `deserialization` time because `serialization` occurs once while `deserialization` occurs many times.
### Why are the changes needed?
The current version is too old. We had better upgrade it to get the performance improvement and bug fixes.
Although `MapStatusesSerDeserBenchmark` is synthetic, the benchmark result is updated with this patch.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Pass the Jenkins or GitHub Action.
Closes#29233 from dongjoon-hyun/SPARK-ROAR.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
`UninterruptibleThread` running functionality is baked into `KafkaOffsetReader` which can be extracted into a class. The main intention is to simplify `KafkaOffsetReader` in order to make easier to solve SPARK-32032. In this PR I've made this extraction without functionality change.
### Why are the changes needed?
`UninterruptibleThread` running functionality is baked into `KafkaOffsetReader`.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Existing + additional unit tests.
Closes#29187 from gaborgsomogyi/SPARK-32387.
Authored-by: Gabor Somogyi <gabor.g.somogyi@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
I wasn't able to reproduce the failure but the best I can tell is that the allocation manager timer triggers and call doRequest. The timeout is 10s so try to increase that to 30seconds.
### Why are the changes needed?
test failure
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
unit test
Closes#29225 from tgravescs/SPARK-32287.
Authored-by: Thomas Graves <tgraves@nvidia.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
Updates to scalatest 3.2.0. Though it looks large, it is 99% changes to the new location of scalatest classes.
### Why are the changes needed?
3.2.0+ has a fix that is required for Scala 2.13.3+ compatibility.
### Does this PR introduce _any_ user-facing change?
No, only affects tests.
### How was this patch tested?
Existing tests.
Closes#29196 from srowen/SPARK-32398.
Authored-by: Sean Owen <srowen@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
In this change, when dynamic allocation is enabled instead of aborting immediately when there is an unschedulable taskset due to blacklisting, pass an event saying `SparkListenerUnschedulableTaskSetAdded` which will be handled by `ExecutorAllocationManager` and request more executors needed to schedule the unschedulable blacklisted tasks. Once the event is sent, we start the abortTimer similar to [SPARK-22148][SPARK-15815] to abort in the case when no new executors launched either due to max executors reached or cluster manager is out of capacity.
### Why are the changes needed?
This is an improvement. In the case when dynamic allocation is enabled, this would request more executors to schedule the unschedulable tasks instead of aborting the stage without even retrying upto spark.task.maxFailures times (in some cases not retrying at all). This is a potential issue with respect to Spark's Fault tolerance.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Added unit tests both in ExecutorAllocationManagerSuite and TaskSchedulerImplSuite
Closes#28287 from venkata91/SPARK-31418.
Authored-by: Venkata krishnan Sowrirajan <vsowrirajan@linkedin.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
This PR is a giant plumbing PR that plumbs an `ExecutorDecommissionInfo` along
with the DecommissionExecutor message.
### Why are the changes needed?
The primary motivation is to know whether a decommissioned executor
would also be loosing shuffle files -- and thus it is important to know
whether the host would also be decommissioned.
In the absence of this PR, the existing code assumes that decommissioning an executor does not loose the whole host with it, and thus does not clear the shuffle state if external shuffle service is enabled. While this may hold in some cases (like K8s decommissioning an executor pod, or YARN container preemption), it does not hold in others like when the cluster is managed by a Standalone Scheduler (Master). This is similar to the existing `workerLost` field in the `ExecutorProcessLost` message.
In the future, this `ExecutorDecommissionInfo` can be embellished for
knowing how long the executor has to live for scenarios like Cloud spot
kills (or Yarn preemption) and the like.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Tweaked an existing unit test in `AppClientSuite`
Closes#29032 from agrawaldevesh/plumb_decom_info.
Authored-by: Devesh Agrawal <devesh.agrawal@gmail.com>
Signed-off-by: Holden Karau <hkarau@apple.com>
### What changes were proposed in this pull request?
If an executor is lost, the `DAGScheduler` handles the executor loss by removing the executor but does not unregister its outputs if the external shuffle service is used. However, if the node on which the executor runs is lost, the shuffle service may not be able to serve the shuffle files.
In such a case, when fetches from the executor's outputs fail in the same stage, the `DAGScheduler` again removes the executor and by right, should unregister its outputs. It doesn't because the epoch used to track the executor failure has not increased.
We track the epoch for failed executors that result in lost file output separately, so we can unregister the outputs in this scenario. The idea to track a second epoch is due to Attila Zsolt Piros.
### Why are the changes needed?
Without the changes, the loss of a node could require two stage attempts to recover instead of one.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
New unit test. This test fails without the change and passes with it.
Closes#28848 from wypoon/SPARK-32003.
Authored-by: Wing Yew Poon <wypoon@cloudera.com>
Signed-off-by: Imran Rashid <irashid@cloudera.com>
### What changes were proposed in this pull request?
In PR, I propose to create input tables once before executing tests in `JDBCSuite` and `JdbcRDDSuite`. Currently, the table are created before every test in the test suites.
### Why are the changes needed?
This speed up the test suites up 30-40%.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Run the modified test suites
Closes#29176 from MaxGekk/jdbc-suite-before-all.
Authored-by: Max Gekk <max.gekk@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
Catch the `RpcEnvStoppedException` and log debug it when stop is called for a `LocalSparkCluster`.
This PR also contains two small changes to fix the potential issues.
### Why are the changes needed?
Currently, there's always "RpcEnv already stopped" error if we exit spark-shell with local-cluster mode:
```
20/06/07 14:54:18 ERROR TransportRequestHandler: Error while invoking RpcHandler#receive() for one-way message.
org.apache.spark.rpc.RpcEnvStoppedException: RpcEnv already stopped.
