## What changes were proposed in this pull request?

Handle case where number of predictions is less than label set, k in nDCG computation

## How was this patch tested?

New unit test; existing tests

Author: Sean Owen <sowen@cloudera.com>

Closes #12756 from srowen/SPARK-14886.

## What changes were proposed in this pull request?
Minor typo fixes

## How was this patch tested?
local build

Author: Zheng RuiFeng <ruifengz@foxmail.com>

Closes #12755 from zhengruifeng/fix_doc_dataset.

## What changes were proposed in this pull request?
According to the [SPARK-14829](https://issues.apache.org/jira/browse/SPARK-14829), deprecate API of LogisticRegression and LinearRegression using SGD

## How was this patch tested?
manual tests

Author: Zheng RuiFeng <ruifengz@foxmail.com>

Closes #12596 from zhengruifeng/deprecate_sgd.

## What changes were proposed in this pull request?

This PR adds a new function in SparkR called `sparkLapply(list, function)`. This function implements a distributed version of `lapply` using Spark as a backend.

TODO:
 - [x] check documentation
 - [ ] check tests

Trivial example in SparkR:

```R
sparkLapply(1:5, function(x) { 2 * x })
```

Output:

```
[[1]]
[1] 2

[[2]]
[1] 4

[[3]]
[1] 6

[[4]]
[1] 8

[[5]]
[1] 10
```

Here is a slightly more complex example to perform distributed training of multiple models. Under the hood, Spark broadcasts the dataset.

```R
library("MASS")
data(menarche)
families <- c("gaussian", "poisson")
train <- function(family){glm(Menarche ~ Age  , family=family, data=menarche)}
results <- sparkLapply(families, train)
```

## How was this patch tested?

This PR was tested in SparkR. I am unfamiliar with R and SparkR, so any feedback on style, testing, etc. will be much appreciated.

cc falaki davies

Author: Timothy Hunter <timhunter@databricks.com>

Closes #12426 from thunterdb/7264.

## What changes were proposed in this pull request?
This patch removes HiveNativeCommand, so we can continue to remove the dependency on Hive. This pull request also removes the ability to generate golden result file using Hive.

## How was this patch tested?
Updated tests to reflect this.

Author: Reynold Xin <rxin@databricks.com>

Closes #12769 from rxin/SPARK-14991.

## What changes were proposed in this pull request?

`AccumulatorContext` is not thread-safe, that's why all of its methods are synchronized. However, there is one exception: the `AccumulatorContext.originals`. `NewAccumulator` use it to check if it's registered, which is wrong as it's not synchronized.

This PR mark `AccumulatorContext.originals` as `private` and now all access to `AccumulatorContext` is synchronized.

## How was this patch tested?

I verified it locally. To be safe, we can let jenkins test it many times to make sure this problem is gone.

Author: Wenchen Fan <wenchen@databricks.com>

Closes #12773 from cloud-fan/debug.

This reverts commit dae538a4.

## What changes were proposed in this pull request?

<copy form JIRA>

Currently if neither `spark.yarn.jars` nor `spark.yarn.archive` is set (by default), Spark on yarn code will upload all the jars in the folder separately into distributed cache, this is quite time consuming, and very verbose, instead of upload jars separately into distributed cache, here changes to zip all the jars first, and then put into distributed cache.

This will significantly improve the speed of starting time.

## How was this patch tested?

Unit test and local integrated test is done.

Verified with SparkPi both in spark cluster and client mode.

Author: jerryshao <sshao@hortonworks.com>

Closes #12597 from jerryshao/SPARK-14836.

## What changes were proposed in this pull request?

Updated Classifier, DecisionTreeClassifier, RandomForestClassifier, GBTClassifier to not require input column metadata.
* They first check for metadata.
* If numClasses is not specified in metadata, they identify the largest label value (up to a limit).

This functionality is implemented in a new Classifier.getNumClasses method.

Also
* Updated Classifier.extractLabeledPoints to (a) check label values and (b) include a second version which takes a numClasses value for validity checking.

