The security manager adds a lot of overhead to the runtime of the
app, and causes a severe performance regression. Even stubbing out
all unneeded methods (all except checkExit()) does not help.

So, instead, penalize users who do an explicit System.exit() by leaving
them in "undefined behavior" territory: if they do that, the Yarn
backend won't be able to report the final app status to the RM.
The result is that the final status of the application might not match
the user's expectations.

One side-effect of the change is that users who do an explicit
System.exit() will lose the AM retry functionality. Since there is
no way to know if the exit was because of success or failure, the
AM right now errs on the side of it being a successful exit.

Author: Marcelo Vanzin <vanzin@cloudera.com>

Closes #3484 from vanzin/SPARK-4584 and squashes the following commits:

21f2502 [Marcelo Vanzin] Do not retry apps that use System.exit().
4198b3b [Marcelo Vanzin] [SPARK-4584] [yarn] Remove security manager from Yarn AM.

Author: Takuya UESHIN <ueshin@happy-camper.st>

Closes #3058 from ueshin/issues/SPARK-4193 and squashes the following commits:

e096bb1 [Takuya UESHIN] Add a plugin declaration to pluginManagement.
6762ec2 [Takuya UESHIN] Fix usage of -Xdoclint javadoc option.
fdb280a [Takuya UESHIN] Fix Javadoc errors.
4745f3c [Takuya UESHIN] Merge branch 'master' into issues/SPARK-4193
923e2f0 [Takuya UESHIN] Use doclint option `-missing` instead of `none`.
30d6718 [Takuya UESHIN] Fix Javadoc errors.
b548017 [Takuya UESHIN] Disable doclint in Java 8 to prevent from build error.

The old location will return a 404.

Author: Daoyuan Wang <daoyuan.wang@intel.com>

Closes #3504 from adrian-wang/repo and squashes the following commits:

f604e05 [Daoyuan Wang] already in maven central, remove at all
f494fac [Daoyuan Wang] spark staging repo outdated

when building spark by sbt, the function “runAlternateBoot" in sbt/sbt-launch-lib.bash is not used. And this function is not used by spark code. So I think this function is not necessary. And the option of "sbt.boot.properties" can be configured in the command line when building spark, eg:
sbt/sbt assembly -Dsbt.boot.properties=$bootpropsfile.

The file from https://github.com/sbt/sbt-launcher-package is changed. And the function “runAlternateBoot" is deleted in upstream project. I think spark project should delete this function in file sbt/sbt-launch-lib.bash. Thanks.

Author: KaiXinXiaoLei <huleilei1@huawei.com>

Closes #3224 from KaiXinXiaoLei/deleteFunction and squashes the following commits:

e8eac49 [KaiXinXiaoLei] Delete blank lines.
efe36d4 [KaiXinXiaoLei] Delete unnecessary function

This PR disables HiveThriftServer2 asynchronous execution by setting `runInBackground` argument in `ExecuteStatementOperation` to `false`, and reverting `SparkExecuteStatementOperation.run` in Hive 13 shim to Hive 12 version. This change makes Simba ODBC driver v1.0.0.1000 work.

<!-- Reviewable:start -->
[<img src="https://reviewable.io/review_button.png" height=40 alt="Review on Reviewable"/>](https://reviewable.io/reviews/apache/spark/3506)
<!-- Reviewable:end -->

Author: Cheng Lian <lian@databricks.com>

Closes #3506 from liancheng/disable-async-exec and squashes the following commits:

593804d [Cheng Lian] Disables asynchronous execution in Hive 0.13.1 HiveThriftServer2

Time suffix exists in Utils.getUsedTimeMs(startTime), no need to append again, delete that

Author: maji2014 <maji3@asiainfo.com>

Closes #3475 from maji2014/SPARK-4619 and squashes the following commits:

df0da4e [maji2014] delete redundant time suffix

This PR introduces a set of Java APIs for using `JdbcRDD`:

1. Trait (interface) `JdbcRDD.ConnectionFactory`: equivalent to the `getConnection: () => Connection` parameter in `JdbcRDD` constructor.
2. Two overloaded versions of `Jdbc.create`: used to create `JavaRDD` that wraps a `JdbcRDD`.