at org.apache.spark.rpc.netty.Dispatcher.postMessage(Dispatcher.scala:167)
at org.apache.spark.rpc.netty.Dispatcher.postOneWayMessage(Dispatcher.scala:150)
at org.apache.spark.rpc.netty.NettyRpcHandler.receive(NettyRpcEnv.scala:691)
at org.apache.spark.network.server.TransportRequestHandler.processOneWayMessage(TransportRequestHandler.java:253)
at org.apache.spark.network.server.TransportRequestHandler.handle(TransportRequestHandler.java:111)
at org.apache.spark.network.server.TransportChannelHandler.channelRead0(TransportChannelHandler.java:140)
at org.apache.spark.network.server.TransportChannelHandler.channelRead0(TransportChannelHandler.java:53)
at io.netty.channel.SimpleChannelInboundHandler.channelRead(SimpleChannelInboundHandler.java:99)
at io.netty.channel.AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext.java:379)
at io.netty.channel.AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext.java:365)
at io.netty.channel.AbstractChannelHandlerContext.fireChannelRead(AbstractChannelHandlerContext.java:357)
at io.netty.handler.timeout.IdleStateHandler.channelRead(IdleStateHandler.java:286)
at io.netty.channel.AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext.java:379)
at io.netty.channel.AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext.java:365)
at io.netty.channel.AbstractChannelHandlerContext.fireChannelRead(AbstractChannelHandlerContext.java:357)
at io.netty.handler.codec.MessageToMessageDecoder.channelRead(MessageToMessageDecoder.java:102)
at io.netty.channel.AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext.java:379)
at io.netty.channel.AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext.java:365)
at io.netty.channel.AbstractChannelHandlerContext.fireChannelRead(AbstractChannelHandlerContext.java:357)
at org.apache.spark.network.util.TransportFrameDecoder.channelRead(TransportFrameDecoder.java:102)
at io.netty.channel.AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext.java:379)
at io.netty.channel.AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext.java:365)
at io.netty.channel.AbstractChannelHandlerContext.fireChannelRead(AbstractChannelHandlerContext.java:357)
at io.netty.channel.DefaultChannelPipeline$HeadContext.channelRead(DefaultChannelPipeline.java:1410)
at io.netty.channel.AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext.java:379)
at io.netty.channel.AbstractChannelHandlerContext.invokeChannelRead(AbstractChannelHandlerContext.java:365)
at io.netty.channel.DefaultChannelPipeline.fireChannelRead(DefaultChannelPipeline.java:919)
at io.netty.channel.nio.AbstractNioByteChannel$NioByteUnsafe.read(AbstractNioByteChannel.java:163)
at io.netty.channel.nio.NioEventLoop.processSelectedKey(NioEventLoop.java:714)
at io.netty.channel.nio.NioEventLoop.processSelectedKeysOptimized(NioEventLoop.java:650)
at io.netty.channel.nio.NioEventLoop.processSelectedKeys(NioEventLoop.java:576)
at io.netty.channel.nio.NioEventLoop.run(NioEventLoop.java:493)
at io.netty.util.concurrent.SingleThreadEventExecutor$4.run(SingleThreadEventExecutor.java:989)
at io.netty.util.internal.ThreadExecutorMap$2.run(ThreadExecutorMap.java:74)
at io.netty.util.concurrent.FastThreadLocalRunnable.run(FastThreadLocalRunnable.java:30)
at java.lang.Thread.run(Thread.java:748)
```
When we call stop on `StandaloneSchedulerBackend`, the backend will firstly send `UnregisterApplication` to `Master` and then call stop on `LocalSparkCluster` immediately. On the other side, `Master` will send messages to `Worker` when it receives `UnregisterApplication`. However, the rpcEnv of the `Worker` has been already stoped by the backend. Therefore, the error message shows when the `Worker` tries to handle the messages.
It's only an error on shutdown, users would not like to care about it. So we could hide it in debug log and this is also what we've done previously in #18547.
### Does this PR introduce _any_ user-facing change?
Yes, users will not see the error message after this PR.
### How was this patch tested?
Tested manually.
Closes#28746 from Ngone51/fix-spark-31922.
Authored-by: yi.wu <yi.wu@databricks.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
The idea is to improve the performance of HybridStore by adding batch write support to LevelDB. #28412 introduces HybridStore. HybridStore will write data to InMemoryStore at first and use a background thread to dump data to LevelDB once the writing to InMemoryStore is completed. In the comments section of #28412 , mridulm mentioned using batch writing can improve the performance of this dumping process and he wrote the code of writeAll().
### Why are the changes needed?
I did the comparison of the HybridStore switching time between one-by-one write and batch write on an HDD disk. When the disk is free, the batch-write has around 25% improvement, and when the disk is 100% busy, the batch-write has 7x - 10x improvement.
when the disk is at 0% utilization:
| log size, jobs and tasks per job | original switching time, with write() | switching time with writeAll() |
| ---------------------------------- | ------------------------------------- | ------------------------------ |
| 133m, 400 jobs, 100 tasks per job | 16s | 13s |
| 265m, 400 jobs, 200 tasks per job | 30s | 23s |
| 1.3g, 1000 jobs, 400 tasks per job | 136s | 108s |
when the disk is at 100% utilization:
| log size, jobs and tasks per job | original switching time, with write() | switching time with writeAll() |
| --------------------------------- | ------------------------------------- | ------------------------------ |
| 133m, 400 jobs, 100 tasks per job | 116s | 17s |
| 265m, 400 jobs, 200 tasks per job | 251s | 26s |
I also ran some write related benchmarking tests on LevelDBBenchmark.java and measured the total time of writing 1024 objects. The tests were conducted when the disk is at 0% utilization.
| Benchmark test | with write(), ms | with writeAll(), ms |
| ------------------------ | ---------------- | ------------------- |
| randomUpdatesIndexed | 213.06 | 157.356 |
| randomUpdatesNoIndex | 57.869 | 35.439 |
| randomWritesIndexed | 298.854 | 229.274 |
| randomWritesNoIndex | 66.764 | 38.361 |
| sequentialUpdatesIndexed | 87.019 | 56.219 |
| sequentialUpdatesNoIndex | 61.851 | 41.942 |
| sequentialWritesIndexed | 94.044 | 56.534 |
| sequentialWritesNoIndex | 118.345 | 66.483 |
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Manually tested.
Closes#29149 from baohe-zhang/SPARK-32350.
Authored-by: Baohe Zhang <baohe.zhang@verizonmedia.com>
Signed-off-by: Jungtaek Lim (HeartSaVioR) <kabhwan.opensource@gmail.com>
### What is changed?
This pull request adds the ability to migrate shuffle files during Spark's decommissioning. The design document associated with this change is at https://docs.google.com/document/d/1xVO1b6KAwdUhjEJBolVPl9C6sLj7oOveErwDSYdT-pE .
To allow this change the `MapOutputTracker` has been extended to allow the location of shuffle files to be updated with `updateMapOutput`. When a shuffle block is put, a block update message will be sent which triggers the `updateMapOutput`.
Instead of rejecting remote puts of shuffle blocks `BlockManager` delegates the storage of shuffle blocks to it's shufflemanager's resolver (if supported). A new, experimental, trait is added for shuffle resolvers to indicate they handle remote putting of blocks.
The existing block migration code is moved out into a separate file, and a producer/consumer model is introduced for migrating shuffle files from the host as quickly as possible while not overwhelming other executors.
### Why are the changes needed?
Recomputting shuffle blocks can be expensive, we should take advantage of our decommissioning time to migrate these blocks.
### Does this PR introduce any user-facing change?
This PR introduces two new configs parameters, `spark.storage.decommission.shuffleBlocks.enabled` & `spark.storage.decommission.rddBlocks.enabled` that control which blocks should be migrated during storage decommissioning.
### How was this patch tested?
New unit test & expansion of the Spark on K8s decom test to assert that decommisioning with shuffle block migration means that the results are not recomputed even when the original executor is terminated.
This PR is a cleaned-up version of the previous WIP PR I made https://github.com/apache/spark/pull/28331 (thanks to attilapiros for his very helpful reviewing on it :)).
Closes#28708 from holdenk/SPARK-20629-copy-shuffle-data-when-nodes-are-being-shutdown-cleaned-up.
Lead-authored-by: Holden Karau <hkarau@apple.com>
Co-authored-by: Holden Karau <holden@pigscanfly.ca>
Co-authored-by: “attilapiros” <piros.attila.zsolt@gmail.com>
Co-authored-by: Attila Zsolt Piros <attilazsoltpiros@apiros-mbp16.lan>
Signed-off-by: Holden Karau <hkarau@apple.com>
### What changes were proposed in this pull request?
This PR adds functionality to consider the running tasks on decommission executors based on some config.
In spark-on-cloud , we sometimes already know that an executor won't be alive for more than fix amount of time. Ex- In AWS Spot nodes, once we get the notification, we know that a node will be gone in 120 seconds.
So if the running tasks on the decommissioning executors may run beyond currentTime+120 seconds, then they are candidate for speculation.
