## What changes were proposed in this pull request?
coalesce doesn't handle UnionRDD with partial locality properly. I had a user who had a UnionRDD that was made up of mapPartitionRDD without preferred locations and a checkpointedRDD with preferred locations (getting from hdfs). It took the driver over 20 minutes to setup the groups and put the partitions into those groups before it even started any tasks. Even perhaps worse is it didn't end up with the number of partitions he was asking for because it didn't put a partition in each of the groups properly.
The changes in this patch get rid of a n^2 while loop that was causing the 20 minutes, it properly distributes the partitions to have at least one per group, and it changes from using the rotation iterator which got the preferred locations many times to get all the preferred locations once up front.
Note that the n^2 while loop that I removed in setupGroups took so long because all of the partitions with preferred locations were already assigned to group, so it basically looped through every single one and wasn't ever able to assign it. At the time I had 960 partitions with preferred locations and 1020 without and did the outer while loop 319 times because that is the # of groups left to create. Note that each of those times through the inner while loop is going off to hdfs to get the block locations, so this is extremely inefficient.
## How was the this patch tested?
Added unit tests for this case and ran existing ones that applied to make sure no regressions.
Also manually tested on the users production job to make sure it fixed their issue. It created the proper number of partitions and now it takes about 6 seconds rather then 20 minutes.
I did also run some basic manual tests with spark-shell doing coalesced to smaller number, same number, and then greater with shuffle.
Author: Thomas Graves <tgraves@prevailsail.corp.gq1.yahoo.com>
Closes#11327 from tgravescs/SPARK-11316.
## What changes were proposed in this pull request?
Observed stackOverflowError in Kryo when executing TPC-DS Query27. Spark thrift server disables kryo reference tracking (if not specified in conf). When "spark.kryo.referenceTracking" is set to true explicitly in spark-defaults.conf, query executes successfully. The root cause is that the TaskMemoryManager inside MemoryConsumer and LongToUnsafeRowMap were not transient and thus were serialized and broadcast around from within LongHashedRelation, which could potentially cause circular reference inside Kryo. But the TaskMemoryManager is per task and should not be passed around at the first place. This fix makes it transient.
## How was this patch tested?
core/test, hive/test, sql/test, catalyst/test, dev/lint-scala, org.apache.spark.sql.hive.execution.HiveCompatibilitySuite, dev/scalastyle,
manual test of TBC-DS Query 27 with 1GB data but without the "limit 100" which would cause a NPE due to SPARK-14752.
Author: yzhou2001 <yzhou_1999@yahoo.com>
Closes#12598 from yzhou2001/master.
## What changes were proposed in this pull request?
Resetting the task interruption status before updating the task status.
## How was this patch tested?
I have verified it manually by running multiple applications, Executor doesn't crash and updates the status to the driver without any exceptions with the patch changes.
Author: Devaraj K <devaraj@apache.org>
Closes#12031 from devaraj-kavali/SPARK-14234.
## What changes were proposed in this pull request?
In some cases, fine-grained lock have race condition with class-loader lock and have caused dead lock issue. It is safe to drop this fine grained lock and load all classes by single class-loader lock.
Author: Zheng Tan <zheng.tan@hulu.com>
Closes#12857 from tankkyo/master.
## What changes were proposed in this pull request?
Added tests for ListAccumulator and LegacyAccumulatorWrapper, test for ListAccumulator is one similar to old Collection Accumulators
## How was this patch tested?
Ran tests locally.
cc rxin
Author: Sandeep Singh <sandeep@techaddict.me>
Closes#12862 from techaddict/SPARK-15082.
- that reduceFunc and invReduceFunc should be associative
- that the intermediate result in iterated applications of inverseReduceFunc
is its first argument
Author: François Garillot <francois@garillot.net>
Closes#8103 from huitseeker/issue/invReduceFuncDoc.
## What changes were proposed in this pull request?
Remove AccumulatorV2.localValue and keep only value
## How was this patch tested?
existing tests
Author: Sandeep Singh <sandeep@techaddict.me>
Closes#12865 from techaddict/SPARK-15087.