<!-- Reviewable:start -->
[<img src="https://reviewable.io/review_button.png" height=40 alt="Review on Reviewable"/>](https://reviewable.io/reviews/apache/spark/3478)
<!-- Reviewable:end -->

Author: Cheng Lian <lian@databricks.com>

Closes #3478 from liancheng/japi-jdbc-rdd and squashes the following commits:

9a54625 [Cheng Lian] Only shutdowns a single DB rather than the whole Derby driver
d4cedc5 [Cheng Lian] Moves Java JdbcRDD test case to a separate test suite
ffcdf2e [Cheng Lian] Java API for JdbcRDD

Author: roxchkplusony <roxchkplusony@gmail.com>

Closes #3483 from roxchkplusony/bugfix/4626 and squashes the following commits:

aba9184 [roxchkplusony] replace warning message per review
5e7fdea [roxchkplusony] [SPARK-4626] Kill a task only if the executorId is (still) registered with the scheduler

Warn against subclassing scala.App, and remove one instance of this in examples

Author: Sean Owen <sowen@cloudera.com>

Closes #3497 from srowen/SPARK-4170 and squashes the following commits:

4a6131f [Sean Owen] Restore multiline string formatting
a8ca895 [Sean Owen] Warn against subclassing scala.App, and remove one instance of this in examples

This commit provides a script that computes the contributors list
by linking the github commits with JIRA issues. Automatically
translating github usernames remains a TODO at this point.

https://issues.apache.org/jira/browse/SPARK-3628

In current implementation, the accumulator will be updated for every successfully finished task, even the task is from a resubmitted stage, which makes the accumulator counter-intuitive

In this patch, I changed the way for the DAGScheduler to update the accumulator,

DAGScheduler maintains a HashTable, mapping the stage id to the received <accumulator_id , value> pairs. Only when the stage becomes independent, (no job needs it any more), we accumulate the values of the <accumulator_id , value> pairs, when a task finished, we check if the HashTable has contained such stageId, it saves the accumulator_id, value only when the task is the first finished task of a new stage or the stage is running for the first attempt...

Author: CodingCat <zhunansjtu@gmail.com>

Closes #2524 from CodingCat/SPARK-732-1 and squashes the following commits:

701a1e8 [CodingCat] roll back change on Accumulator.scala
1433e6f [CodingCat] make MIMA happy
b233737 [CodingCat] address Matei's comments
02261b8 [CodingCat] rollback  some changes
6b0aff9 [CodingCat] update document
2b2e8cf [CodingCat] updateAccumulator
83b75f8 [CodingCat] style fix
84570d2 [CodingCat] re-enable  the bad accumulator guard
1e9e14d [CodingCat] add NPE guard
21b6840 [CodingCat] simplify the patch
88d1f03 [CodingCat] fix rebase error
f74266b [CodingCat] add test case for resubmitted result stage
5cf586f [CodingCat] de-duplicate on task level
138f9b3 [CodingCat] make MIMA happy
67593d2 [CodingCat] make if allowing duplicate update as an option of accumulator

Before we have a full picture of the operators we want to add, it might be safer to hide `Matrix.transposeMultiply` in 1.2.0. Another update we want to change is `Matrix.randn` and `Matrix.rand`, both of which should take a `Random` implementation. Otherwise, it is very likely to produce inconsistent RDDs. I also added some unit tests for matrix factory methods. All APIs are new in 1.2, so there is no incompatible changes.

brkyvz

Author: Xiangrui Meng <meng@databricks.com>

Closes #3468 from mengxr/SPARK-4614 and squashes the following commits:

3b0e4e2 [Xiangrui Meng] add mima excludes
6bfd8a4 [Xiangrui Meng] hide transposeMultiply; add rng to rand and randn; add unit tests

After additional discussion with rxin, I think having all the possible `TripletField` options is confusing.  This pull request reduces the triplet fields to:

```java
  /**
   * None of the triplet fields are exposed.
   */
  public static final TripletFields None = new TripletFields(false, false, false);

  /**
   * Expose only the edge field and not the source or destination field.
   */
  public static final TripletFields EdgeOnly = new TripletFields(false, false, true);

  /**
   * Expose the source and edge fields but not the destination field. (Same as Src)
   */
  public static final TripletFields Src = new TripletFields(true, false, true);

  /**
   * Expose the destination and edge fields but not the source field. (Same as Dst)
   */
  public static final TripletFields Dst = new TripletFields(false, true, true);

  /**
   * Expose all the fields (source, edge, and destination).
   */
  public static final TripletFields All = new TripletFields(true, true, true);
```

