## What changes were proposed in this pull request?
This is a follow-up for https://github.com/apache/spark/pull/18488, to simplify the code.
The major change is, we should map java enum to string type, instead of a struct type with a single string field.
## How was this patch tested?
existing tests
Author: Wenchen Fan <wenchen@databricks.com>
Closes#19066 from cloud-fan/fix.
## What changes were proposed in this pull request?
Add trait UserDefinedExpression to identify user-defined functions.
UDF can be expensive. In optimizer we may need to avoid executing UDF multiple times.
E.g.
```scala
table.select(UDF as 'a).select('a, ('a + 1) as 'b)
```
If UDF is expensive in this case, optimizer should not collapse the project to
```scala
table.select(UDF as 'a, (UDF+1) as 'b)
```
Currently UDF classes like PythonUDF, HiveGenericUDF are not defined in catalyst.
This PR is to add a new trait to make it easier to identify user-defined functions.
## How was this patch tested?
Unit test
Author: Wang Gengliang <ltnwgl@gmail.com>
Closes#19064 from gengliangwang/UDFType.
## What changes were proposed in this pull request?
As mentioned at https://github.com/apache/spark/pull/18680#issuecomment-316820409, when we have more `ColumnVector` implementations, it might (or might not) have huge performance implications because it might disable inlining, or force virtual dispatches.
As for read path, one of the major paths is the one generated by `ColumnBatchScan`. Currently it refers `ColumnVector` so the penalty will be bigger as we have more classes, but we can know the concrete type from its usage, e.g. vectorized Parquet reader uses `OnHeapColumnVector`. We can use the concrete type in the generated code directly to avoid the penalty.
## How was this patch tested?
Existing tests.
Author: Takuya UESHIN <ueshin@databricks.com>
Closes#18989 from ueshin/issues/SPARK-21781.
## What changes were proposed in this pull request?
Add 4 traits, using the following hierarchy:
LogisticRegressionSummary
LogisticRegressionTrainingSummary: LogisticRegressionSummary
BinaryLogisticRegressionSummary: LogisticRegressionSummary
BinaryLogisticRegressionTrainingSummary: LogisticRegressionTrainingSummary, BinaryLogisticRegressionSummary
and the public method such as `def summary` only return trait type listed above.
and then implement 4 concrete classes:
LogisticRegressionSummaryImpl (multiclass case)
LogisticRegressionTrainingSummaryImpl (multiclass case)
BinaryLogisticRegressionSummaryImpl (binary case).
BinaryLogisticRegressionTrainingSummaryImpl (binary case).
## How was this patch tested?
Existing tests & added tests.
Author: WeichenXu <WeichenXu123@outlook.com>
Closes#15435 from WeichenXu123/mlor_summary.
## What changes were proposed in this pull request?
Fair Scheduler can be built via one of the following options:
- By setting a `spark.scheduler.allocation.file` property,
- By setting `fairscheduler.xml` into classpath.
These options are checked **in order** and fair-scheduler is built via first found option. If invalid path is found, `FileNotFoundException` will be expected.
This PR aims unit test coverage of these use cases and a minor documentation change has been added for second option(`fairscheduler.xml` into classpath) to inform the users.
Also, this PR was related with #16813 and has been created separately to keep patch content as isolated and to help the reviewers.
## How was this patch tested?
Added new Unit Tests.
Author: erenavsarogullari <erenavsarogullari@gmail.com>
Closes#16992 from erenavsarogullari/SPARK-19662.
History Server Launch uses SparkClassCommandBuilder for launching the server. It is observed that SPARK_CLASSPATH has been removed and deprecated. For spark-submit this takes a different route and spark.driver.extraClasspath takes care of specifying additional jars in the classpath that were previously specified in the SPARK_CLASSPATH. Right now the only way specify the additional jars for launching daemons such as history server is using SPARK_DIST_CLASSPATH (https://spark.apache.org/docs/latest/hadoop-provided.html) but this I presume is a distribution classpath. It would be nice to have a similar config like spark.driver.extraClasspath for launching daemons similar to history server.
Added new environment variable SPARK_DAEMON_CLASSPATH to set classpath for launching daemons. Tested and verified for History Server and Standalone Mode.