### Why are the changes needed?
Currently when an executor is decommission, we stop scheduling new tasks on those executors but the already running tasks keeps on running on them. Based on the cloud, we might know beforehand that an executor won't be alive for more than a preconfigured time. Different cloud providers gives different timeouts before they take away the nodes. For Ex- In case of AWS spot nodes, an executor won't be alive for more than 120 seconds. We can utilize this information in cloud environments and take better decisions about speculating the already running tasks on decommission executors.
### Does this PR introduce _any_ user-facing change?
Yes. This PR adds a new config "spark.executor.decommission.killInterval" which they can explicitly set based on the cloud environment where they are running.
### How was this patch tested?
Added UT.
Closes#28619 from prakharjain09/SPARK-21040-speculate-decommission-exec-tasks.
Authored-by: Prakhar Jain <prakharjain09@gmail.com>
Signed-off-by: Holden Karau <hkarau@apple.com>
### What changes were proposed in this pull request?
This PR allows an external agent to inform the Master that certain hosts
are being decommissioned.
### Why are the changes needed?
The current decommissioning is triggered by the Worker getting getting a SIGPWR
(out of band possibly by some cleanup hook), which then informs the Master
about it. This approach may not be feasible in some environments that cannot
trigger a clean up hook on the Worker. In addition, when a large number of
worker nodes are being decommissioned then the master will get a flood of
messages.
So we add a new post endpoint `/workers/kill` on the MasterWebUI that allows an
external agent to inform the master about all the nodes being decommissioned in
bulk. The list of nodes is specified by providing a list of hostnames. All workers on those
hosts will be decommissioned.
This API is merely a new entry point into the existing decommissioning
logic. It does not change how the decommissioning request is handled in
its core.
### Does this PR introduce _any_ user-facing change?
Yes, a new endpoint `/workers/kill` is added to the MasterWebUI. By default only
requests originating from an IP address local to the MasterWebUI are allowed.
### How was this patch tested?
Added unit tests
Closes#29015 from agrawaldevesh/master_decom_endpoint.
Authored-by: Devesh Agrawal <devesh.agrawal@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
Support fetching taskList by status as below:
```
/applications/[app-id]/stages/[stage-id]/[stage-attempt-id]/taskList?status=failed
```
### Why are the changes needed?
When there're large number of tasks in one stage, current api is hard to get taskList by status
### Does this PR introduce _any_ user-facing change?
Yes. Updated monitoring doc.
### How was this patch tested?
Added tests in `HistoryServerSuite`
Closes#28942 from warrenzhu25/SPARK-32125.
Authored-by: Warren Zhu <zhonzh@microsoft.com>
Signed-off-by: Gengliang Wang <gengliang.wang@databricks.com>
### What changes were proposed in this pull request?
This PR will remove references to these "blacklist" and "whitelist" terms besides the blacklisting feature as a whole, which can be handled in a separate JIRA/PR.
This touches quite a few files, but the changes are straightforward (variable/method/etc. name changes) and most quite self-contained.
### Why are the changes needed?
As per discussion on the Spark dev list, it will be beneficial to remove references to problematic language that can alienate potential community members. One such reference is "blacklist" and "whitelist". While it seems to me that there is some valid debate as to whether these terms have racist origins, the cultural connotations are inescapable in today's world.
### Does this PR introduce _any_ user-facing change?
In the test file `HiveQueryFileTest`, a developer has the ability to specify the system property `spark.hive.whitelist` to specify a list of Hive query files that should be tested. This system property has been renamed to `spark.hive.includelist`. The old property has been kept for compatibility, but will log a warning if used. I am open to feedback from others on whether keeping a deprecated property here is unnecessary given that this is just for developers running tests.
### How was this patch tested?
Existing tests should be suitable since no behavior changes are expected as a result of this PR.
Closes#28874 from xkrogen/xkrogen-SPARK-32036-rename-blacklists.
Authored-by: Erik Krogen <ekrogen@linkedin.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
Add a new class HybridStore to make the history server faster when loading event files. When rebuilding the application state from event logs, HybridStore will write data to InMemoryStore at first and use a background thread to dump data to LevelDB once the writing to InMemoryStore is completed. HybridStore is to make content serving faster by using more memory. It's only safe to enable it when the cluster is not having a heavy load.
### Why are the changes needed?
HybridStore can greatly reduce the event logs loading time, especially for large log files. In general, it has 4x - 6x UI loading speed improvement for large log files. The detailed result is shown in comments.
### Does this PR introduce any user-facing change?
This PR adds new configs `spark.history.store.hybridStore.enabled` and `spark.history.store.hybridStore.maxMemoryUsage`.
### How was this patch tested?
A test suite for HybridStore is added. I also manually tested it on 3.1.0 on mac os.
This is a follow-up for the work done by Hieu Huynh in 2019.
Closes#28412 from baohe-zhang/SPARK-31608.
Authored-by: Baohe Zhang <baohe.zhang@verizonmedia.com>
Signed-off-by: Jungtaek Lim (HeartSaVioR) <kabhwan.opensource@gmail.com>
### What changes were proposed in this pull request?
This PR aims to drop Python 2.7, 3.4 and 3.5.
Roughly speaking, it removes all the widely known Python 2 compatibility workarounds such as `sys.version` comparison, `__future__`. Also, it removes the Python 2 dedicated codes such as `ArrayConstructor` in Spark.
### Why are the changes needed?
1. Unsupport EOL Python versions
2. Reduce maintenance overhead and remove a bit of legacy codes and hacks for Python 2.
3. PyPy2 has a critical bug that causes a flaky test, SPARK-28358 given my testing and investigation.
4. Users can use Python type hints with Pandas UDFs without thinking about Python version
5. Users can leverage one latest cloudpickle, https://github.com/apache/spark/pull/28950. With Python 3.8+ it can also leverage C pickle.
### Does this PR introduce _any_ user-facing change?
Yes, users cannot use Python 2.7, 3.4 and 3.5 in the upcoming Spark version.
### How was this patch tested?
Manually tested and also tested in Jenkins.
Closes#28957 from HyukjinKwon/SPARK-32138.
Authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
This change replaces the world slave with alternatives matching the context.
### Why are the changes needed?
There is no need to call things slave, we might as well use better clearer names.
### Does this PR introduce _any_ user-facing change?
Yes, the ouput JSON does change. To allow backwards compatibility this is an additive change.
The shell scripts for starting & stopping workers are renamed, and for backwards compatibility old scripts are added to call through to the new ones while printing a deprecation message to stderr.
### How was this patch tested?
Existing tests.
Closes#28864 from holdenk/SPARK-32004-drop-references-to-slave.
Lead-authored-by: Holden Karau <hkarau@apple.com>
Co-authored-by: Holden Karau <holden@pigscanfly.ca>
Signed-off-by: Holden Karau <hkarau@apple.com>
### What changes were proposed in this pull request?
When last partition's splitFile's split size is larger then maxSize, this partition will be lost
Origin logic error like below as 1, 2, 3, 4, 5
```scala
// 5. since index = partition.size now, jump out of the loop , then the last partition is lost since we won't call updatePartition() again.
while (index < partitions.size) {
// 1. we assume that when index = partitions.length -1(the last partition)
val partition = partitions(index)
val fileSplit =
partition.asInstanceOf[HadoopPartition].inputSplit.value.asInstanceOf[FileSplit]
val splitSize = fileSplit.getLength
// 2. if this partition's splitSize > maxSize
if (currentSum + splitSize < maxSize) {
addPartition(partition, splitSize)
index += 1
if (index == partitions.size) {
updateGroups
}
} else {
// 3. if currentGroup.partitions.size >0, this situation is possiable
if (currentGroup.partitions.size == 0) {
addPartition(partition, splitSize)
index += 1
} else {
// 4. then will just call updateGroups() here first, and index won't update in group
updateGroups
}
}
}
groups.toArray
}
}
```
### Why are the changes needed?
Fix bug
### Does this PR introduce any user-facing change?
NO
### How was this patch tested?
Manual code review.