Author: Joseph E. Gonzalez <joseph.e.gonzalez@gmail.com>

Closes #3472 from jegonzal/SimplifyTripletFields and squashes the following commits:

91796b5 [Joseph E. Gonzalez] removing confusing triplet fields

[SPARK-4612] Reduce task latency and increase scheduling throughput by making configuration initialization lazy

https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/executor/Executor.scala#L337 creates a configuration object for every task that is launched, even if there is no new dependent file/JAR to update. This is a heavy-weight creation that should be avoided if there is no new file/JAR to update. This PR makes that creation lazy. Quick local test in spark-perf scheduling throughput tests gives the following numbers in a local standalone scheduler mode.
1 job with 10000 tasks: before 7.8395 seconds, after 2.6415 seconds = 3x increase in task scheduling throughput

pwendell JoshRosen

Author: Tathagata Das <tathagata.das1565@gmail.com>

Closes #3463 from tdas/lazy-config and squashes the following commits:

c791c1e [Tathagata Das] Reduce task latency by making configuration initialization lazy

Turns out we are allocating an allocator pool for every TransportClient (which means that the number increases with the number of nodes in the cluster), when really we should just reuse one for all clients.

This patch, as expected, greatly decreases off-heap memory allocation, and appears to make allocation only proportional to the number of cores.

Author: Aaron Davidson <aaron@databricks.com>

Closes #3465 from aarondav/fewer-pools and squashes the following commits:

36c49da [Aaron Davidson] [SPARK-4516] Avoid allocating unnecessarily Netty PooledByteBufAllocators

In practice, only 2-4 cores should be required to transfer roughly 10 Gb/s, and each core that we use will have an initial overhead of roughly 32 MB of off-heap memory, which comes at a premium.

Thus, this value should still retain maximum throughput and reduce wasted off-heap memory allocation. It can be overridden by setting the number of serverThreads and clientThreads manually in Spark's configuration.

Author: Aaron Davidson <aaron@databricks.com>

Closes #3469 from aarondav/fewer-pools2 and squashes the following commits:

087c59f [Aaron Davidson] [SPARK-4516] Cap default number of Netty threads at 8

User could construct an MF model directly. I added a note about the performance.

Author: Xiangrui Meng <meng@databricks.com>

Closes #3459 from mengxr/SPARK-4604 and squashes the following commits:

f64bcd3 [Xiangrui Meng] organize imports
ed08214 [Xiangrui Meng] check preconditions and unit tests
a624c12 [Xiangrui Meng] make MatrixFactorizationModel public

Currently, the LogLoss used by GradientBoostedTrees has 2 issues:
* the gradient (and therefore loss) does not match that used by Friedman (1999)
* the error computation uses 0/1 accuracy, not log loss

This PR updates LogLoss.
It also adds some doc for boosting and forests.

I tested it on sample data and made sure the log loss is monotonically decreasing with each boosting iteration.

CC: mengxr manishamde codedeft

Author: Joseph K. Bradley <joseph@databricks.com>

Closes #3439 from jkbradley/gbt-loss-fix and squashes the following commits:

cfec17e [Joseph K. Bradley] removed forgotten temp comments
a27eb6d [Joseph K. Bradley] corrections to last log loss commit
ed5da2c [Joseph K. Bradley] updated LogLoss (boosting) for numerical stability
5e52bff [Joseph K. Bradley] * Removed the 1/2 from SquaredError.  This also required updating the test suite since it effectively doubles the gradient and loss. * Added doc for developers within RandomForest. * Small cleanup in test suite (generating data only once)
e57897a [Joseph K. Bradley] Fixed LogLoss for GradientBoostedTrees, and updated doc for losses, forests, and boosting

This PR reverts changes related to tag-based cluster membership. As discussed in SPARK-3332, we didn't figure out a safe strategy to use tags to determine cluster membership, because tagging is not atomic. The following changes are reverted:

SPARK-2333: 94053a7b
SPARK-3213: 7faf755a
SPARK-3608: 78d4220f.

I tested launch, login, and destroy. It is easy to check the diff by comparing it to Josh's patch for branch-1.1:

https://github.com/apache/spark/pull/2225/files

JoshRosen I sent the PR to master. It might be easier for us to keep master and branch-1.2 the same at this time. We can always re-apply the patch once we figure out a stable solution.