## How was this patch tested?
Initially, history server start script would fail for the reason being that it could not find the required jars for launching the server in the java classpath. Same was true for running Master and Worker in standalone mode. By adding the environment variable SPARK_DAEMON_CLASSPATH to the java classpath, both the daemons(History Server, Standalone daemons) are starting up and running.
Author: pgandhi <pgandhi@yahoo-inc.com>
Author: pgandhi999 <parthkgandhi9@gmail.com>
Closes#19047 from pgandhi999/master.
## What changes were proposed in this pull request?
Because of numerical error, MultivariateOnlineSummarizer.variance is possible to generate negative variance.
**This is a serious bug because many algos in MLLib**
**use stddev computed from** `sqrt(variance)`
**it will generate NaN and crash the whole algorithm.**
we can reproduce this bug use the following code:
```
val summarizer1 = (new MultivariateOnlineSummarizer)
.add(Vectors.dense(3.0), 0.7)
val summarizer2 = (new MultivariateOnlineSummarizer)
.add(Vectors.dense(3.0), 0.4)
val summarizer3 = (new MultivariateOnlineSummarizer)
.add(Vectors.dense(3.0), 0.5)
val summarizer4 = (new MultivariateOnlineSummarizer)
.add(Vectors.dense(3.0), 0.4)
val summarizer = summarizer1
.merge(summarizer2)
.merge(summarizer3)
.merge(summarizer4)
println(summarizer.variance(0))
```
This PR fix the bugs in `mllib.stat.MultivariateOnlineSummarizer.variance` and `ml.stat.SummarizerBuffer.variance`, and several places in `WeightedLeastSquares`
## How was this patch tested?
test cases added.
Author: WeichenXu <WeichenXu123@outlook.com>
Closes#19029 from WeichenXu123/fix_summarizer_var_bug.
Signed-off-by: iamhumanbeing <iamhumanbeinggmail.com>
## What changes were proposed in this pull request?
testNameNote = "(minNumPostShufflePartitions: 3) is not correct.
it should be "(minNumPostShufflePartitions: " + numPartitions + ")" in ExchangeCoordinatorSuite
## How was this patch tested?
unit tests
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: iamhumanbeing <iamhumanbeing@gmail.com>
Closes#19058 from iamhumanbeing/testnote.
## What changes were proposed in this pull request?
This PR proposes both:
- Add information about Javadoc, SQL docs and few more information in `docs/README.md` and a comment in `docs/_plugins/copy_api_dirs.rb` related with Javadoc.
- Adds some commands so that the script always runs the SQL docs build under `./sql` directory (for directly running `./sql/create-docs.sh` in the root directory).
## How was this patch tested?
Manual tests with `jekyll build` and `./sql/create-docs.sh` in the root directory.
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#19019 from HyukjinKwon/minor-doc-build.
## What changes were proposed in this pull request?
[SPARK-19025](https://github.com/apache/spark/pull/16869) removes SQLBuilder, so we don't need the following in HiveCompatibilitySuite.
```scala
// Ensures that the plans generation use metastore relation and not OrcRelation
// Was done because SqlBuilder does not work with plans having logical relation
TestHive.setConf(HiveUtils.CONVERT_METASTORE_ORC, false)
```
## How was this patch tested?
Pass the existing Jenkins tests.
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#19043 from dongjoon-hyun/SPARK-21831.
## What changes were proposed in this pull request?
Adjust Local UDTs test to assert about results, and fix index of vector column. See JIRA for details.
## How was this patch tested?
Existing tests.
Author: Sean Owen <sowen@cloudera.com>
Closes#19053 from srowen/SPARK-21837.
## What changes were proposed in this pull request?
This patch adds allowUnquotedControlChars option in JSON data source to allow JSON Strings to contain unquoted control characters (ASCII characters with value less than 32, including tab and line feed characters)
## How was this patch tested?
Add new test cases
Author: vinodkc <vinod.kc.in@gmail.com>
Closes#19008 from vinodkc/br_fix_SPARK-21756.
This is a follow up to cba826d0; that commit set the app handle state
to "LOST" when the child process exited, but that can be ambiguous. This
change sets the state to "FAILED" if the exit code was non-zero and
the handle state wasn't a failure state, or "LOST" if the exit status
was zero.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#19012 from vanzin/SPARK-17742.
## What changes were proposed in this pull request?
With SPARK-10643, Spark supports download resources from remote in client deploy mode. But the implementation overrides variables which representing added resources (like `args.jars`, `args.pyFiles`) to local path, And yarn client leverage this local path to re-upload resources to distributed cache. This is unnecessary to break the semantics of putting resources in a shared FS. So here proposed to fix it.