Closes#27988 from AngersZhuuuu/SPARK-31226.
Authored-by: angerszhu <angers.zhu@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
The purpose of this PR is to partly resolve SPARK-29292, and fully resolve SPARK-30010, which should allow Spark to compile vs Scala 2.13 in Spark Core and up through GraphX (not SQL, Streaming, etc).
Note that we are not trying to determine here whether this makes Spark work on 2.13 yet, just compile, as a prerequisite for assessing test outcomes. However, of course, we need to ensure that the change does not break 2.12.
The changes are, in the main, adding .toSeq and .toMap calls where mutable collections / maps are returned as Seq / Map, which are immutable by default in Scala 2.13. The theory is that it should be a no-op for Scala 2.12 (these return themselves), and required for 2.13.
There are a few non-trivial changes highlighted below.
In particular, to get Core to compile, we need to resolve SPARK-30010 which removes a deprecated SparkConf method
### Why are the changes needed?
Eventually, we need to support a Scala 2.13 build, perhaps in Spark 3.1.
### Does this PR introduce _any_ user-facing change?
Yes, removal of the deprecated SparkConf.setAll overload, which isn't legal in Scala 2.13 anymore.
### How was this patch tested?
Existing tests. (2.13 was not _tested_; this is about getting it to compile without breaking 2.12)
Closes#28971 from srowen/SPARK-29292.1.
Authored-by: Sean Owen <srowen@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR proposes to change the HistoryServer to redirect to the history page when we access to /history without application id.
### Why are the changes needed?
In the current master, status code 400 will be returned when we access to /history.
So I wonder it's better to redirect to the history page for the better UX.
### Does this PR introduce _any_ user-facing change?
Yes. In the current master, if we access to /history without application id, we will see like the following page.
After this change applied, we will be redirected to the history page.
### How was this patch tested?
New test added.
Closes#29016 from sarutak/history-redirect.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR aims to run the Spark tests in Github Actions.
To briefly explain the main idea:
- Reuse `dev/run-tests.py` with SBT build
- Reuse the modules in `dev/sparktestsupport/modules.py` to test each module
- Pass the modules to test into `dev/run-tests.py` directly via `TEST_ONLY_MODULES` environment variable. For example, `pyspark-sql,core,sql,hive`.
- `dev/run-tests.py` _does not_ take the dependent modules into account but solely the specified modules to test.
Another thing to note might be `SlowHiveTest` annotation. Running the tests in Hive modules takes too much so the slow tests are extracted and it runs as a separate job. It was extracted from the actual elapsed time in Jenkins:
So, Hive tests are separated into to jobs. One is slow test cases, and the other one is the other test cases.
_Note that_ the current GitHub Actions build virtually copies what the default PR builder on Jenkins does (without other profiles such as JDK 11, Hadoop 2, etc.). The only exception is Kinesis https://github.com/apache/spark/pull/29057/files#diff-04eb107ee163a50b61281ca08f4e4c7bR23
### Why are the changes needed?
Last week and onwards, the Jenkins machines became very unstable for many reasons:
- Apparently, the machines became extremely slow. Almost all tests can't pass.
- One machine (worker 4) started to have the corrupt `.m2` which fails the build.
- Documentation build fails time to time for an unknown reason in Jenkins machine specifically. This is disabled for now at https://github.com/apache/spark/pull/29017.
- Almost all PRs are basically blocked by this instability currently.
The advantages of using Github Actions:
- To avoid depending on few persons who can access to the cluster.
- To reduce the elapsed time in the build - we could split the tests (e.g., SQL, ML, CORE), and run them in parallel so the total build time will significantly reduce.
- To control the environment more flexibly.
- Other contributors can test and propose to fix Github Actions configurations so we can distribute this build management cost.
Note that:
- The current build in Jenkins takes _more than 7 hours_. With Github actions it takes _less than 2 hours_
- We can now control the environments especially for Python easily.
- The test and build look more stable than the Jenkins'.
### Does this PR introduce _any_ user-facing change?
No, dev-only change.
### How was this patch tested?
Tested at https://github.com/HyukjinKwon/spark/pull/4Closes#29057 from HyukjinKwon/migrate-to-github-actions.
Authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
In IPv6 scenario, the current logic to split hostname and port is not correct.
### Why are the changes needed?
to support IPV6 deployment scenario
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
UT and IPV6 spark deployment with yarn
Closes#28931 from PavithraRamachandran/ipv6_issue.
Authored-by: Pavithraramachandran <pavi.rams@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR adds the check to see whether the executor is lost (by asking the `CoarseGrainedSchedulerBackend`) after timeout error raised in `BlockManagerMasterEndponit` due to removing blocks(e.g. RDD, broadcast, shuffle). If the executor is lost, we will ignore the error. Otherwise, throw the error.
### Why are the changes needed?
When removing blocks(e.g. RDD, broadcast, shuffle), `BlockManagerMaserEndpoint` will make RPC calls to each known `BlockManagerSlaveEndpoint` to remove the specific blocks. The PRC call sometimes could end in a timeout when the executor has been lost, but only notified the `BlockManagerMasterEndpoint` after the removing call has already happened. The timeout error could therefore fail the whole job.
In this case, we actually could just ignore the error since those blocks on the lost executor could be considered as removed already.
### Does this PR introduce _any_ user-facing change?
Yes. In case of users hits this issue, they will have the job executed successfully instead of throwing the exception.
### How was this patch tested?
Added unit tests.
Closes#28924 from Ngone51/ignore-timeout-error-for-inactive-executor.
Authored-by: yi.wu <yi.wu@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
This PR proposes to disallow to create `SparkContext` in executors, e.g., in UDFs.
### Why are the changes needed?
Currently executors can create SparkContext, but shouldn't be able to create it.
```scala
sc.range(0, 1).foreach { _ =>
new SparkContext(new SparkConf().setAppName("test").setMaster("local"))
}
```
### Does this PR introduce _any_ user-facing change?
Yes, users won't be able to create `SparkContext` in executors.
### How was this patch tested?
Addes tests.
Closes#28986 from ueshin/issues/SPARK-32160/disallow_spark_context_in_executors.
Authored-by: Takuya UESHIN <ueshin@databricks.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
This patch fixes the test failure due to the missing when statements for destination path. Note that it didn't fail on master branch, because 245aee9 got rid of size call in destination path, but still good to not depend on 245aee9.
### Why are the changes needed?
The build against branch-3.0 / branch-2.4 starts to fail after merging SPARK-32024 (#28859) and this patch will fix it.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Ran modified UT against master / branch-3.0 / branch-2.4.
Closes#29046 from HeartSaVioR/QUICKFIX-SPARK-32024.