Author: Xiangrui Meng <meng@databricks.com>

Closes #3453 from mengxr/SPARK-4509 and squashes the following commits:

f0b708b [Xiangrui Meng] revert 94053a7b
4298ea5 [Xiangrui Meng] revert 7faf755a
35963a1 [Xiangrui Meng] Revert "SPARK-3608 Break if the instance tag naming succeeds"

Fix [SPARK-4471](https://issues.apache.org/jira/browse/SPARK-4471): blockManagerIdFromJson function throws exception while BlockManagerId be null in MetadataFetchFailedException

Author: hushan[胡珊] <hushan@xiaomi.com>

Closes #3340 from suyanNone/fix-blockmanagerId-jnothing-2 and squashes the following commits:

159f9a3 [hushan[胡珊]] Refine test code for blockmanager is null
4380d73 [hushan[胡珊]] remove useless blank line
3ccf651 [hushan[胡珊]] Fix SPARK-4471: blockManagerIdFromJson function throws exception while metadata fetch failed

The documentation points the user to run the following
```
sbin/start-history-server.sh
```
The first thing this does is throw an exception that complains a log directory is not specified. The exception message itself does not say anything about what to set. Instead we should have a default and a landing page with a better message. The new default log directory is `file:/tmp/spark-events`.

This is what it looks like as of this PR:

![after](https://issues.apache.org/jira/secure/attachment/12682985/after.png)

Author: Andrew Or <andrew@databricks.com>

Closes #3411 from andrewor14/minor-history-improvements and squashes the following commits:

f33d6b3 [Andrew Or] Point user to set config if default log dir does not exist
fc4c17a [Andrew Or] Improve HistoryServer UX

**Summary.** On failover, the Master may receive duplicate registrations from the same worker, causing the worker to exit. This is caused by this commit https://github.com/apache/spark/commit/4afe9a4852ebeb4cc77322a14225cd3dec165f3f, which adds logic for the worker to re-register with the master in case of failures. However, the following race condition may occur:

(1) Master A fails and Worker attempts to reconnect to all masters
(2) Master B takes over and notifies Worker
(3) Worker responds by registering with Master B
(4) Meanwhile, Worker's previous reconnection attempt reaches Master B, causing the same Worker to register with Master B twice

**Fix.** Instead of attempting to register with all known masters, the worker should re-register with only the one that it has been communicating with. This is safe because the fact that a failover has occurred means the old master must have died. Then, when the worker is finally notified of a new master, it gives up on the old one in favor of the new one.

**Caveat.** Even this fix is subject to more obscure race conditions. For instance, if Master B fails and Master A recovers immediately, then Master A may still observe duplicate worker registrations. However, this and other potential race conditions summarized in [SPARK-4592](https://issues.apache.org/jira/browse/SPARK-4592), are much, much less likely than the one described above, which is deterministically reproducible.

Author: Andrew Or <andrew@databricks.com>

Closes #3447 from andrewor14/standalone-failover and squashes the following commits:

0d9716c [Andrew Or] Move re-registration logic to actor for thread-safety
79286dc [Andrew Or] Preserve old behavior for initial retries
83b321c [Andrew Or] Tweak wording
1fce6a9 [Andrew Or] Active master actor could be null in the beginning
b6f269e [Andrew Or] Avoid duplicate worker registrations

[SPARK-4196][SPARK-4602][Streaming] Fix serialization issue in PairDStreamFunctions.saveAsNewAPIHadoopFiles

Solves two JIRAs in one shot
- Makes the ForechDStream created by saveAsNewAPIHadoopFiles serializable for checkpoints
- Makes the default configuration object used saveAsNewAPIHadoopFiles be the Spark's hadoop configuration

Author: Tathagata Das <tathagata.das1565@gmail.com>

Closes #3457 from tdas/savefiles-fix and squashes the following commits:

bb4729a [Tathagata Das] Same treatment for saveAsHadoopFiles
b382ea9 [Tathagata Das] Fix serialization issue in PairDStreamFunctions.saveAsNewAPIHadoopFiles.

The following optimizations are done to improve the StandardScaler model
transformation performance.