## How was this patch tested?
This is manually verified with jars, pyFiles in local and remote storage, both in client and cluster mode.
Author: jerryshao <sshao@hortonworks.com>
Closes#18962 from jerryshao/SPARK-21714.
## What changes were proposed in this pull request?
After [SPARK-19025](https://github.com/apache/spark/pull/16869), there is no need to keep SQLBuilderTest.
ExpressionSQLBuilderSuite is the only place to use it.
This PR aims to remove SQLBuilderTest.
## How was this patch tested?
Pass the updated `ExpressionSQLBuilderSuite`.
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#19044 from dongjoon-hyun/SPARK-21832.
## What changes were proposed in this pull request?
Fix build warnings and Java lint errors. This just helps a bit in evaluating (new) warnings in another PR I have open.
## How was this patch tested?
Existing tests
Author: Sean Owen <sowen@cloudera.com>
Closes#19051 from srowen/JavaWarnings.
## What changes were proposed in this pull request?
Right now, ChunkedByteBuffer#writeFully do not slice bytes first.We observe code in java nio Util#getTemporaryDirectBuffer below:
BufferCache cache = bufferCache.get();
ByteBuffer buf = cache.get(size);
if (buf != null) {
return buf;
} else {
// No suitable buffer in the cache so we need to allocate a new
// one. To avoid the cache growing then we remove the first
// buffer from the cache and free it.
if (!cache.isEmpty()) {
buf = cache.removeFirst();
free(buf);
}
return ByteBuffer.allocateDirect(size);
}
If we slice first with a fixed size, we can use buffer cache and only need to allocate at the first write call.
Since we allocate new buffer, we can not control the free time of this buffer.This once cause memory issue in our production cluster.
In this patch, i supply a new api which will slice with fixed size for buffer writing.
## How was this patch tested?
Unit test and test in production.
Author: zhoukang <zhoukang199191@gmail.com>
Author: zhoukang <zhoukang@xiaomi.com>
Closes#18730 from caneGuy/zhoukang/improve-chunkwrite.
## What changes were proposed in this pull request?
Fixed NPE when creating encoder for enum.
When you try to create an encoder for Enum type (or bean with enum property) via Encoders.bean(...), it fails with NullPointerException at TypeToken:495.
I did a little research and it turns out, that in JavaTypeInference following code
```
def getJavaBeanReadableProperties(beanClass: Class[_]): Array[PropertyDescriptor] = {
val beanInfo = Introspector.getBeanInfo(beanClass)
beanInfo.getPropertyDescriptors.filterNot(_.getName == "class")
.filter(_.getReadMethod != null)
}
```
filters out properties named "class", because we wouldn't want to serialize that. But enum types have another property of type Class named "declaringClass", which we are trying to inspect recursively. Eventually we try to inspect ClassLoader class, which has property "defaultAssertionStatus" with no read method, which leads to NPE at TypeToken:495.
I added property name "declaringClass" to filtering to resolve this.
## How was this patch tested?
Unit test in JavaDatasetSuite which creates an encoder for enum
Author: mike <mike0sv@gmail.com>
Author: Mikhail Sveshnikov <mike0sv@gmail.com>
Closes#18488 from mike0sv/enum-support.
## What changes were proposed in this pull request?
convert LinearSVC to new aggregator framework
## How was this patch tested?
existing unit test.
Author: Yuhao Yang <yuhao.yang@intel.com>
Closes#18315 from hhbyyh/svcAggregator.
## What changes were proposed in this pull request?
This PR bumps the ANTLR version to 4.7, and fixes a number of small parser related issues uncovered by the bump.
The main reason for upgrading is that in some cases the current version of ANTLR (4.5) can exhibit exponential slowdowns if it needs to parse boolean predicates. For example the following query will take forever to parse:
```sql
SELECT *
FROM RANGE(1000)
WHERE
TRUE
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
AND NOT upper(DESCRIPTION) LIKE '%FOO%'
```
This is caused by a know bug in ANTLR (https://github.com/antlr/antlr4/issues/994), which was fixed in version 4.6.
## How was this patch tested?
Existing tests.
Author: Herman van Hovell <hvanhovell@databricks.com>
Closes#19042 from hvanhovell/SPARK-21830.