## What changes were proposed in this pull request?
This PR updates `QueryStatusCollector.reset` to create Waiter instead of calling `await(1 milliseconds)` to bypass an ScalaTest's issue that Waiter.await may block forever.
## How was this patch tested?
I created a local stress test to call codes in `test("event ordering")` 100 times. It cannot pass without this patch.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#12623 from zsxwing/flaky-test.
# What changes were proposed in this pull request?
Support partitioning in the file stream sink. This is implemented using a new, but simpler code path for writing parquet files - both unpartitioned and partitioned. This new code path does not use Output Committers, as we will eventually write the file names to the metadata log for "committing" them.
This patch duplicates < 100 LOC from the WriterContainer. But its far simpler that WriterContainer as it does not involve output committing. In addition, it introduces the new APIs in FileFormat and OutputWriterFactory in an attempt to simplify the APIs (not have Job in the `FileFormat` API, not have bucket and other stuff in the `OutputWriterFactory.newInstance()` ).
# Tests
- New unit tests to test the FileStreamSinkWriter for partitioned and unpartitioned files
- New unit test to partially test the FileStreamSink for partitioned files (does not test recovery of partition column data, as that requires change in the StreamFileCatalog, future PR).
- Updated FileStressSuite to test number of records read from partitioned output files.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#12409 from tdas/streaming-partitioned-parquet.
## What changes were proposed in this pull request?
This patch removes SparkSqlSerializer. I believe this is now dead code.
## How was this patch tested?
Removed a test case related to it.
Author: Reynold Xin <rxin@databricks.com>
Closes#12864 from rxin/SPARK-15088.
## What changes were proposed in this pull request?
Fix warnings and a failure in SparkR test cases with testthat version 1.0.1
## How was this patch tested?
SparkR unit test cases.
Author: Sun Rui <sunrui2016@gmail.com>
Closes#12867 from sun-rui/SPARK-15091.
## What changes were proposed in this pull request?
PySpark ML Params setter code clean up.
For examples,
```setInputCol``` can be simplified from
```
self._set(inputCol=value)
return self
```
to:
```
return self._set(inputCol=value)
```
This is a pretty big sweeps, and we cleaned wherever possible.
## How was this patch tested?
Exist unit tests.
Author: Yanbo Liang <ybliang8@gmail.com>
Closes#12749 from yanboliang/spark-14971.
## What changes were proposed in this pull request?
This PR removes a stale TODO comment in `GraphGenerators.scala`
## How was this patch tested?
Just comment removed.
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#12839 from dongjoon-hyun/SPARK-15057.
## What changes were proposed in this pull request?
Update Jetty 8.1 to the latest 2016/02 release, from a 2013/10 release, for security and bug fixes. This does not resolve the JIRA necessarily, as it's still worth considering an update to 9.3.
## How was this patch tested?
Jenkins tests
Author: Sean Owen <sowen@cloudera.com>
Closes#12842 from srowen/SPARK-14897.
## What changes were proposed in this pull request?
This patch moves AccumulatorV2 and subclasses into util package.
## How was this patch tested?
Updated relevant tests.
Author: Reynold Xin <rxin@databricks.com>
Closes#12863 from rxin/SPARK-15081.
## What changes were proposed in this pull request?
When ALS is run with a checkpoint interval, during the checkpoint materialize the current state and cleanup the previous shuffles (non-blocking).
## How was this patch tested?
Existing ALS unit tests, new ALS checkpoint cleanup unit tests added & shuffle files checked after ALS w/checkpointing run.
Author: Holden Karau <holden@us.ibm.com>
Author: Holden Karau <holden@pigscanfly.ca>
Closes#11919 from holdenk/SPARK-6717-clear-shuffle-files-after-checkpointing-in-ALS.
## What changes were proposed in this pull request?
This is a follow up PR for #11583. It makes 3 lazy vals into just vals and adds unit test coverage.
## How was this patch tested?
Existing unit tests and additional unit tests.
Author: Andrew Ray <ray.andrew@gmail.com>
Closes#12861 from aray/fast-pivot-follow-up.