## How was this patch tested?

* Unit tests in ClassifierSuite for helper methods
* Unit tests for DecisionTreeClassifier, RandomForestClassifier, GBTClassifier with toy datasets lacking label metadata

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

Closes #12663 from jkbradley/trees-no-metadata.

## What changes were proposed in this pull request?

This PR adds `since` tag into the matrix and vector classes in spark-mllib-local.

## How was this patch tested?

Scala-style checks passed.

Author: Pravin Gadakh <prgadakh@in.ibm.com>

Closes #12416 from pravingadakh/SPARK-14613.

## What changes were proposed in this pull request?

This PR adds Python APIs for:
 - `ContinuousQueryManager`
 - `ContinuousQueryException`

The `ContinuousQueryException` is a very basic wrapper, it doesn't provide the functionality that the Scala side provides, but it follows the same pattern for `AnalysisException`.

For `ContinuousQueryManager`, all APIs are provided except for registering listeners.

This PR also attempts to fix test flakiness by stopping all active streams just before tests.

## How was this patch tested?

Python Doc tests and unit tests

Author: Burak Yavuz <brkyvz@gmail.com>

Closes #12673 from brkyvz/pyspark-cqm.

## What changes were proposed in this pull request?
support avgMetrics in CrossValidatorModel with Python
## How was this patch tested?
Doctest and `test_save_load` in `pyspark/ml/test.py`
[JIRA](https://issues.apache.org/jira/browse/SPARK-12810)

Author: Kai Jiang <jiangkai@gmail.com>

Closes #12464 from vectorijk/spark-12810.

[SPARK-14970][SQL] Prevent DataSource from enumerates all files in a directory if there is user specified schema

## What changes were proposed in this pull request?
The FileCatalog object gets created even if the user specifies schema, which means files in the directory is enumerated even thought its not necessary. For large directories this is very slow. User would want to specify schema in such scenarios of large dirs, and this defeats the purpose quite a bit.

## How was this patch tested?
Hard to test this with unit test.

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

Closes #12748 from tdas/SPARK-14970.

## What changes were proposed in this pull request?

jira: https://issues.apache.org/jira/browse/SPARK-14916
FreqItemset as the result of FPGrowth should have a more friendly toString(), to help users and developers.
sample:
{a, b}: 5
{x, y, z}: 4

## How was this patch tested?

existing unit tests.

Author: Yuhao Yang <hhbyyh@gmail.com>

Closes #12698 from hhbyyh/freqtos.

## What changes were proposed in this pull request?

This splits GeneralizedLinearRegressionSummary into 2 summary types:
* GeneralizedLinearRegressionSummary, which does not store info from fitting (diagInvAtWA)
* GeneralizedLinearRegressionTrainingSummary, which is a subclass of GeneralizedLinearRegressionSummary and stores info from fitting

This also add a method evaluate() which can produce a GeneralizedLinearRegressionSummary on a new dataset.

The summary no longer provides the model itself as a public val.

Also:
* Fixes bug where GeneralizedLinearRegressionTrainingSummary was created with model, not summaryModel.
* Adds hasSummary method.
* Renames findSummaryModelAndPredictionCol -> getSummaryModel and simplifies that method.
* In summary, extract values from model immediately in case user later changes those (e.g., predictionCol).
* Pardon the style fixes; that is IntelliJ being obnoxious.

## How was this patch tested?

Existing unit tests + updated test for evaluate and hasSummary

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

Closes #12624 from jkbradley/model-summary-api.

## What changes were proposed in this pull request?

Fix to ScalaDoc for StructType.

## How was this patch tested?

Built locally.

Author: Gregory Hart <greg.hart@thinkbiganalytics.com>

Closes #12758 from freastro/hotfix/SPARK-14965.