## What changes were proposed in this pull request?
TCP parameters like SO_RCVBUF and SO_SNDBUF can be set in SparkConf, and `org.apache.spark.network.server.TransportServe`r can use those parameters to build server by leveraging netty. But for TransportClientFactory, there is no such way to set those parameters from SparkConf. This could be inconsistent in server and client side when people set parameters in SparkConf. So this PR make RPC client to be enable to use those TCP parameters as well.
## How was this patch tested?
Existing tests.
Author: xu.zhang <xu.zhang@hulu.com>
Closes#18964 from neoremind/add_client_param.
## What changes were proposed in this pull request?
Add more cases we should view as a normal query stop rather than a failure.
## How was this patch tested?
The new unit tests.
Author: Shixiong Zhu <zsxwing@gmail.com>
Closes#18997 from zsxwing/SPARK-21788.
## What changes were proposed in this pull request?
With the check for structural integrity proposed in SPARK-21726, it is found that the optimization rule `PullupCorrelatedPredicates` can produce unresolved plans.
For a correlated IN query looks like:
SELECT t1.a FROM t1
WHERE
t1.a IN (SELECT t2.c
FROM t2
WHERE t1.b < t2.d);
The query plan might look like:
Project [a#0]
+- Filter a#0 IN (list#4 [b#1])
: +- Project [c#2]
: +- Filter (outer(b#1) < d#3)
: +- LocalRelation <empty>, [c#2, d#3]
+- LocalRelation <empty>, [a#0, b#1]
After `PullupCorrelatedPredicates`, it produces query plan like:
'Project [a#0]
+- 'Filter a#0 IN (list#4 [(b#1 < d#3)])
: +- Project [c#2, d#3]
: +- LocalRelation <empty>, [c#2, d#3]
+- LocalRelation <empty>, [a#0, b#1]
Because the correlated predicate involves another attribute `d#3` in subquery, it has been pulled out and added into the `Project` on the top of the subquery.
When `list` in `In` contains just one `ListQuery`, `In.checkInputDataTypes` checks if the size of `value` expressions matches the output size of subquery. In the above example, there is only `value` expression and the subquery output has two attributes `c#2, d#3`, so it fails the check and `In.resolved` returns `false`.
We should not let `In.checkInputDataTypes` wrongly report unresolved plans to fail the structural integrity check.
## How was this patch tested?
Added test.
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes#18968 from viirya/SPARK-21759.
## What changes were proposed in this pull request?
This is a refactoring of `ColumnVector` hierarchy and related classes.
1. make `ColumnVector` read-only
2. introduce `WritableColumnVector` with write interface
3. remove `ReadOnlyColumnVector`
## How was this patch tested?
Existing tests.
Author: Takuya UESHIN <ueshin@databricks.com>
Closes#18958 from ueshin/issues/SPARK-21745.
## What changes were proposed in this pull request?
While preparing to take over https://github.com/apache/spark/pull/16537, I realised a (I think) better approach to make the exception handling in one point.
This PR proposes to fix `_to_java_column` in `pyspark.sql.column`, which most of functions in `functions.py` and some other APIs use. This `_to_java_column` basically looks not working with other types than `pyspark.sql.column.Column` or string (`str` and `unicode`).
If this is not `Column`, then it calls `_create_column_from_name` which calls `functions.col` within JVM:
42b9eda80e/sql/core/src/main/scala/org/apache/spark/sql/functions.scala (L76)
And it looks we only have `String` one with `col`.
So, these should work:
```python
>>> from pyspark.sql.column import _to_java_column, Column
>>> _to_java_column("a")
JavaObject id=o28
>>> _to_java_column(u"a")
JavaObject id=o29
>>> _to_java_column(spark.range(1).id)
JavaObject id=o33
```
whereas these do not:
```python
>>> _to_java_column(1)
```
```
...
py4j.protocol.Py4JError: An error occurred while calling z:org.apache.spark.sql.functions.col. Trace:
py4j.Py4JException: Method col([class java.lang.Integer]) does not exist
...
```
```python
>>> _to_java_column([])
```
```
...
py4j.protocol.Py4JError: An error occurred while calling z:org.apache.spark.sql.functions.col. Trace:
py4j.Py4JException: Method col([class java.util.ArrayList]) does not exist
...
```
```python
>>> class A(): pass
>>> _to_java_column(A())
```
```
...