Authored-by: Jungtaek Lim (HeartSaVioR) <kabhwan.opensource@gmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
Enable test("static relative links are prefixed with uiRoot (spark.ui.proxyBase)")
### Why are the changes needed?
In Jira, the failed test is another one test("ajax rendered relative links are prefixed with uiRoot (spark.ui.proxyBase)"). This test has been fixed in 6a895d0
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Fix UT
Closes#28970 from warrenzhu25/31723.
Authored-by: Warren Zhu <zhonzh@microsoft.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
Update ApplicationStoreInfo.size to real size during HistoryServerDiskManager initializing.
### Why are the changes needed?
This PR is for fixing bug [32024](https://issues.apache.org/jira/browse/SPARK-32024). We found after history server restart, below error would randomly happen: "java.lang.IllegalStateException: Disk usage tracker went negative (now = -***, delta = -***)" from `HistoryServerDiskManager`.
**Cause**: Reading data from level db would trigger table file compaction, which may also trigger size of level db directory changes. This size change may not be recorded in LevelDB (`ApplicationStoreInfo` in `listing`). When service restarts, `currentUsage` is calculated from real directory size, but `ApplicationStoreInfo` are loaded from leveldb, then `currentUsage` may be less then sum of `ApplicationStoreInfo.size`. In `makeRoom()` function, `ApplicationStoreInfo.size` is used to update usage. Then `currentUsage` becomes negative after several round of `release()` and `lease()` (`makeRoom()`).
**Reproduce**: we can reproduce this issue in dev environment by reducing config value of "spark.history.retainedApplications" and "spark.history.store.maxDiskUsage" to some small values. Here are steps: 1. start history server, load some applications and access some pages (maybe "stages" page to trigger leveldb compaction). 2. restart HS, and refresh pages.
I also added an UT to simulate this case in `HistoryServerDiskManagerSuite`.
**Benefit**: this change would help improve history server reliability.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Add unit test and manually tested it.
Closes#28859 from zhli1142015/update-ApplicationStoreInfo.size-during-disk-manager-initialize.
Authored-by: Zhen Li <zhli@microsoft.com>
Signed-off-by: Jungtaek Lim (HeartSaVioR) <kabhwan.opensource@gmail.com>
### What changes were proposed in this pull request?
Use the invalid value Int.MinValue to fill the map index when the event logs from the old Spark version.
### Why are the changes needed?
Follow up PR for #28941.
### Does this PR introduce _any_ user-facing change?
When we use the Spark version 3.0 history server reading the event log written by the old Spark version, we use the invalid value -2 to fill the map index.
### How was this patch tested?
Existing UT.
Closes#28965 from xuanyuanking/follow-up.
Authored-by: Yuanjian Li <xyliyuanjian@gmail.com>
Signed-off-by: Jungtaek Lim (HeartSaVioR) <kabhwan.opensource@gmail.com>
### What changes were proposed in this pull request?
Use Files.createDirectory() to create local directory instead of File.mkdir() in DiskBlockManager.
Many times, we will see such error log information like "Failed to create local dir in xxxxxx". But there is no clear information indicating why the directory creation failed.
When Files.createDirectory() fails to create a local directory, it can give specific error information for subsequent troubleshooting(also throws IOException).
### Why are the changes needed?
Throw clear error message when creating directory fails.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
`DiskBlockManagerSuite`
Closes#28997 from sidedoorleftroad/SPARK-32172.
Authored-by: sidedoorleftroad <sidedoorleftroad@163.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR aims to reduce the required test resources in WorkerDecommissionExtendedSuite.
### Why are the changes needed?
When Jenkins farms is crowded, the following failure happens currently [here](https://amplab.cs.berkeley.edu/jenkins/view/Spark%20QA%20Test%20(Dashboard)/job/spark-master-test-sbt-hadoop-3.2-hive-2.3/890/testReport/junit/org.apache.spark.scheduler/WorkerDecommissionExtendedSuite/Worker_decommission_and_executor_idle_timeout/)
```
java.util.concurrent.TimeoutException: Can't find 20 executors before 60000 milliseconds elapsed
at org.apache.spark.TestUtils$.waitUntilExecutorsUp(TestUtils.scala:326)
at org.apache.spark.scheduler.WorkerDecommissionExtendedSuite.$anonfun$new$2(WorkerDecommissionExtendedSuite.scala:45)
```
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Pass the Jenkins.
Closes#29001 from dongjoon-hyun/SPARK-32100-2.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR changes `webui.css` to fix a style issue on moving mouse cursor on the Spark logo.
### Why are the changes needed?
In the webui, the Spark logo is on the top right side.
When we move mouse cursor on the logo, a weird underline appears near the logo.
<img width="209" alt="logo_with_line" src="https://user-images.githubusercontent.com/4736016/86542828-3c6a9f00-bf54-11ea-9b9d-cc50c12c2c9b.png">
### Does this PR introduce _any_ user-facing change?
Yes. After this change applied, no more weird line shown even if mouse cursor moves on the logo.
<img width="207" alt="removed-line-from-logo" src="https://user-images.githubusercontent.com/4736016/86542877-98cdbe80-bf54-11ea-8695-ee39689673ab.png">
### How was this patch tested?
By moving mouse cursor on the Spark logo and confirmed no more weird line there.
Closes#29003 from sarutak/fix-logo-underline.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR aims to add `PrometheusServletSuite`.
### Why are the changes needed?
This improves the test coverage.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Pass the newly added test suite.
Closes#28865 from erenavsarogullari/spark_driver_prometheus_metrics_improvement.
Authored-by: Eren Avsarogullari <erenavsarogullari@gmail.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
`formatDate` in utils.js `org/apache/spark/ui/static/utils.js` is partly refactored.
### Why are the changes needed?
In branch-2.4，task launch time is returned as html string from driver,
while in branch-3.x，this is returned in JSON Object as`Date`type from `org.apache.spark.status.api.v1.TaskData`
Due to：
LaunchTime from jersey server in spark driver is correct, which will be converted to date string like `2020-06-28T02:57:42.605GMT` in json object, then the formatDate in utils.js treat it as date.split(".")[0].replace("T", " ").
So `2020-06-28T02:57:42.605GMT` will be converted to `2020-06-28 02:57:42`, but correct is `2020-06-28 10:57:42` in GMT+8 timezone.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Manual test.
Closes#28918 from TJX2014/master-SPARK-32068-ui-task-lauch-time-tz.
Authored-by: TJX2014 <xiaoxingstack@gmail.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
### What changes were proposed in this pull request?
This PR tries to unify the method `getReader` and `getReaderForRange` in `ShuffleManager`.
### Why are the changes needed?
Reduce the duplicate codes, simplify the implementation, and for better maintenance.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Covered by existing tests.
Closes#28895 from Ngone51/unify-getreader.
Authored-by: yi.wu <yi.wu@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
Fix bug of exception when parse event log of fetch failed task end reason without `Map Index`
### Why are the changes needed?
When Spark history server read event log produced by older version of spark 2.4 (which don't have `Map Index` field), parsing of TaskEndReason will fail. This will cause TaskEnd event being ignored.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
JsonProtocolSuite.test("FetchFailed Map Index backwards compatibility")
Closes#28941 from warrenzhu25/shs-task.
Authored-by: Warren Zhu <zhonzh@microsoft.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
**First**
`DAGSchedulerSuite` provides `completeNextStageWithFetchFailure` to make all tasks in non first stage occur `FetchFailed`.