## What changes were proposed in this pull request?
This patch removes AverageAccumulator and adds the ability to compute average to LongAccumulator and DoubleAccumulator. The patch also improves documentation for the two accumulators.
## How was this patch tested?
Added unit tests for this.
Author: Reynold Xin <rxin@databricks.com>
Closes#12858 from rxin/SPARK-15079.
## What changes were proposed in this pull request?
This updates the java-/scala- doc for setLocalProperty to document heritability of localProperties. This also adds tests for that behaviour.
## How was this patch tested?
Tests pass. New tests were added.
Author: Marcin Tustin <marcin.tustin@gmail.com>
Closes#12455 from marcintustin/SPARK-14685.
## What changes were proposed in this pull request?
Right now `StreamExecution.awaitBatchLock` uses an unfair lock. `StreamExecution.awaitOffset` may run too long and fail some test because `StreamExecution.constructNextBatch` keeps getting the lock.
See: https://amplab.cs.berkeley.edu/jenkins/job/spark-master-test-sbt-hadoop-2.4/865/testReport/junit/org.apache.spark.sql.streaming/FileStreamSourceStressTestSuite/file_source_stress_test/
This PR uses a fair ReentrantLock to resolve the thread starvation issue.
## How was this patch tested?
Modified `FileStreamSourceStressTestSuite.test("file source stress test")` to run the test codes 100 times locally. It always fails because of timeout without this patch.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#12852 from zsxwing/SPARK-15077.
## What changes were proposed in this pull request?
Make serializer correctly inferred if the input type is `List[_]`, since `List[_]` is type of `Seq[_]`, before it was matched to different case (`case t if definedByConstructorParams(t)`).
## How was this patch tested?
New test case was added.
Author: bomeng <bmeng@us.ibm.com>
Closes#12849 from bomeng/SPARK-15062.
## What changes were proposed in this pull request?
This PR addresses a few minor issues in SQL parser:
- Removes some unused rules and keywords in the grammar.
- Removes code path for fallback SQL parsing (was needed for Hive native parsing).
- Use `UnresolvedGenerator` instead of hard-coding `Explode` & `JsonTuple`.
- Adds a more generic way of creating error messages for unsupported Hive features.
- Use `visitFunctionName` as much as possible.
- Interpret a `CatalogColumn`'s `DataType` directly instead of parsing it again.
## How was this patch tested?
Existing tests.
Author: Herman van Hovell <hvanhovell@questtec.nl>
Closes#12826 from hvanhovell/SPARK-15047.
## What changes were proposed in this pull request?
https://issues.apache.org/jira/browse/SPARK-15050
This PR adds function parameters for Python API for reading and writing `csv()`.
## How was this patch tested?
This was tested by `./dev/run_tests`.
Author: hyukjinkwon <gurwls223@gmail.com>
Author: Hyukjin Kwon <gurwls223@gmail.com>
Closes#12834 from HyukjinKwon/SPARK-15050.
## What changes were proposed in this pull request?
In this PR we add support for correlated scalar subqueries. An example of such a query is:
```SQL
select * from tbl1 a where a.value > (select max(value) from tbl2 b where b.key = a.key)
```
The implementation adds the `RewriteCorrelatedScalarSubquery` rule to the Optimizer. This rule plans these subqueries using `LEFT OUTER` joins. It currently supports rewrites for `Project`, `Aggregate` & `Filter` logical plans.
I could not find a well defined semantics for the use of scalar subqueries in an `Aggregate`. The current implementation currently evaluates the scalar subquery *before* aggregation. This means that you either have to make scalar subquery part of the grouping expression, or that you have to aggregate it further on. I am open to suggestions on this.
The implementation currently forces the uniqueness of a scalar subquery by enforcing that it is aggregated and that the resulting column is wrapped in an `AggregateExpression`.
## How was this patch tested?
Added tests to `SubquerySuite`.
Author: Herman van Hovell <hvanhovell@questtec.nl>
Closes#12822 from hvanhovell/SPARK-14785.
The contribution is my original work and that I license the work to the project under the project's open source license.