AttributeError: 'A' object has no attribute '_get_object_id'
```
Meaning most of functions using `_to_java_column` such as `udf` or `to_json` or some other APIs throw an exception as below:
```python
>>> from pyspark.sql.functions import udf
>>> udf(lambda x: x)(None)
```
```
...
py4j.protocol.Py4JJavaError: An error occurred while calling z:org.apache.spark.sql.functions.col.
: java.lang.NullPointerException
...
```
```python
>>> from pyspark.sql.functions import to_json
>>> to_json(None)
```
```
...
py4j.protocol.Py4JJavaError: An error occurred while calling z:org.apache.spark.sql.functions.col.
: java.lang.NullPointerException
...
```
**After this PR**:
```python
>>> from pyspark.sql.functions import udf
>>> udf(lambda x: x)(None)
...
```
```
TypeError: Invalid argument, not a string or column: None of type <type 'NoneType'>. For column literals, use 'lit', 'array', 'struct' or 'create_map' functions.
```
```python
>>> from pyspark.sql.functions import to_json
>>> to_json(None)
```
```
...
TypeError: Invalid argument, not a string or column: None of type <type 'NoneType'>. For column literals, use 'lit', 'array', 'struct' or 'create_map' functions.
```
## How was this patch tested?
Unit tests added in `python/pyspark/sql/tests.py` and manual tests.
Author: hyukjinkwon <gurwls223@gmail.com>
Author: zero323 <zero323@users.noreply.github.com>
Closes#19027 from HyukjinKwon/SPARK-19165.
## What changes were proposed in this pull request?
When json_tuple in extracting values from JSON it returns null values within repeated columns except the first one as below:
``` scala
scala> spark.sql("""SELECT json_tuple('{"a":1, "b":2}', 'a', 'b', 'a')""").show()
+---+---+----+
| c0| c1| c2|
+---+---+----+
| 1| 2|null|
+---+---+----+
```
I think this should be consistent with Hive's implementation:
```
hive> SELECT json_tuple('{"a": 1, "b": 2}', 'a', 'a');
...
1 1
```
In this PR, we located all the matched indices in `fieldNames` instead of returning the first matched index, i.e., indexOf.
## How was this patch tested?
Added test in JsonExpressionsSuite.
Author: Jen-Ming Chung <jenmingisme@gmail.com>
Closes#19017 from jmchung/SPARK-21804.
## What changes were proposed in this pull request?
The given example in the comment of Class ExchangeCoordinator is exist four post-shuffle partitions,but the current comment is “three”.
## How was this patch tested?
Author: lufei <lu.fei80@zte.com.cn>
Closes#19028 from figo77/SPARK-21816.
JIRA ticket: https://issues.apache.org/jira/browse/SPARK-21694
## What changes were proposed in this pull request?
Spark already supports launching containers attached to a given CNI network by specifying it via the config `spark.mesos.network.name`.
This PR adds support to pass in network labels to CNI plugins via a new config option `spark.mesos.network.labels`. These network labels are key-value pairs that are set in the `NetworkInfo` of both the driver and executor tasks. More details in the related Mesos documentation: http://mesos.apache.org/documentation/latest/cni/#mesos-meta-data-to-cni-plugins
## How was this patch tested?
Unit tests, for both driver and executor tasks.
Manual integration test to submit a job with the `spark.mesos.network.labels` option, hit the mesos/state.json endpoint, and check that the labels are set in the driver and executor tasks.
ArtRand skonto
Author: Susan X. Huynh <xhuynh@mesosphere.com>
Closes#18910 from susanxhuynh/sh-mesos-cni-labels.
## What changes were proposed in this pull request?
Code in vignettes requires winutils on windows to run, when publishing to CRAN or building from source, winutils might not be available, so it's better to disable code run (so resulting vigenttes will not have output from code, but text is still there and code is still there)
fix * checking re-building of vignette outputs ... WARNING
and
> %LOCALAPPDATA% not found. Please define the environment variable or restart and enter an installation path in localDir.
## How was this patch tested?
jenkins, appveyor, r-hub
before: https://artifacts.r-hub.io/SparkR_2.2.0.tar.gz-49cecef3bb09db1db130db31604e0293/SparkR.Rcheck/00check.log
after: https://artifacts.r-hub.io/SparkR_2.2.0.tar.gz-86a066c7576f46794930ad114e5cff7c/SparkR.Rcheck/00check.log
Author: Felix Cheung <felixcheung_m@hotmail.com>
Closes#19016 from felixcheung/rvigwind.