## What changes were proposed in this pull request?

```
Welcome to
      ____              __
     / __/__  ___ _____/ /__
    _\ \/ _ \/ _ `/ __/  '_/
   /__ / .__/\_,_/_/ /_/\_\   version 2.0.0-SNAPSHOT
      /_/

Using Python version 2.7.5 (default, Mar  9 2014 22:15:05)
SparkSession available as 'spark'.
>>> spark
<pyspark.sql.session.SparkSession object at 0x101f3bfd0>
>>> spark.sql("SHOW TABLES").show()
...
+---------+-----------+
|tableName|isTemporary|
+---------+-----------+
|      src|      false|
+---------+-----------+

>>> spark.range(1, 10, 2).show()
+---+
| id|
+---+
|  1|
|  3|
|  5|
|  7|
|  9|
+---+
```
**Note**: This API is NOT complete in its current state. In particular, for now I left out the `conf` and `catalog` APIs, which were added later in Scala. These will be added later before 2.0.

## How was this patch tested?

Python tests.

Author: Andrew Or <andrew@databricks.com>

Closes #12746 from andrewor14/python-spark-session.

https://issues.apache.org/jira/browse/SPARK-14935

In DistributedSuite, the "local-cluster format" test actually launches a bunch of clusters, but this doesn't seem necessary for what should just be a unit test of a regex. We should clean up the code so that this is testable without actually launching a cluster, which should buy us about 20 seconds per build.

Passed unit test on my local machine

Author: Xin Ren <iamshrek@126.com>

Closes #12744 from keypointt/SPARK-14935.

## What changes were proposed in this pull request?

Add simple clarification that Spark can be cross-built for other Scala versions.

## How was this patch tested?

Automated doc build

Author: Sean Owen <sowen@cloudera.com>

Closes #12757 from srowen/SPARK-14882.

This work is based on twinkle-sachdeva 's proposal. In parallel to such mechanism for AM failures, here add similar mechanism for executor failure tracking, this is useful for long running Spark service to mitigate the executor failure problems.

Please help to review, tgravescs sryza and vanzin

Author: jerryshao <sshao@hortonworks.com>

Closes #10241 from jerryshao/SPARK-6735.

Make the behavior of mutate more consistent with that in dplyr, besides support for replacing existing columns.
1. Throw error message when there are duplicated column names in the DataFrame being mutated.
2. when there are duplicated column names in specified columns by arguments, the last column of the same name takes effect.

Author: Sun Rui <rui.sun@intel.com>

Closes #10220 from sun-rui/SPARK-12235.

## What changes were proposed in this pull request?
This is a proposal to print the Spark Driver UI link when spark-shell is launched.

## How was this patch tested?
Launched spark-shell in local mode and cluster mode. Spark-shell console output included following line:
"Spark context Web UI available at <Spark web url>"

Author: Ergin Seyfe <eseyfe@fb.com>

Closes #12341 from seyfe/spark_console_display_webui_link.

## What changes were proposed in this pull request?

Currently we use `SQLUserDefinedType` annotation to register UDTs for user classes. However, by doing this, we add Spark dependency to user classes.

For some user classes, it is unnecessary to add such dependency that will increase deployment difficulty.

We should provide alternative approach to register UDTs for user classes without `SQLUserDefinedType` annotation.

## How was this patch tested?

`UserDefinedTypeSuite`

Author: Liang-Chi Hsieh <simonh@tw.ibm.com>

Closes #12259 from viirya/improve-sql-usertype.

## What changes were proposed in this pull request?

This PR introduces a new accumulator API  which is much simpler than before:

1. the type hierarchy is simplified, now we only have an `Accumulator` class
2. Combine `initialValue` and `zeroValue` concepts into just one concept: `zeroValue`
3. there in only one `register` method, the accumulator registration and cleanup registration are combined.
4. the `id`,`name` and `countFailedValues` are combined into an `AccumulatorMetadata`, and is provided during registration.

`SQLMetric` is a good example to show the simplicity of this new API.

What we break:

1. no `setValue` anymore. In the new API, the intermedia type can be different from the result type, it's very hard to implement a general `setValue`
2. accumulator can't be serialized before registered.