But many test case uses complete directly as follows:
```scala
complete(taskSets(1), Seq(
(FetchFailed(makeBlockManagerId("hostA"),
shuffleDep1.shuffleId, 0L, 0, 0, "ignored"), null)))
```
We need to reuse `completeNextStageWithFetchFailure`.
**Second**
`DAGSchedulerSuite` also check the results show below:
```scala
complete(taskSets(0), Seq((Success, 42)))
assert(results === Map(0 -> 42))
```
We can extract it as a generic method of `checkAnswer`.
### Why are the changes needed?
Reuse `completeNextStageWithFetchFailure`
### Does this PR introduce _any_ user-facing change?
'No'.
### How was this patch tested?
Jenkins test
Closes#28866 from beliefer/reuse-completeNextStageWithFetchFailure.
Authored-by: gengjiaan <gengjiaan@360.cn>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR aims to add `WorkerDecomissionExtendedSuite` for various worker decommission combinations.
### Why are the changes needed?
This will improve the test coverage.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Pass the Jenkins.
Closes#28929 from dongjoon-hyun/SPARK-WD-TEST.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
Fix app id link for multi attempts application in history summary page
If attempt id is available (yarn), app id link url will contain correct attempt id, like `/history/application_1561589317410_0002/1/jobs/`.
If attempt id is not available (standalone), app id link url will not contain fake attempt id, like `/history/app-20190404053606-0000/jobs/`.
### Why are the changes needed?
This PR is for fixing [32028](https://issues.apache.org/jira/browse/SPARK-32028). App id link use application attempt count as attempt id. this would cause link url wrong for below cases:
1. there are multi attempts, all links point to last attempt
2. if there is one attempt, but attempt id is not 1 (before attempt maybe crash or fail to gerenerate event file). link url points to worng attempt (1) here.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Tested this manually.
Closes#28867 from zhli1142015/fix-appid-link-in-history-page.
Authored-by: Zhen Li <zhli@microsoft.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
Introduces the concept of a `MapOutputMetadata` opaque object that can be returned from map output writers.
Note that this PR only proposes the API changes on the shuffle writer side. Following patches will be proposed for actually accepting the metadata on the driver and persisting it in the driver's shuffle metadata storage plugin.
### Why are the changes needed?
For a more complete design discussion on this subject as a whole, refer to [this design document](https://docs.google.com/document/d/1Aj6IyMsbS2sdIfHxLvIbHUNjHIWHTabfknIPoxOrTjk/edit#).
### Does this PR introduce any user-facing change?
Enables additional APIs for the shuffle storage plugin tree. Usage will become more apparent as the API evolves.
### How was this patch tested?
No tests here, since this is only an API-side change that is not consumed by core Spark itself.
Closes#28616 from mccheah/return-map-output-metadata.
Authored-by: mcheah <mcheah@palantir.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
This PR changes the test to get an active executorId and set it as preferred location instead of setting a fixed preferred location.
### Why are the changes needed?
The test is flaky. After checking the [log](https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/124086/artifact/core/), I find the root cause is:
Two test cases from different test suites got submitted at the same time because of concurrent execution. In this particular case, the two test cases (from DistributedSuite and BarrierTaskContextSuite) both launch under local-cluster mode. The two applications are submitted at the SAME time so they have the same applications(app-20200615210132-0000). Thus, when the cluster of BarrierTaskContextSuite is launching executors, it failed to create the directory for the executor 0, because the path (/home/jenkins/workspace/work/app-app-20200615210132-0000/0) has been used by the cluster of DistributedSuite. Therefore, it has to launch executor 1 and 2 instead, that lead to non of the tasks can get preferred locality thus they got scheduled together and lead to the test failure.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
The test can not be reproduced locally. We can only know it's been fixed when it's no longer flaky on Jenkins.
Closes#28849 from Ngone51/fix-spark-32000.
Authored-by: yi.wu <yi.wu@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
This aims to replace `executorIdsToBeRemoved` with `executorIdToBeRemoved`.
### Why are the changes needed?
Since a wrong variable is used currently, `ArrayBuffer()` is always displayed.
```
20/06/16 19:33:31 WARN ExecutorAllocationManager: Not removing executor ArrayBuffer() because the ResourceProfile was UNKNOWN!
```
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Manual.
Closes#28847 from dongjoon-hyun/SPARK-29148.
Authored-by: Dongjoon Hyun <dongjoon@apache.org>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
Close LevelDBIterator when LevelDB.close() is called.
### Why are the changes needed?
This pull request would prevent JNI resources leaking from Level DB instance and its' iterators. In before implementation JNI resources from LevelDBIterator are cleaned by finalize() function. This behavior is also mentioned in comments of ["LevelDBIterator.java"](https://github.com/apache/spark/blob/master/common/kvstore/src/main/java/org/apache/spark/util/kvstore/LevelDBIterator.java) by squito . But if DB instance is already closed, then iterator's close method would be ignored. LevelDB's iterator would keep level db files opened (for the case table cache is filled up), till iterator.close() is called. Then these JNI resources (file handle) would be leaked.
This JNI resource leaking issue would cause the problem described in [SPARK-31929](https://issues.apache.org/jira/browse/SPARK-31929) on Windows: in spark history server, leaked file handle for level db files would trigger "IOException" when HistoryServerDiskManager try to remove them for releasing disk space.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Add unit test and manually tested it.
Closes#28769 from zhli1142015/close-leveldbiterator-when-leveldb.close.
Authored-by: Zhen Li <zhli@microsoft.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
`DAGSchedulerSuite `provides `completeShuffleMapStageSuccessfully `to make `ShuffleMapStage `successfully.
But many test case uses complete directly as follows:
`complete(taskSets(0), Seq((Success, makeMapStatus("hostA", 1))))`
We need to improve `completeShuffleMapStageSuccessfully `and reuse it.
`completeShuffleMapStageSuccessfully(0, 0, 1, Some(0), Seq("hostA"))`
### Why are the changes needed?
Improve and reuse completeShuffleMapStageSuccessfully
### Does this PR introduce _any_ user-facing change?
'No'.
### How was this patch tested?
Jenkins test
Closes#28641 from beliefer/improve-and-reuse-method.
Authored-by: gengjiaan <gengjiaan@360.cn>
Signed-off-by: Xingbo Jiang <xingbo.jiang@databricks.com>
### What changes were proposed in this pull request?
This PR proposes to use "mdc.XXX" as the consistent key for both `sc.setLocalProperty` and `log4j.properties` when setting up configurations for MDC.
### Why are the changes needed?
It's weird that we use "mdc.XXX" as key to set MDC value via `sc.setLocalProperty` while we use "XXX" as key to set MDC pattern in log4j.properties. It could also bring extra burden to the user.
### Does this PR introduce _any_ user-facing change?
No, as MDC feature is added in version 3.1, which hasn't been released.
### How was this patch tested?
Tested manually.
Closes#28801 from Ngone51/consistent-mdc.
Authored-by: yi.wu <yi.wu@databricks.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### 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.
-->
This PR enriches the exception message when application summary is not available.
#28444 covers the case when application information is not available but the case application summary is not available is not covered.
### 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.
-->
To complement #28444 .
### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such as the documentation fix.
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 possible, please also clarify if this is a user-facing change compared to the released Spark versions or within the unreleased branches such as master.
If no, write 'No'.
-->
Yes.