## What changes were proposed in this pull request?
The getAliasedConstraints fuction in LogicalPlan.scala will clone the expression set when an element added,
and it will take a long time. This PR add a function to add multiple elements at once to reduce the clone time.
Before modified, the cost of getAliasedConstraints is:
100 expressions: 41 seconds
150 expressions: 466 seconds
After modified, the cost of getAliasedConstraints is:
100 expressions: 1.8 seconds
150 expressions: 6.5 seconds
The test is like this:
test("getAliasedConstraints") {
val expressionNum = 150
val aggExpression = (1 to expressionNum).map(i => Alias(Count(Literal(1)), s"cnt$i")())
val aggPlan = Aggregate(Nil, aggExpression, LocalRelation())
val beginTime = System.currentTimeMillis()
val expressions = aggPlan.validConstraints
println(s"validConstraints cost: ${System.currentTimeMillis() - beginTime}ms")
// The size of Aliased expression is n * (n - 1) / 2 + n
assert( expressions.size === expressionNum * (expressionNum - 1) / 2 + expressionNum)
}
(Please fill in changes proposed in this fix)
## How was this patch tested?
(Please explain how this patch was tested. E.g. unit tests, integration tests, manual tests)
(If this patch involves UI changes, please attach a screenshot; otherwise, remove this)
Run new added test.
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: 10129659 <chen.yanshan@zte.com.cn>
Closes#19022 from eatoncys/getAliasedConstraints.
## What changes were proposed in this pull request?
This pr changed the default value of `maxLinesPerFunction` into `4000`. In #18810, we had this new option to disable code generation for too long functions and I found this option only affected `Q17` and `Q66` in TPC-DS. But, `Q66` had some performance regression:
```
Q17 w/o #18810, 3224ms --> q17 w/#18810, 2627ms (improvement)
Q66 w/o #18810, 1712ms --> q66 w/#18810, 3032ms (regression)
```
To keep the previous performance in TPC-DS, we better set higher value at `maxLinesPerFunction` by default.
## How was this patch tested?
Existing tests.
Author: Takeshi Yamamuro <yamamuro@apache.org>
Closes#19021 from maropu/SPARK-21603-FOLLOWUP-1.
Right now the spark shuffle service has a cache for index files. It is based on a # of files cached (spark.shuffle.service.index.cache.entries). This can cause issues if people have a lot of reducers because the size of each entry can fluctuate based on the # of reducers.
We saw an issues with a job that had 170000 reducers and it caused NM with spark shuffle service to use 700-800MB or memory in NM by itself.
We should change this cache to be memory based and only allow a certain memory size used. When I say memory based I mean the cache should have a limit of say 100MB.
https://issues.apache.org/jira/browse/SPARK-21501
Manual Testing with 170000 reducers has been performed with cache loaded up to max 100MB default limit, with each shuffle index file of size 1.3MB. Eviction takes place as soon as the total cache size reaches the 100MB limit and the objects will be ready for garbage collection there by avoiding NM to crash. No notable difference in runtime has been observed.
Author: Sanket Chintapalli <schintap@yahoo-inc.com>
Closes#18940 from redsanket/SPARK-21501.
## What changes were proposed in this pull request?
Modify MLP model to inherit `ProbabilisticClassificationModel` and so that it can expose the probability column when transforming data.
## How was this patch tested?
Test added.
Author: WeichenXu <WeichenXu123@outlook.com>
Closes#17373 from WeichenXu123/expose_probability_in_mlp_model.
## What changes were proposed in this pull request?
Add a new listener event when a speculative task is created and notify it to ExecutorAllocationManager for requesting more executor.
## How was this patch tested?
- Added Unittests.
- For the test snippet in the jira:
val n = 100
val someRDD = sc.parallelize(1 to n, n)
someRDD.mapPartitionsWithIndex( (index: Int, it: Iterator[Int]) => {
if (index == 1) {
Thread.sleep(Long.MaxValue) // fake long running task(s)
}
it.toList.map(x => index + ", " + x).iterator
}).collect
With this code change, spark indicates 101 jobs are running (99 succeeded, 2 running and 1 is speculative job)
Author: Jane Wang <janewang@fb.com>
Closes#18492 from janewangfb/speculated_task_not_launched.