1) Covert Breeze dense vector to primitive vector to reduce the overhead.
2) Since mean can be potentially a sparse vector, we explicitly convert it to dense primitive vector.
3) Have a local reference to `shift` and `factor` array so JVM can locate the value with one operation call.
4) In pattern matching part, we use the mllib SparseVector/DenseVector instead of breeze's vector to
make the codebase cleaner.

Benchmark with mnist8m dataset:

Before,
DenseVector withMean and withStd: 50.97secs
DenseVector withMean and withoutStd: 42.11secs
DenseVector withoutMean and withStd: 8.75secs
SparseVector withoutMean and withStd: 5.437secs

With this PR,
DenseVector withMean and withStd: 5.76secs
DenseVector withMean and withoutStd: 5.28secs
DenseVector withoutMean and withStd: 5.30secs
SparseVector withoutMean and withStd: 1.27secs

Note that without the local reference copy of `factor` and `shift` arrays,
the runtime is almost three time slower.

DenseVector withMean and withStd: 18.15secs
DenseVector withMean and withoutStd: 18.05secs
DenseVector withoutMean and withStd: 18.54secs
SparseVector withoutMean and withStd: 2.01secs

The following code,
```scala
while (i < size) {
   values(i) = (values(i) - shift(i)) * factor(i)
   i += 1
}
```
will generate the bytecode
```
   L13
    LINENUMBER 106 L13
   FRAME FULL [org/apache/spark/mllib/feature/StandardScalerModel org/apache/spark/mllib/linalg/Vector org/apache/spark/mllib/linalg/Vector org/apache/spark/mllib/linalg/DenseVector T [D I I] []
    ILOAD 7
    ILOAD 6
    IF_ICMPGE L14
   L15
    LINENUMBER 107 L15
    ALOAD 5
    ILOAD 7
    ALOAD 5
    ILOAD 7
    DALOAD
    ALOAD 0
    INVOKESPECIAL org/apache/spark/mllib/feature/StandardScalerModel.shift ()[D
    ILOAD 7
    DALOAD
    DSUB
    ALOAD 0
    INVOKESPECIAL org/apache/spark/mllib/feature/StandardScalerModel.factor ()[D
    ILOAD 7
    DALOAD
    DMUL
    DASTORE
   L16
    LINENUMBER 108 L16
    ILOAD 7
    ICONST_1
    IADD
    ISTORE 7
    GOTO L13
```
, while with the local reference of the `shift` and `factor` arrays, the bytecode will be
```
   L14
    LINENUMBER 107 L14
    ALOAD 0
    INVOKESPECIAL org/apache/spark/mllib/feature/StandardScalerModel.factor ()[D
    ASTORE 9
   L15
    LINENUMBER 108 L15
   FRAME FULL [org/apache/spark/mllib/feature/StandardScalerModel org/apache/spark/mllib/linalg/Vector [D org/apache/spark/mllib/linalg/Vector org/apache/spark/mllib/linalg/DenseVector T [D I I [D] []
    ILOAD 8
    ILOAD 7
    IF_ICMPGE L16
   L17
    LINENUMBER 109 L17
    ALOAD 6
    ILOAD 8
    ALOAD 6
    ILOAD 8
    DALOAD
    ALOAD 2
    ILOAD 8
    DALOAD
    DSUB
    ALOAD 9
    ILOAD 8
    DALOAD
    DMUL
    DASTORE
   L18
    LINENUMBER 110 L18
    ILOAD 8
    ICONST_1
    IADD
    ISTORE 8
    GOTO L15
```

You can see that with local reference, the both of the arrays will be in the stack, so JVM can access the value without calling `INVOKESPECIAL`.

Author: DB Tsai <dbtsai@alpinenow.com>

Closes #3435 from dbtsai/standardscaler and squashes the following commits:

85885a9 [DB Tsai] revert to have lazy in shift array.
daf2b06 [DB Tsai] Address the feedback
cdb5cef [DB Tsai] small change
9c51eef [DB Tsai] style
fc795e4 [DB Tsai] update
5bffd3d [DB Tsai] first commit

[SPARK-4601][Streaming] Set correct call site for streaming jobs so that it is displayed correctly on the Spark UI

When running the NetworkWordCount, the description of the word count jobs are set as "getCallsite at DStream:xxx" . This should be set to the line number of the streaming application that has the output operation that led to the job being created. This is because the callsite is incorrectly set in the thread launching the jobs. This PR fixes that.