Problems need to be addressed in follow-ups:

1. with this new API, `AccumulatorInfo` doesn't make a lot of sense, the partial output is not partial updates, we need to expose the intermediate value.
2. `ExceptionFailure` should not carry the accumulator updates. Why do users care about accumulator updates for failed cases? It looks like we only use this feature to update the internal metrics, how about we sending a heartbeat to update internal metrics after the failure event?
3. the public event `SparkListenerTaskEnd` carries a `TaskMetrics`. Ideally this `TaskMetrics` don't need to carry external accumulators, as the only method of `TaskMetrics` that can access external accumulators is `private[spark]`. However, `SQLListener` use it to retrieve sql metrics.

## How was this patch tested?

existing tests

Author: Wenchen Fan <wenchen@databricks.com>

Closes #12612 from cloud-fan/acc.

## What changes were proposed in this pull request?
The current signal handlers have a subtle bug that stops evaluating registered actions as soon as one of them returns true, this is because `forall` is short-circuited.
This PR adds a strict mapping stage before evaluating returned result.
There are no known occurrences of the bug and this is a preemptive fix.

## How was this patch tested?
As with the original introduction of signal handlers, this was tested manually (unit testing with signals is not straightforward).

Author: Jakob Odersky <jakob@odersky.com>

Closes #12745 from jodersky/SPARK-10001-hotfix.

## What changes were proposed in this pull request?

Currently, LongToUnsafeRowMap use byte array as the underlying page, which can't be larger 1G.

This PR improves LongToUnsafeRowMap  to scale up to 8G bytes by using array of Long instead of array of byte.

## How was this patch tested?

Manually ran a test to confirm that both UnsafeHashedRelation and LongHashedRelation could build a map that larger than 2G.

Author: Davies Liu <davies@databricks.com>

Closes #12740 from davies/larger_broadcast.

## What changes were proposed in this pull request?

https://issues.apache.org/jira/browse/SPARK-12143

This PR adds the support for conversion between `SparkRow` in Spark and `RowSet` in Hive for `BinaryType` as `Array[Byte]` (JDBC)
## How was this patch tested?

Unittests in `HiveThriftBinaryServerSuite` (regression test)

Closes #10139

Author: hyukjinkwon <gurwls223@gmail.com>

Closes #12733 from HyukjinKwon/SPARK-12143.

## What changes were proposed in this pull request?

The contents of `.gitignore` have been sorted to make it more readable. The actual contents of the file have not been changed.

## How was this patch tested?

Does not require any tests.

Author: Arun Allamsetty <arun@instructure.com>

Closes #12742 from aa8y/gitignore.

In local profiling, I noticed SizeEstimator spending tons of time estimating the size of objects which contain TypeTag or ClassTag fields. The problem with these tags is that they reference global Scala reflection objects, which, in turn, reference many singletons, such as TestHive. This throws off the accuracy of the size estimation and wastes tons of time traversing a huge object graph.

As a result, I think that SizeEstimator should ignore any classes in the `scala.reflect` package.

Author: Josh Rosen <joshrosen@databricks.com>

Closes #12741 from JoshRosen/ignore-scala-reflect-in-size-estimator.

## What changes were proposed in this pull request?

Pipeline.setStages failed for some code examples which worked in 1.5 but fail in 1.6.  This tends to occur when using a mix of transformers from ml.feature. It is because Java Arrays are non-covariant and the addition of MLWritable to some transformers means the stages0/1 arrays above are not of type Array[PipelineStage].  This PR modifies the following to accept subclasses of PipelineStage:
* Pipeline.setStages()
* Params.w()

## How was this patch tested?

Unit test which fails to compile before this fix.

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

Closes #12430 from jkbradley/pipeline-setstages.

This reverts commit 3f49afee.

Added parameter drop to subsetting operator [. This is useful to get a Column from a DataFrame, given its name. R supports it.