Before this change, we can get the following error message when we access to `/jobs` if application summary is not available.
<img width="707" alt="no-such-element-exception-error-message" src="https://user-images.githubusercontent.com/4736016/84562182-6aadf200-ad8d-11ea-8980-d63edde6fad6.png">
After this change, we can get the following error message. It's like #28444 does.
<img width="1349" alt="enriched-errorm-message" src="https://user-images.githubusercontent.com/4736016/84562189-85806680-ad8d-11ea-8346-4da2ec11df2b.png">
### 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.
-->
I checked with the following procedure.
1. Set breakpoint in the line of `kvstore.write(appSummary)` in `AppStatusListener#onStartApplicatin`. Only the thread reaching this line should be suspended.
2. Start spark-shell and wait few seconds.
3. Access to `/jobs`
Closes#28820 from sarutak/fix-no-such-element.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
After #28192, the job list page becomes very slow.
For example, after the following operation, the UI loading can take >40 sec.
```
(1 to 1000).foreach(_ => sc.parallelize(1 to 10).collect)
```
This is caused by a [performance issue of `vis-timeline`](https://github.com/visjs/vis-timeline/issues/379). The serious issue affects both branch-3.0 and branch-2.4
I tried a different version 4.21.0 from https://cdnjs.com/libraries/vis
The infinite drawing issue seems also fixed if the zoom is disabled as default.
### Why are the changes needed?
Fix the serious perf issue in web UI by falling back vis-timeline-graph2d to an ealier version.
### Does this PR introduce _any_ user-facing change?
Yes, fix the UI perf regression
### How was this patch tested?
Manual test
Closes#28806 from gengliangwang/downgradeVis.
Authored-by: Gengliang Wang <gengliang.wang@databricks.com>
Signed-off-by: Gengliang Wang <gengliang.wang@databricks.com>
### What changes were proposed in this pull request?
#28747 reverted #28439 due to some flaky test case. This PR fixes the flaky test and adds pagination support.
### Why are the changes needed?
To support pagination for streaming tab
### Does this PR introduce _any_ user-facing change?
Yes, Now streaming tab tables will be paginated.
### How was this patch tested?
Manually.
Closes#28748 from iRakson/fixstreamingpagination.
Authored-by: iRakson <raksonrakesh@gmail.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
This PR upgrades HtmlUnit.
Selenium and Jetty also upgraded because of dependency.
### Why are the changes needed?
Recently, a security issue which affects HtmlUnit is reported.
https://nvd.nist.gov/vuln/detail/CVE-2020-5529
According to the report, arbitrary code can be run by malicious users.
HtmlUnit is used for test so the impact might not be large but it's better to upgrade it just in case.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Existing testcases.
Closes#28585 from sarutak/upgrade-htmlunit.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
This PR is a followup of #26624. This PR cleans up MDC properties if the original value is empty.
Besides, this PR adds a warning and ignore the value when the user tries to override the value of `taskName`.
### Why are the changes needed?
Before this PR, running the following jobs:
```
sc.setLocalProperty("mdc.my", "ABC")
sc.parallelize(1 to 100).count()
sc.setLocalProperty("mdc.my", null)
sc.parallelize(1 to 100).count()
```
there's still MDC value "ABC" in the log of the second count job even if we've unset the value.
### Does this PR introduce _any_ user-facing change?
Yes, user will 1) no longer see the MDC values after unsetting the value; 2) see a warning if he/she tries to override the value of `taskName`.
### How was this patch tested?
Tested Manaually.
Closes#28756 from Ngone51/followup-8981.
Authored-by: yi.wu <yi.wu@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
After SPARK-31632 SparkException is thrown from def applicationInfo
`def applicationInfo(): v1.ApplicationInfo = {
try {
// The ApplicationInfo may not be available when Spark is starting up.
store.view(classOf[ApplicationInfoWrapper]).max(1).iterator().next().info
} catch {
case _: NoSuchElementException =>
throw new SparkException("Failed to get the application information. " +
"If you are starting up Spark, please wait a while until it's ready.")
}
}`
Where as the caller for this method def getSparkUser in Spark UI is not handling SparkException in the catch
`def getSparkUser: String = {
try {
Option(store.applicationInfo().attempts.head.sparkUser)
.orElse(store.environmentInfo().systemProperties.toMap.get("user.name"))
.getOrElse("<unknown>")
} catch {
case _: NoSuchElementException => "<unknown>"
}
}`
So On using this method (getSparkUser )we can get the application erred out.
As the part of this PR we will replace SparkException to NoSuchElementException for applicationInfo in AppStatusStore
### Why are the changes needed?
On invoking the method getSparkUser, we can get the SparkException on calling store.applicationInfo(). And this is not handled in the catch block and getSparkUser will error out in this scenario
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Done the manual testing using the spark-shell and spark-submit
Closes#28768 from SaurabhChawla100/SPARK-31941.
Authored-by: SaurabhChawla <saurabhc@qubole.com>
Signed-off-by: Kousuke Saruta <sarutak@oss.nttdata.com>
### What changes were proposed in this pull request?
This PR replaces constant config with the `ConfigEntry`.
### Why are the changes needed?
For better code maintenance.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Tested manually.
Closes#28775 from Ngone51/followup-SPARK-31486.
Authored-by: yi.wu <yi.wu@databricks.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
This PR proposes to make `PythonFunction` holds `Seq[Byte]` instead of `Array[Byte]` to be able to compare if the byte array has the same values for the cache manager.
### Why are the changes needed?
Currently the cache manager doesn't use the cache for `udf` if the `udf` is created again even if the functions is the same.
```py
>>> func = lambda x: x
>>> df = spark.range(1)
>>> df.select(udf(func)("id")).cache()
```
```py
>>> df.select(udf(func)("id")).explain()
== Physical Plan ==
*(2) Project [pythonUDF0#14 AS <lambda>(id)#12]
+- BatchEvalPython [<lambda>(id#0L)], [pythonUDF0#14]
+- *(1) Range (0, 1, step=1, splits=12)
```
This is because `PythonFunction` holds `Array[Byte]`, and `equals` method of array equals only when the both array is the same instance.
### Does this PR introduce _any_ user-facing change?
Yes, if the user reuse the Python function for the UDF, the cache manager will detect the same function and use the cache for it.
### How was this patch tested?
I added a test case and manually.
```py
>>> df.select(udf(func)("id")).explain()
== Physical Plan ==
InMemoryTableScan [<lambda>(id)#12]
+- InMemoryRelation [<lambda>(id)#12], StorageLevel(disk, memory, deserialized, 1 replicas)
+- *(2) Project [pythonUDF0#5 AS <lambda>(id)#3]
+- BatchEvalPython [<lambda>(id#0L)], [pythonUDF0#5]
+- *(1) Range (0, 1, step=1, splits=12)
```
Closes#28774 from ueshin/issues/SPARK-31945/udf_cache.
Authored-by: Takuya UESHIN <ueshin@databricks.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
This PR adds the check to see whether the allocated executors for the waiting application is empty before recognizing it as a possible hang application.
### Why are the changes needed?
It's a bugfix. The warning means there are not enough resources for the application to launch at least one executor. But we can still successfully run a job under this warning, which means it does have launched executor.