Author: Tathagata Das <tathagata.das1565@gmail.com>

Closes #3455 from tdas/streaming-callsite-fix and squashes the following commits:

69fc26f [Tathagata Das] Set correct call site for streaming jobs so that it is displayed correctly on the Spark UI

The documentation for the two parameters is the same with a pointer from the standalone parameter to the yarn parameter

Author: arahuja <aahuja11@gmail.com>

Closes #3209 from arahuja/yarn-classpath-first-param and squashes the following commits:

51cb9b2 [arahuja] [SPARK-4344][DOCS] adding documentation for YARN on userClassPathFirst

[SPARK-4381][Streaming]Add warning log when user set spark.master to local in Spark Streaming and there's no job executed

Author: jerryshao <saisai.shao@intel.com>

Closes #3244 from jerryshao/SPARK-4381 and squashes the following commits:

d2486c7 [jerryshao] Improve the warning log
d726e85 [jerryshao] Add local[1] to the filter condition
eca428b [jerryshao] Add warning log

change `NetworkInputDStream` to `ReceiverInputDStream`
change `ReceiverInputTracker` to `ReceiverTracker`

Author: q00251598 <qiyadong@huawei.com>

Closes #3400 from watermen/fix-comments and squashes the following commits:

75d795c [q00251598] change 'NetworkInputDStream' to 'ReceiverInputDStream' && change 'ReceiverInputTracker' to 'ReceiverTracker'

This is caused by the miniBatchSize parameter.The number of `RDD.sample` returns is not fixed.
cc mengxr

Author: GuoQiang Li <witgo@qq.com>

Closes #3399 from witgo/GradientDescent and squashes the following commits:

13cb228 [GuoQiang Li] review commit
668ab66 [GuoQiang Li] Double to Long
b6aa11a [GuoQiang Li] Check miniBatchSize is greater than 0
0b5c3e3 [GuoQiang Li] Minor fix
12e7424 [GuoQiang Li] GradientDescent get a wrong gradient value according to the gradient formula, which is caused by the miniBatchSize parameter.

In this refactoring, the performance will be slightly increased due to removing
the overhead from breeze vector. The bottleneck is still in breeze norm
which is implemented by activeIterator.

This inefficiency of breeze norm will be addressed in next PR. At least,
this PR makes the code more consistent in the codebase.

Author: DB Tsai <dbtsai@alpinenow.com>

Closes #3446 from dbtsai/normalizer and squashes the following commits:

e20a2b9 [DB Tsai] first commit

Author: wangfei <wangfei1@huawei.com>

Closes #3335 from scwf/patch-10 and squashes the following commits:

d343113 [wangfei] add '-Phive'
60d595e [wangfei] [DOC] Wrong cmd for build spark with apache hadoop 2.4.X and Hive 12 support

```timeTaken``` should not count the time of printing result.

Author: w00228970 <wangfei1@huawei.com>

Closes #3423 from scwf/time-taken-bug and squashes the following commits:

da7e102 [w00228970] compute time taken correctly

This is #3309 for the master branch.

e.g. clustering

Author: tkaessmann <tobias.kaessmanns24.com>

Closes #3309 from tkaessmann/branch-1.2 and squashes the following commits:

e3a3142 [tkaessmann] changes the comment for getVectors
58d3d83 [tkaessmann] removes sign from comment
a5be213 [tkaessmann] fixes getVectors to fit code guidelines
3782fa9 [tkaessmann] get raw vectors for further processing

Author: tkaessmann <tobias.kaessmann@s24.com>

Closes #3437 from mengxr/SPARK-4582 and squashes the following commits:

6c666b4 [tkaessmann] get raw vectors for further processing in Word2Vec

Functionally, this is just a small change on top of #3393 (by jongyoul). The issue being addressed is discussed in the comments there. I have not yet added a test for the bug there. I will add one shortly.

I've also done some minor renaming/clean-up of variables in this class and tests.

Author: Patrick Wendell <pwendell@gmail.com>
Author: Jongyoul Lee <jongyoul@gmail.com>

Closes #3436 from pwendell/mesos-issue and squashes the following commits:

58c35b5 [Patrick Wendell] Adding unit test for this situation
c4f0697 [Patrick Wendell] Additional clean-up and fixes on top of existing fix
f20f1b3 [Jongyoul Lee] [SPARK-4525] MesosSchedulerBackend.resourceOffers cannot decline unused offers from acceptedOffers - Added code for declining unused offers among acceptedOffers - Edited testCase for checking declining unused offers

This reverts commit b043c274.

I accidentally committed this using my own authorship credential. However,
I should have given authoriship to the original author: Jongyoul Lee.

Functionally, this is just a small change on top of #3393 (by jongyoul). The issue being addressed is discussed in the comments there. I have not yet added a test for the bug there. I will add one shortly.

I've also done some minor renaming/clean-up of variables in this class and tests.

Author: Patrick Wendell <pwendell@gmail.com>
Author: Jongyoul Lee <jongyoul@gmail.com>

Closes #3436 from pwendell/mesos-issue and squashes the following commits:

58c35b5 [Patrick Wendell] Adding unit test for this situation
c4f0697 [Patrick Wendell] Additional clean-up and fixes on top of existing fix
f20f1b3 [Jongyoul Lee] [SPARK-4525] MesosSchedulerBackend.resourceOffers cannot decline unused offers from acceptedOffers - Added code for declining unused offers among acceptedOffers - Edited testCase for checking declining unused offers

The commit changes the java script used to show/hide additional
metrics in order to reduce page load time. SPARK-4016 significantly
increased page load time for the stage page when stages had a lot
(thousands or tens of thousands) of tasks, due to the additional
Javascript to hide some metrics by default and stripe the tables.
This commit reduces page load time in two ways:

(1) Now, all of the metrics that are hidden by default are
hidden by setting "display: none;" using CSS for the page,
rather than hiding them using javascript after the page loads.
Without this change, for stages with thousands of tasks, there
was a few second delay after page load, where first the additional
metrics were shown, and then after a delay were hidden once the
relevant JS finished running.

(2) CSS is used to stripe all of the tables except for the summary
table. The summary table needs javascript to do the striping because
some rows are hidden, but the javascript striping is slower, which
again resulted in a delay when it was used for the task table (where
for a few seconds after page load, all of the rows in the task table
would be white, while the browser finished running the JS to stripe
the table).

cc pwendell

This change is intended to be backported to 1.2 to avoid a regression in
UI performance when users run large jobs.

Author: Kay Ousterhout <kayousterhout@gmail.com>

Closes #3328 from kayousterhout/SPARK-4266 and squashes the following commits:

f964091 [Kay Ousterhout] [SPARK-4266] [Web-UI] Reduce stage page load time.

Re-implement the Python broadcast using file:

1) serialize the python object using cPickle, write into disks.
2) Create a wrapper in JVM (for the dumped file), it read data from during serialization
3) Using TorrentBroadcast or HttpBroadcast to transfer the data (compressed) into executors
4) During deserialization, writing the data into disk.
5) Passing the path into Python worker, read data from disk and unpickle it into python object, until the first access.

It fixes the performance regression introduced in #2659, has similar performance as 1.1, but support object larger than 2G, also improve the memory efficiency (only one compressed copy in driver and executor).

Testing with a 500M broadcast and 4 tasks (excluding the benefit from reused worker in 1.2):

         name |   1.1   | 1.2 with this patch |  improvement
---------|--------|---------|--------
      python-broadcast-w-bytes  |	25.20  |	9.33   |	170.13% |
        python-broadcast-w-set	  |     4.13	   |    4.50  |	-8.35%  |

Testing with 100 tasks (16 CPUs):

         name |   1.1   | 1.2 with this patch |  improvement
---------|--------|---------|--------
     python-broadcast-w-bytes	| 38.16	| 8.40	 | 353.98%
        python-broadcast-w-set	| 23.29	| 9.59 |	142.80%

Author: Davies Liu <davies@databricks.com>

Closes #3417 from davies/pybroadcast and squashes the following commits:

50a58e0 [Davies Liu] address comments
b98de1d [Davies Liu] disable gc while unpickle
e5ee6b9 [Davies Liu] support large string
09303b8 [Davies Liu] read all data into memory
dde02dd [Davies Liu] improve performance of python broadcast

The Row object is created on the fly once the field is accessed, so we should access them by getattr() in asDict(0

Author: Davies Liu <davies@databricks.com>

Closes #3434 from davies/fix_asDict and squashes the following commits:

b20f1e7 [Davies Liu] fix asDict() with nested Row()