In R:
```
> name <- "Sepal_Length"
> class(iris[, name])
[1] "numeric"
```
Currently, in SparkR:
```
> name <- "Sepal_Length"
> class(irisDF[, name])
[1] "DataFrame"
```

Previous code returns a DataFrame, which is inconsistent with R's behavior. SparkR should return a Column instead. Currently, in order for the user to return a Column given a column name as a character variable would be through `eval(parse(x))`, where x is the string `"irisDF$Sepal_Length"`. That itself is pretty hacky. `SparkR:::getColumn() `is another choice, but I don't see why this method should be externalized. Instead, following R's way to do things, the proposed implementation allows this:

```
> name <- "Sepal_Length"
> class(irisDF[, name, drop=T])
[1] "Column"

> class(irisDF[, name, drop=F])
[1] "DataFrame"
```

This is consistent with R:

```
> name <- "Sepal_Length"
> class(iris[, name])
[1] "numeric"
> class(iris[, name, drop=F])
[1] "data.frame"
```

Author: Oscar D. Lara Yejas <odlaraye@oscars-mbp.usca.ibm.com>
Author: Oscar D. Lara Yejas <odlaraye@oscars-mbp.attlocal.net>

Closes #11318 from olarayej/SPARK-13436.

## What changes were proposed in this pull request?

`interfaces.scala` was getting big. This just moves the biggest class in there to a new file for cleanliness.

## How was this patch tested?

Just moving things around.

Author: Andrew Or <andrew@databricks.com>

Closes #12721 from andrewor14/move-external-catalog.

## What changes were proposed in this pull request?
Since [SPARK-10574](https://issues.apache.org/jira/browse/SPARK-10574) breaks behavior of ```HashingTF```, we should try to enforce good practice by removing the "native" hashAlgorithm option in spark.ml and pyspark.ml. We can leave spark.mllib and pyspark.mllib alone.

## How was this patch tested?
Unit tests.

cc jkbradley

Author: Yanbo Liang <ybliang8@gmail.com>

Closes #12702 from yanboliang/spark-14899.

Currently, we can only create persisted partitioned and/or bucketed data source tables using the Dataset API but not using SQL DDL. This PR implements the following syntax to add partitioning and bucketing support to the SQL DDL:

```
CREATE TABLE <table-name>
USING <provider> [OPTIONS (<key1> <value1>, <key2> <value2>, ...)]
[PARTITIONED BY (col1, col2, ...)]
[CLUSTERED BY (col1, col2, ...) [SORTED BY (col1, col2, ...)] INTO <n> BUCKETS]
AS SELECT ...
```

Test cases are added in `MetastoreDataSourcesSuite` to check the newly added syntax.

Author: Cheng Lian <lian@databricks.com>
Author: Yin Huai <yhuai@databricks.com>

Closes #12734 from liancheng/spark-14954.

## What changes were proposed in this pull request?

Currently, `build/mvn` provides a convenient option, `--force`, in order to use the recommended version of maven without changing PATH environment variable. However, there were two problems.

- `dev/lint-java` does not use the newly installed maven.

  ```bash
$ ./build/mvn --force clean
$ ./dev/lint-java
Using `mvn` from path: /usr/local/bin/mvn
```
- It's not easy to type `--force` option always.

If '--force' option is used once, we had better prefer the installed maven recommended by Spark.
This PR makes `build/mvn` check the existence of maven installed by `--force` option first.

According to the comments, this PR aims to the followings:
- Detect the maven version from `pom.xml`.
- Install maven if there is no or old maven.
- Remove `--force` option.

## How was this patch tested?

Manual.

```bash
$ ./build/mvn --force clean
$ ./dev/lint-java
Using `mvn` from path: /Users/dongjoon/spark/build/apache-maven-3.3.9/bin/mvn
...
$ rm -rf ./build/apache-maven-3.3.9/
$ ./dev/lint-java
Using `mvn` from path: /usr/local/bin/mvn
```

Author: Dongjoon Hyun <dongjoon@apache.org>

Closes #12631 from dongjoon-hyun/SPARK-14867.

## What changes were proposed in this pull request?

This PR aims to implement decimal aggregation optimization for window queries by improving existing `DecimalAggregates`. Historically, `DecimalAggregates` optimizer is designed to transform general `sum/avg(decimal)`, but it breaks recently added windows queries like the followings. The following queries work well without the current `DecimalAggregates` optimizer.