## What changes were proposed in this pull request?
```sharedParams.scala``` was generated by ```SharedParamsCodeGen```, but it's not updated in master. Maybe someone manual update ```sharedParams.scala```, this PR fix this issue.
## How was this patch tested?
Offline check.
Author: Yanbo Liang <ybliang8@gmail.com>
Closes#19011 from yanboliang/sharedParams.
## What changes were proposed in this pull request?
All streaming logical plans will now have isStreaming set. This involved adding isStreaming as a case class arg in a few cases, since a node might be logically streaming depending on where it came from.
## How was this patch tested?
Existing unit tests - no functional change is intended in this PR.
Author: Jose Torres <joseph-torres@databricks.com>
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#18973 from joseph-torres/SPARK-21765.
## What changes were proposed in this pull request?
Added call to copy values of Params from Estimator to Model after fit in PySpark ML. This will copy values for any params that are also defined in the Model. Since currently most Models do not define the same params from the Estimator, also added method to create new Params from looking at the Java object if they do not exist in the Python object. This is a temporary fix that can be removed once the PySpark models properly define the params themselves.
## How was this patch tested?
Refactored the `check_params` test to optionally check if the model params for Python and Java match and added this check to an existing fitted model that shares params between Estimator and Model.
Author: Bryan Cutler <cutlerb@gmail.com>
Closes#17849 from BryanCutler/pyspark-models-own-params-SPARK-10931.
## What changes were proposed in this pull request?
fix bug of MLOR do not work correctly when featureStd contains zero
We can reproduce the bug through such dataset (features including zero variance), will generate wrong result (all coefficients becomes 0)
```
val multinomialDatasetWithZeroVar = {
val nPoints = 100
val coefficients = Array(
-0.57997, 0.912083, -0.371077,
-0.16624, -0.84355, -0.048509)
val xMean = Array(5.843, 3.0)
val xVariance = Array(0.6856, 0.0) // including zero variance
val testData = generateMultinomialLogisticInput(
coefficients, xMean, xVariance, addIntercept = true, nPoints, seed)
val df = sc.parallelize(testData, 4).toDF().withColumn("weight", lit(1.0))
df.cache()
df
}
```
## How was this patch tested?
testcase added.
Author: WeichenXu <WeichenXu123@outlook.com>
Closes#18896 from WeichenXu123/fix_mlor_stdvalue_zero_bug.
## What changes were proposed in this pull request?
For Hive-serde tables, we always respect the schema stored in Hive metastore, because the schema could be altered by the other engines that share the same metastore. Thus, we always trust the metastore-controlled schema for Hive-serde tables when the schemas are different (without considering the nullability and cases). However, in some scenarios, Hive metastore also could INCORRECTLY overwrite the schemas when the serde and Hive metastore built-in serde are different.
The proposed solution is to introduce a table-specific option for such scenarios. For a specific table, users can make Spark always respect Spark-inferred/controlled schema instead of trusting metastore-controlled schema. By default, we trust Hive metastore-controlled schema.
## How was this patch tested?
Added a cross-version test case
Author: gatorsmile <gatorsmile@gmail.com>
Closes#19003 from gatorsmile/respectSparkSchema.
## What changes were proposed in this pull request?
This PR is to enable users to create persistent Scala UDAF (that extends UserDefinedAggregateFunction).
```SQL
CREATE FUNCTION myDoubleAvg AS 'test.org.apache.spark.sql.MyDoubleAvg'
```
Before this PR, Spark UDAF only can be registered through the API `spark.udf.register(...)`
## How was this patch tested?
Added test cases
Author: gatorsmile <gatorsmile@gmail.com>
Closes#18700 from gatorsmile/javaUDFinScala.
There're two code in Launcher and SparkSubmit will will explicitly list all the Spark submodules, newly added kvstore module is missing in this two parts, so submitting a minor PR to fix this.
Author: jerryshao <sshao@hortonworks.com>
Closes#19014 from jerryshao/missing-kvstore.
## What changes were proposed in this pull request?
We do not have any Hive-specific parser. It does not make sense to keep a parser-specific test suite `HiveDDLCommandSuite.scala` in the Hive package. This PR is to remove it.
## How was this patch tested?
N/A
Author: gatorsmile <gatorsmile@gmail.com>
Closes#19015 from gatorsmile/combineDDL.