Author: poolis <gmichalopoulos@gmail.com>
Author: Greg Michalopoulos <gmichalopoulos@gmail.com>
Closes#10899 from poolis/spark-12928.
## What changes were proposed in this pull request?
This patch creates a builder pattern for creating SparkSession. The new code is unused and mostly deadcode. I'm putting it up here for feedback.
There are a few TODOs that can be done as follow-up pull requests:
- [ ] Update tests to use this
- [ ] Update examples to use this
- [ ] Clean up SQLContext code w.r.t. this one (i.e. SparkSession shouldn't call into SQLContext.getOrCreate; it should be the other way around)
- [ ] Remove SparkSession.withHiveSupport
- [ ] Disable the old constructor (by making it private) so the only way to start a SparkSession is through this builder pattern
## How was this patch tested?
Part of the future pull request is to clean this up and switch existing tests to use this.
Author: Reynold Xin <rxin@databricks.com>
Closes#12830 from rxin/sparksession-builder.
## What changes were proposed in this pull request?
This patch deprecates the old accumulator API.
## How was this patch tested?
N/A
Author: Reynold Xin <rxin@databricks.com>
Closes#12832 from rxin/SPARK-15054.
## What changes were proposed in this pull request?
parquet datasource and ColumnarBatch tests fail on big-endian platforms This patch adds support for the little-endian byte arrays being correctly interpreted on a big-endian platform
## How was this patch tested?
Spark test builds ran on big endian z/Linux and regression build on little endian amd64
Author: Pete Robbins <robbinspg@gmail.com>
Closes#12397 from robbinspg/master.
## What changes were proposed in this pull request?
In order to support nested predicate subquery, this PR introduce an internal join type ExistenceJoin, which will emit all the rows from left, plus an additional column, which presents there are any rows matched from right or not (it's not null-aware right now). This additional column could be used to replace the subquery in Filter.
In theory, all the predicate subquery could use this join type, but it's slower than LeftSemi and LeftAnti, so it's only used for nested subquery (subquery inside OR).
For example, the following SQL:
```sql
SELECT a FROM t WHERE EXISTS (select 0) OR EXISTS (select 1)
```
This PR also fix a bug in predicate subquery push down through join (they should not).
Nested null-aware subquery is still not supported. For example, `a > 3 OR b NOT IN (select bb from t)`
After this, we could run TPCDS query Q10, Q35, Q45
## How was this patch tested?
Added unit tests.
Author: Davies Liu <davies@databricks.com>
Closes#12820 from davies/or_exists.
## What changes were proposed in this pull request?
#12339 didn't fix the race condition. MemorySinkSuite is still flaky: https://amplab.cs.berkeley.edu/jenkins/job/spark-master-test-maven-hadoop-2.2/814/testReport/junit/org.apache.spark.sql.streaming/MemorySinkSuite/registering_as_a_table/
Here is an execution order to reproduce it.
| Time |Thread 1 | MicroBatchThread |
|:-------------:|:-------------:|:-----:|
| 1 | | `MemorySink.getOffset` |
| 2 | | availableOffsets ++= newData (availableOffsets is not changed here) |
| 3 | addData(newData) | |
| 4 | Set `noNewData` to `false` in processAllAvailable | |
| 5 | | `dataAvailable` returns `false` |
| 6 | | noNewData = true |
| 7 | `noNewData` is true so just return | |
| 8 | assert results and fail | |
| 9 | | `dataAvailable` returns true so process the new batch |
This PR expands the scope of `awaitBatchLock.synchronized` to eliminate the above race.
## How was this patch tested?
test("stress test"). It always failed before this patch. And it will pass after applying this patch. Ignore this test in the PR as it takes several minutes to finish.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#12582 from zsxwing/SPARK-14579-2.
## What changes were proposed in this pull request?
The existing implementation of pivot translates into a single aggregation with one aggregate per distinct pivot value. When the number of distinct pivot values is large (say 1000+) this can get extremely slow since each input value gets evaluated on every aggregate even though it only affects the value of one of them.