**Sum**
```scala
scala> sql("select sum(a) over () from (select explode(array(1.0,2.0)) a) t").head
java.lang.RuntimeException: Unsupported window function: MakeDecimal((sum(UnscaledValue(a#31)),mode=Complete,isDistinct=false),12,1)
scala> sql("select sum(a) over () from (select explode(array(1.0,2.0)) a) t").explain()
== Physical Plan ==
WholeStageCodegen
:  +- Project [sum(a) OVER (  ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING)#23]
:     +- INPUT
+- Window [MakeDecimal((sum(UnscaledValue(a#21)),mode=Complete,isDistinct=false),12,1) windowspecdefinition(ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING) AS sum(a) OVER (  ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING)#23]
   +- Exchange SinglePartition, None
      +- Generate explode([1.0,2.0]), false, false, [a#21]
         +- Scan OneRowRelation[]
```

**Average**
```scala
scala> sql("select avg(a) over () from (select explode(array(1.0,2.0)) a) t").head
java.lang.RuntimeException: Unsupported window function: cast(((avg(UnscaledValue(a#40)),mode=Complete,isDistinct=false) / 10.0) as decimal(6,5))
scala> sql("select avg(a) over () from (select explode(array(1.0,2.0)) a) t").explain()
== Physical Plan ==
WholeStageCodegen
:  +- Project [avg(a) OVER (  ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING)#44]
:     +- INPUT
+- Window [cast(((avg(UnscaledValue(a#42)),mode=Complete,isDistinct=false) / 10.0) as decimal(6,5)) windowspecdefinition(ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING) AS avg(a) OVER (  ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING)#44]
   +- Exchange SinglePartition, None
      +- Generate explode([1.0,2.0]), false, false, [a#42]
         +- Scan OneRowRelation[]
```

After this PR, those queries work fine and new optimized physical plans look like the followings.

**Sum**
```scala
scala> sql("select sum(a) over () from (select explode(array(1.0,2.0)) a) t").explain()
== Physical Plan ==
WholeStageCodegen
:  +- Project [sum(a) OVER (  ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING)#35]
:     +- INPUT
+- Window [MakeDecimal((sum(UnscaledValue(a#33)),mode=Complete,isDistinct=false) windowspecdefinition(ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING),12,1) AS sum(a) OVER (  ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING)#35]
   +- Exchange SinglePartition, None
      +- Generate explode([1.0,2.0]), false, false, [a#33]
         +- Scan OneRowRelation[]
```

**Average**
```scala
scala> sql("select avg(a) over () from (select explode(array(1.0,2.0)) a) t").explain()
== Physical Plan ==
WholeStageCodegen
:  +- Project [avg(a) OVER (  ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING)#47]
:     +- INPUT
+- Window [cast(((avg(UnscaledValue(a#45)),mode=Complete,isDistinct=false) windowspecdefinition(ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING) / 10.0) as decimal(6,5)) AS avg(a) OVER (  ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING)#47]
   +- Exchange SinglePartition, None
      +- Generate explode([1.0,2.0]), false, false, [a#45]
         +- Scan OneRowRelation[]
```

In this PR, *SUM over window* pattern matching is based on the code of hvanhovell ; he should be credited for the work he did.

## How was this patch tested?

Pass the Jenkins tests (with newly added testcases)

Author: Dongjoon Hyun <dongjoon@apache.org>

Closes #12421 from dongjoon-hyun/SPARK-14664.

[SPARK-14937][ML][DOCUMENT] spark.ml LogisticRegression sqlCtx in scala is inconsistent with java and python

## What changes were proposed in this pull request?
In spark.ml document, the LogisticRegression scala example uses sqlCtx. It is inconsistent with java and python examples which use sqlContext. In addition, a user can't copy & paste to run the example in spark-shell as sqlCtx doesn't exist in spark-shell while sqlContext exists.

Change the scala example referred by the spark.ml example.

## How was this patch tested?

Compile the example scala file and it passes compilation.

Author: wm624@hotmail.com <wm624@hotmail.com>

Closes #12717 from wangmiao1981/doc.