### Does this PR introduce _any_ user-facing change?
Yes. Before this PR, when using local cluster mode to start spark-shell, e.g. `./bin/spark-shell --master "local-cluster[2, 1, 1024]"`, the user would always see the warning:
```
20/06/06 22:21:02 WARN Utils: Your hostname, C02ZQ051LVDR resolves to a loopback address: 127.0.0.1; using 192.168.1.6 instead (on interface en0)
20/06/06 22:21:02 WARN Utils: Set SPARK_LOCAL_IP if you need to bind to another address
20/06/06 22:21:02 WARN NativeCodeLoader: Unable to load native-hadoop library for your platform... using builtin-java classes where applicable
Setting default log level to "WARN".
To adjust logging level use sc.setLogLevel(newLevel). For SparkR, use setLogLevel(newLevel).
NOTE: SPARK_PREPEND_CLASSES is set, placing locally compiled Spark classes ahead of assembly.
NOTE: SPARK_PREPEND_CLASSES is set, placing locally compiled Spark classes ahead of assembly.
Spark context Web UI available at http://192.168.1.6:4040
Spark context available as 'sc' (master = local-cluster[2, 1, 1024], app id = app-20200606222107-0000).
Spark session available as 'spark'.
20/06/06 22:21:07 WARN Master: App app-20200606222107-0000 requires more resource than any of Workers could have.
20/06/06 22:21:07 WARN Master: App app-20200606222107-0000 requires more resource than any of Workers could have.
```
After this PR, the warning has gone.
### How was this patch tested?
Tested manually.
Closes#28742 from Ngone51/fix_warning.
Authored-by: yi.wu <yi.wu@databricks.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
These changes implement an application wait mechanism which will allow spark-submit to wait until the application finishes in Standalone Spark Mode. This will delay the exit of spark-submit JVM until the job is completed. This implementation will keep monitoring the application until it is either finished, failed or killed. This will be controlled via a flag (spark.submit.waitForCompletion) which will be set to false by default.
### Why are the changes needed?
Currently, Livy API for Standalone Cluster Mode doesn't know when the job has finished. If this flag is enabled, this can be used by Livy API (/batches/{batchId}/state) to find out when the application has finished/failed. This flag is Similar to spark.yarn.submit.waitAppCompletion.
### Does this PR introduce any user-facing change?
Yes, this PR introduces a new flag but it will be disabled by default.
### How was this patch tested?
Couldn't implement unit tests since the pollAndReportStatus method has System.exit() calls. Please provide any suggestions.
Tested spark-submit locally for the following scenarios:
1. With the flag enabled, spark-submit exits once the job is finished.
2. With the flag enabled and job failed, spark-submit exits when the job fails.
3. With the flag disabled, spark-submit exists post submitting the job (existing behavior).
4. Existing behavior is unchanged when the flag is not added explicitly.
Closes#28258 from akshatb1/master.
Lead-authored-by: Akshat Bordia <akshat.bordia31@gmail.com>
Co-authored-by: Akshat Bordia <akshat.bordia@citrix.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
Ignore internal accumulators that use unrecognized types rather than crashing so that an event log containing such accumulators can still be converted to JSON and logged.
### Why are the changes needed?
A user may use internal accumulators by adding the `internal.metrics.` prefix to the accumulator name to hide sensitive information from UI (Accumulators except internal ones will be shown in Spark UI).
However, `org.apache.spark.util.JsonProtocol.accumValueToJson` assumes an internal accumulator has only 3 possible types: `int`, `long`, and `java.util.List[(BlockId, BlockStatus)]`. When an internal accumulator uses an unexpected type, it will crash.
An event log that contains such accumulator will be dropped because it cannot be converted to JSON, and it will cause weird UI issue when rendering in Spark History Server. For example, if `SparkListenerTaskEnd` is dropped because of this issue, the user will see the task is still running even if it was finished.
It's better to make `accumValueToJson` more robust because it's up to the user to pick up the accumulator name.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
The new unit tests.
Closes#28744 from zsxwing/fix-internal-accum.
Authored-by: Shixiong Zhu <zsxwing@gmail.com>
Signed-off-by: Shixiong Zhu <zsxwing@gmail.com>
This PR reverts #28439 due to that PR breaks QA build.
Closes#28747 from sarutak/revert-SPARK-30119.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Kousuke Saruta <sarutak@oss.nttdata.com>
### What changes were proposed in this pull request?
* Pagination Support is added to all tables of streaming page in spark web UI.
For adding pagination support, existing classes from #7399 were used.
* Earlier streaming page has two tables `Active Batches` and `Completed Batches`. Now, we will have three tables `Running Batches`, `Waiting Batches` and `Completed Batches`. If we have large number of waiting and running batches then keeping track in a single table is difficult. Also other pages have different table for different type type of data.
* Earlier empty tables were shown. Now only non-empty tables will be shown.
`Active Batches` table used to show details of waiting batches followed by running batches.
### Why are the changes needed?
Pagination will allow users to analyse the table in much better way. All spark web UI pages support pagination apart from streaming pages, so this will add consistency as well. Also it might fix the potential OOM errors that can arise.
### Does this PR introduce _any_ user-facing change?
Yes. `Active Batches` table is split into two tables `Running Batches` and `Waiting Batches`. Pagination Support is added to the all the tables. Every other functionality is unchanged.
### How was this patch tested?
Manually.
Before changes:
<img width="1667" alt="Screenshot 2020-05-03 at 7 07 14 PM" src="https://user-images.githubusercontent.com/15366835/80915680-8fb44b80-8d71-11ea-9957-c4a3769b8b67.png">
After Changes:
<img width="1669" alt="Screenshot 2020-05-03 at 6 51 22 PM" src="https://user-images.githubusercontent.com/15366835/80915694-a9ee2980-8d71-11ea-8fc5-246413a4951d.png">
Closes#28439 from iRakson/streamingPagination.
Authored-by: iRakson <raksonrakesh@gmail.com>
Signed-off-by: Kousuke Saruta <sarutak@oss.nttdata.com>
### What changes were proposed in this pull request?
This PR fixes a wrong coloring issue in the DAG-viz.
In the Job Page and Stage Page, nodes which are associated with "barrier mode" in the DAG-viz will be colored pale green.
But, with some type of jobs, nodes which are not associated with the mode will also colored.
You can reproduce with the following operation.
```
sc.parallelize(1 to 10).barrier.mapPartitions(identity).repartition(1).collect()
```
<img width="376" alt="wrong-coloring" src="https://user-images.githubusercontent.com/4736016/83403670-1711df00-a444-11ea-9457-c683f75bc566.png">
In the screen shot above, `repartition` in `Stage 1` is not associated with barrier mode so the corresponding node should not be colored pale green.
The cause of this issue is the logic which chooses HTML elements to be colored is wrong.
The logic chooses such elements based on whether each element is associated with a style class (`clusterId` in the code).
But when an operation crosses over shuffle (like `repartition` above), a `clusterId` can be duplicated and non-barrier mode node is also associated with the same `clusterId`.
### Why are the changes needed?
This is a bug.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Newly added test case with the following command.
```
build/sbt -Dtest.default.exclude.tags= -Dspark.test.webdriver.chrome.driver=/path/to/chromedriver "testOnly org.apache.spark.ui.ChromeUISeleniumSuite -- -z SPARK-31886"
```
Closes#28694 from sarutak/fix-wrong-barrier-color.
Authored-by: Kousuke Saruta <sarutak@oss.nttdata.com>
Signed-off-by: Gengliang Wang <gengliang.wang@databricks.com>