I'm proposing an alternate strategy for when there are 10+ (somewhat arbitrary threshold) distinct pivot values. We do two phases of aggregation. In the first we group by the grouping columns plus the pivot column and perform the specified aggregations (one or sometimes more). In the second aggregation we group by the grouping columns and use the new (non public) PivotFirst aggregate that rearranges the outputs of the first aggregation into an array indexed by the pivot value. Finally we do a project to extract the array entries into the appropriate output column.
## How was this patch tested?
Additional unit tests in DataFramePivotSuite and manual larger scale testing.
Author: Andrew Ray <ray.andrew@gmail.com>
Closes#11583 from aray/fast-pivot.
## What changes were proposed in this pull request?
initialize SparkSubmitArgument#files first from spark-submit arguments then from properties file, so that sys property spark.yarn.dist.files will be set correctly.
```
OptionAssigner(args.files, YARN, ALL_DEPLOY_MODES, sysProp = "spark.yarn.dist.files"),
```
## How was this patch tested?
manul test. file defined in properties file is also distributed to driver in yarn-cluster mode.
Author: Jeff Zhang <zjffdu@apache.org>
Closes#12656 from zjffdu/SPARK-14845.
## What changes were proposed in this pull request?
Simplify and clean up some object expressions:
1. simplify the logic to handle `propagateNull`
2. add `propagateNull` parameter to `Invoke`
3. simplify the unbox logic in `Invoke`
4. other minor cleanup
TODO: simplify `MapObjects`
## How was this patch tested?
existing tests.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#12399 from cloud-fan/object.
## What changes were proposed in this pull request?
NewAccumulator isn't the best name if we ever come up with v3 of the API.
## How was this patch tested?
Updated tests to reflect the change.
Author: Reynold Xin <rxin@databricks.com>
Closes#12827 from rxin/SPARK-15049.
## What changes were proposed in this pull request?
This PR adds the explanation and documentation for CSV options for reading and writing.
## How was this patch tested?
Style tests with `./dev/run_tests` for documentation style.
Author: hyukjinkwon <gurwls223@gmail.com>
Author: Hyukjin Kwon <gurwls223@gmail.com>
Closes#12817 from HyukjinKwon/SPARK-13425.
## What changes were proposed in this pull request?
This PR is an update for [https://github.com/apache/spark/pull/12738] which:
* Adds a generic unit test for JavaParams wrappers in pyspark.ml for checking default Param values vs. the defaults in the Scala side
* Various fixes for bugs found
* This includes changing classes taking weightCol to treat unset and empty String Param values the same way.
Defaults changed:
* Scala
* LogisticRegression: weightCol defaults to not set (instead of empty string)
* StringIndexer: labels default to not set (instead of empty array)
* GeneralizedLinearRegression:
* maxIter always defaults to 25 (simpler than defaulting to 25 for a particular solver)
* weightCol defaults to not set (instead of empty string)
* LinearRegression: weightCol defaults to not set (instead of empty string)
* Python
* MultilayerPerceptron: layers default to not set (instead of [1,1])
* ChiSqSelector: numTopFeatures defaults to 50 (instead of not set)
## How was this patch tested?
Generic unit test. Manually tested that unit test by changing defaults and verifying that broke the test.
Author: Joseph K. Bradley <joseph@databricks.com>
Author: yinxusen <yinxusen@gmail.com>
Closes#12816 from jkbradley/yinxusen-SPARK-14931.
## What changes were proposed in this pull request?
This is caused by https://github.com/apache/spark/pull/12776, which removes the `synchronized` from all methods in `AccumulatorContext`.
However, a test in `CachedTableSuite` synchronize on `AccumulatorContext` and expecting no one else can change it, which is not true anymore.
This PR update that test to not require to lock on `AccumulatorContext`.
## How was this patch tested?
N/A
Author: Wenchen Fan <wenchen@databricks.com>
Closes#12811 from cloud-fan/flaky.
1. Adds the following options for parsing NaNs: nanValue
2. Adds the following options for parsing infinity: positiveInf, negativeInf.
`TypeCast.castTo` is unit tested and an end-to-end test is added to `CSVSuite`
Author: Hossein <hossein@databricks.com>
Closes#11947 from falaki/SPARK-14143.