CheckpointWriter.stop() is prone to a race condition: if one thread calls `stop()` right as a checkpoint write task begins to execute, that write task may become blocked when trying to access `fs`, the shared Hadoop FileSystem, since both the `fs` getter and `stop` method synchronize on the same lock. Here's a thread-dump excerpt which illustrates the problem:

```java
"pool-31-thread-1" #156 prio=5 os_prio=31 tid=0x00007fea02cd2000 nid=0x5c0b waiting for monitor entry [0x000000013bc4c000]
   java.lang.Thread.State: BLOCKED (on object monitor)
    at org.apache.spark.streaming.CheckpointWriter.org$apache$spark$streaming$CheckpointWriter$$fs(Checkpoint.scala:302)
    - waiting to lock <0x00000007bf53ee78> (a org.apache.spark.streaming.CheckpointWriter)
    at org.apache.spark.streaming.CheckpointWriter$CheckpointWriteHandler.run(Checkpoint.scala:224)
    at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
    at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
    at java.lang.Thread.run(Thread.java:745)

"pool-1-thread-1-ScalaTest-running-MapWithStateSuite" #11 prio=5 os_prio=31 tid=0x00007fe9ff879800 nid=0x5703 waiting on condition [0x000000012e54c000]
   java.lang.Thread.State: TIMED_WAITING (parking)
    at sun.misc.Unsafe.park(Native Method)
    - parking to wait for  <0x00000007bf564568> (a java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject)
    at java.util.concurrent.locks.LockSupport.parkNanos(LockSupport.java:215)
    at java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.awaitNanos(AbstractQueuedSynchronizer.java:2078)
    at java.util.concurrent.ThreadPoolExecutor.awaitTermination(ThreadPoolExecutor.java:1465)
    at org.apache.spark.streaming.CheckpointWriter.stop(Checkpoint.scala:291)
    - locked <0x00000007bf53ee78> (a org.apache.spark.streaming.CheckpointWriter)
    at org.apache.spark.streaming.scheduler.JobGenerator.stop(JobGenerator.scala:159)
    - locked <0x00000007bf53ea90> (a org.apache.spark.streaming.scheduler.JobGenerator)
    at org.apache.spark.streaming.scheduler.JobScheduler.stop(JobScheduler.scala:115)
    - locked <0x00000007bf53d3f0> (a org.apache.spark.streaming.scheduler.JobScheduler)
    at org.apache.spark.streaming.StreamingContext$$anonfun$stop$1.apply$mcV$sp(StreamingContext.scala:680)
    at org.apache.spark.util.Utils$.tryLogNonFatalError(Utils.scala:1219)
    at org.apache.spark.streaming.StreamingContext.stop(StreamingContext.scala:679)
    - locked <0x00000007bf516a70> (a org.apache.spark.streaming.StreamingContext)
    at org.apache.spark.streaming.StreamingContext.stop(StreamingContext.scala:644)
    - locked <0x00000007bf516a70> (a org.apache.spark.streaming.StreamingContext)
[...]
```

We can fix this problem by having `stop` and `fs` be synchronized on different locks: the synchronization on `stop` only needs to guard against multiple threads calling `stop` at the same time, whereas the synchronization on `fs` is only necessary for cross-thread visibility. There's only ever a single active checkpoint writer thread at a time, so we don't need to guard against concurrent access to `fs`. Thus, `fs` can simply become a `volatile` var, similar to `lastCheckpointTime`.

This change should fix [SPARK-13693](https://issues.apache.org/jira/browse/SPARK-13693), a flaky `MapWithStateSuite` test suite which has recently been failing several times per day. It also results in a huge test speedup: prior to this patch, `MapWithStateSuite` took about 80 seconds to run, whereas it now runs in less than 10 seconds. For the `streaming` project's tests as a whole, they now run in ~220 seconds vs. ~354 before.

/cc zsxwing and tdas for review.

Author: Josh Rosen <joshrosen@databricks.com>

Closes #12712 from JoshRosen/fix-checkpoint-writer-race.