## What changes were proposed in this pull request?
SPARK-21100 introduced a new `summary` method to the Scala/Java Dataset API that included expanded statistics (vs `describe`) and control over which statistics to compute. Currently in the R API `summary` acts as an alias for `describe`. This patch updates the R API to call the new `summary` method in the JVM that includes additional statistics and ability to select which to compute.
This does not break the current interface as the present `summary` method does not take additional arguments like `describe` and the output was never meant to be used programmatically.
## How was this patch tested?
Modified and additional unit tests.
Author: Andrew Ray <ray.andrew@gmail.com>
Closes#18786 from aray/summary-r.
## What changes were proposed in this pull request?
Based on https://github.com/apache/spark/pull/18282 by rgbkrk this PR attempts to update to the current released cloudpickle and minimize the difference between Spark cloudpickle and "stock" cloud pickle with the goal of eventually using the stock cloud pickle.
Some notable changes:
* Import submodules accessed by pickled functions (cloudpipe/cloudpickle#80)
* Support recursive functions inside closures (cloudpipe/cloudpickle#89, cloudpipe/cloudpickle#90)
* Fix ResourceWarnings and DeprecationWarnings (cloudpipe/cloudpickle#88)
* Assume modules with __file__ attribute are not dynamic (cloudpipe/cloudpickle#85)
* Make cloudpickle Python 3.6 compatible (cloudpipe/cloudpickle#72)
* Allow pickling of builtin methods (cloudpipe/cloudpickle#57)
* Add ability to pickle dynamically created modules (cloudpipe/cloudpickle#52)
* Support method descriptor (cloudpipe/cloudpickle#46)
* No more pickling of closed files, was broken on Python 3 (cloudpipe/cloudpickle#32)
* ** Remove non-standard __transient__check (cloudpipe/cloudpickle#110)** -- while we don't use this internally, and have no tests or documentation for its use, downstream code may use __transient__, although it has never been part of the API, if we merge this we should include a note about this in the release notes.
* Support for pickling loggers (yay!) (cloudpipe/cloudpickle#96)
* BUG: Fix crash when pickling dynamic class cycles. (cloudpipe/cloudpickle#102)
## How was this patch tested?
Existing PySpark unit tests + the unit tests from the cloudpickle project on their own.
Author: Holden Karau <holden@us.ibm.com>
Author: Kyle Kelley <rgbkrk@gmail.com>
Closes#18734 from holdenk/holden-rgbkrk-cloudpickle-upgrades.
## What changes were proposed in this pull request?
MLlib ```LinearRegression/LogisticRegression/LinearSVC``` always standardize the data during training to improve the rate of convergence regardless of _standardization_ is true or false. If _standardization_ is false, we perform reverse standardization by penalizing each component differently to get effectively the same objective function when the training dataset is not standardized. We should keep these comments in the code to let developers understand how we handle it correctly.
## How was this patch tested?
Existing tests, only adding some comments in code.
Author: Yanbo Liang <ybliang8@gmail.com>
Closes#18992 from yanboliang/SPARK-19762.
For Hive tables, the current "replace the schema" code is the correct
path, except that an exception in that path should result in an error, and
not in retrying in a different way.
For data source tables, Spark may generate a non-compatible Hive table;
but for that to work with Hive 2.1, the detection of data source tables needs
to be fixed in the Hive client, to also consider the raw tables used by code
such as `alterTableSchema`.
Tested with existing and added unit tests (plus internal tests with a 2.1 metastore).
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#18849 from vanzin/SPARK-21617.
## What changes were proposed in this pull request?
The previous PR(https://github.com/apache/spark/pull/19000) removed filter pushdown verification, This PR add them back.
## How was this patch tested?
manual tests
Author: Yuming Wang <wgyumg@gmail.com>
Closes#19002 from wangyum/SPARK-21790-follow-up.
Add Python API for `FeatureHasher` transformer.
## How was this patch tested?
New doc test.
Author: Nick Pentreath <nickp@za.ibm.com>
Closes#18970 from MLnick/SPARK-21468-pyspark-hasher.
## What changes were proposed in this pull request?
Reduce 'Skipping partitions' message to debug
## How was this patch tested?
Existing tests
Author: Sean Owen <sowen@cloudera.com>
Closes#19010 from srowen/SPARK-21718.