This PR contains three changes:
1. We will use spark.sql.warehouse.dir set warehouse location. We will not use hive.metastore.warehouse.dir.
2. SessionCatalog needs to set the location to default db. Otherwise, when creating a table in SparkSession without hive support, the default db's path will be an empty string.
3. When we create a database, we need to make the path qualified.
Existing tests and new tests
Author: Yin Huai <yhuai@databricks.com>
Closes#12812 from yhuai/warehouse.
## What changes were proposed in this pull request?
* ```RFormula``` supports empty response variable like ```~ x + y```.
* Support formula in ```spark.kmeans``` in SparkR.
* Fix some outdated docs for SparkR.
## How was this patch tested?
Unit tests.
Author: Yanbo Liang <ybliang8@gmail.com>
Closes#12813 from yanboliang/spark-15030.
#### What changes were proposed in this pull request?
This PR removes three methods the were deprecated in 1.6.0:
- `PortableDataStream.close()`
- `LinearRegression.weights`
- `LogisticRegression.weights`
The rationale for doing this is that the impact is small and that Spark 2.0 is a major release.
#### How was this patch tested?
Compilation succeded.
Author: Herman van Hovell <hvanhovell@questtec.nl>
Closes#12732 from hvanhovell/SPARK-14952.
## What changes were proposed in this pull request?
This PR moves Vector.toJson/fromJson to ml.linalg.VectorEncoder under mllib/ to keep mllib-local's dependency minimal. The json encoding is used by Params. So we still need this feature in SPARK-14615, where we will switch to ml.linalg in spark.ml APIs.
## How was this patch tested?
Copied existing unit tests over.
cc; dbtsai
Author: Xiangrui Meng <meng@databricks.com>
Closes#12802 from mengxr/SPARK-14653.
## What changes were proposed in this pull request?
This PR fixes the bug that generates infinite distances between word vectors. For example,
Before this PR, we have
```
val synonyms = model.findSynonyms("who", 40)
```
will give the following results:
```
to Infinity
and Infinity
that Infinity
with Infinity
```
With this PR, the distance between words is a value between 0 and 1, as follows:
```
scala> model.findSynonyms("who", 10)
res0: Array[(String, Double)] = Array((Harvard-educated,0.5253688097000122), (ex-SAS,0.5213794708251953), (McMutrie,0.5187736749649048), (fellow,0.5166833400726318), (businessman,0.5145374536514282), (American-born,0.5127736330032349), (British-born,0.5062344074249268), (gray-bearded,0.5047978162765503), (American-educated,0.5035858750343323), (mentored,0.49849334359169006))
scala> model.findSynonyms("king", 10)
res1: Array[(String, Double)] = Array((queen,0.6787897944450378), (prince,0.6786158084869385), (monarch,0.659771203994751), (emperor,0.6490438580513), (goddess,0.643266499042511), (dynasty,0.635733425617218), (sultan,0.6166239380836487), (pharaoh,0.6150713562965393), (birthplace,0.6143025159835815), (empress,0.6109727025032043))
scala> model.findSynonyms("queen", 10)
res2: Array[(String, Double)] = Array((princess,0.7670737504959106), (godmother,0.6982434988021851), (raven-haired,0.6877717971801758), (swan,0.684934139251709), (hunky,0.6816608309745789), (Titania,0.6808111071586609), (heroine,0.6794036030769348), (king,0.6787897944450378), (diva,0.67848801612854), (lip-synching,0.6731793284416199))
```
### There are two places changed in this PR:
- Normalize the word vector to avoid overflow when calculating inner product between word vectors. This also simplifies the distance calculation, since the word vectors only need to be normalized once.
- Scale the learning rate by number of iteration, to be consistent with Google Word2Vec implementation
## How was this patch tested?
Use word2vec to train text corpus, and run model.findSynonyms() to get the distances between word vectors.
Author: Junyang <fly.shenjy@gmail.com>
Author: flyskyfly <fly.shenjy@gmail.com>
Closes#11812 from flyjy/TVec.
