## What changes were proposed in this pull request?
This adds support for radix sort of nullable long fields. When a sort field is null and radix sort is enabled, we keep nulls in a separate region of the sort buffer so that radix sort does not need to deal with them. This also has performance benefits when sorting smaller integer types, since the current representation of nulls in two's complement (Long.MIN_VALUE) otherwise forces a full-width radix sort.
This strategy for nulls does mean the sort is no longer stable. cc davies
## How was this patch tested?
Existing randomized sort tests for correctness. I also tested some TPCDS queries and there does not seem to be any significant regression for non-null sorts.
Some test queries (best of 5 runs each).
Before change:
scala> val start = System.nanoTime; spark.range(5000000).selectExpr("if(id > 5, cast(hash(id) as long), NULL) as h").coalesce(1).orderBy("h").collect(); (System.nanoTime - start) / 1e6
start: Long = 3190437233227987
res3: Double = 4716.471091
After change:
scala> val start = System.nanoTime; spark.range(5000000).selectExpr("if(id > 5, cast(hash(id) as long), NULL) as h").coalesce(1).orderBy("h").collect(); (System.nanoTime - start) / 1e6
start: Long = 3190367870952791
res4: Double = 2981.143045
Author: Eric Liang <ekl@databricks.com>
Closes#13161 from ericl/sc-2998.
## What changes were proposed in this pull request?
Spark currently incorrectly continues to use cached data even if the underlying data is overwritten.
Current behavior:
```scala
val dir = "/tmp/test"
sqlContext.range(1000).write.mode("overwrite").parquet(dir)
val df = sqlContext.read.parquet(dir).cache()
df.count() // outputs 1000
sqlContext.range(10).write.mode("overwrite").parquet(dir)
sqlContext.read.parquet(dir).count() // outputs 1000 <---- We are still using the cached dataset
```
This patch fixes this bug by adding support for `REFRESH path` that invalidates and refreshes all the cached data (and the associated metadata) for any dataframe that contains the given data source path.
Expected behavior:
```scala
val dir = "/tmp/test"
sqlContext.range(1000).write.mode("overwrite").parquet(dir)
val df = sqlContext.read.parquet(dir).cache()
df.count() // outputs 1000
sqlContext.range(10).write.mode("overwrite").parquet(dir)
spark.catalog.refreshResource(dir)
sqlContext.read.parquet(dir).count() // outputs 10 <---- We are not using the cached dataset
```
## How was this patch tested?
Unit tests for overwrites and appends in `ParquetQuerySuite` and `CachedTableSuite`.
Author: Sameer Agarwal <sameer@databricks.com>
Closes#13566 from sameeragarwal/refresh-path-2.
## What changes were proposed in this pull request?
The base class `SpecificParquetRecordReaderBase` used for vectorized parquet reader will try to get pushed-down filters from the given configuration. This pushed-down filters are used for RowGroups-level filtering. However, we don't set up the filters to push down into the configuration. In other words, the filters are not actually pushed down to do RowGroups-level filtering. This patch is to fix this and tries to set up the filters for pushing down to configuration for the reader.
## How was this patch tested?
Existing tests should be passed.
Author: Liang-Chi Hsieh <simonh@tw.ibm.com>
Closes#13371 from viirya/vectorized-reader-push-down-filter.
## What changes were proposed in this pull request?
As discussed in https://github.com/apache/spark/pull/12836
we need to override stringArgs method in MapPartitionsInR in order to avoid too large strings generated by "stringArgs" method based on the input arguments.
In this case exclude some of the input arguments: serialized R objects.
## How was this patch tested?
Existing test cases
Author: Narine Kokhlikyan <narine.kokhlikyan@gmail.com>
Closes#13610 from NarineK/dapply_MapPartitionsInR_stringArgs.
## What changes were proposed in this pull request?
Serializer instantiation will consider existing SparkConf
## How was this patch tested?
manual test with `ImmutableList` (Guava) and `kryo-serializers`'s `Immutable*Serializer` implementations.
Added Test Suite.
(If this patch involves UI changes, please attach a screenshot; otherwise, remove this)
Author: Sela <ansela@paypal.com>
Closes#13424 from amitsela/SPARK-15489.
## What changes were proposed in this pull request?
Code generated `SortMergeJoin` failed with wrong results when using structs as keys. This could (eventually) be traced back to the use of a wrong row reference when comparing structs.
## How was this patch tested?
TBD
Author: Herman van Hovell <hvanhovell@databricks.com>
Closes#13589 from hvanhovell/SPARK-15822.
## What changes were proposed in this pull request?
In scala, immutable.List.length is an expensive operation so we should
avoid using Seq.length == 0 or Seq.lenth > 0, and use Seq.isEmpty and Seq.nonEmpty instead.
## How was this patch tested?
existing tests
Author: wangyang <wangyang@haizhi.com>
Closes#13601 from yangw1234/isEmpty.
## What changes were proposed in this pull request?
Replace all occurrences of `None: Option[X]` with `Option.empty[X]`
## How was this patch tested?
Exisiting Tests
Author: Sandeep Singh <sandeep@techaddict.me>
Closes#13591 from techaddict/minor-7.
## What changes were proposed in this pull request?
This PR moves `QueryPlanner.planLater()` method into `GenericStrategy` for extra strategies to be able to use `planLater` in its strategy.
## How was this patch tested?
Existing tests.
Author: Takuya UESHIN <ueshin@happy-camper.st>
Closes#13147 from ueshin/issues/SPARK-6320.
## What changes were proposed in this pull request?
This patch moves some codes in `DataFrameWriter.insertInto` that belongs to `Analyzer`.
## How was this patch tested?
Existing tests.
Author: Liang-Chi Hsieh <simonh@tw.ibm.com>
Closes#13496 from viirya/move-analyzer-stuff.
## What changes were proposed in this pull request?
When the output mode is complete, then the output of a streaming aggregation essentially will contain the complete aggregates every time. So this is not different from a batch dataset within an incremental execution. Other non-streaming operations should be supported on this dataset. In this PR, I am just adding support for sorting, as it is a common useful functionality. Support for other operations will come later.
## How was this patch tested?
Additional unit tests.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#13549 from tdas/SPARK-15812.
## What changes were proposed in this pull request?
With very wide tables, e.g. thousands of fields, the plan output is unreadable and often causes OOMs due to inefficient string processing. This truncates all struct and operator field lists to a user configurable threshold to limit performance impact.
It would also be nice to optimize string generation to avoid these sort of O(n^2) slowdowns entirely (i.e. use StringBuilder everywhere including expressions), but this is probably too large of a change for 2.0 at this point, and truncation has other benefits for usability.
## How was this patch tested?
Added a microbenchmark that covers this case particularly well. I also ran the microbenchmark while varying the truncation threshold.
```
numFields = 5
wide shallowly nested struct field r/w: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
2000 wide x 50 rows (write in-mem) 2336 / 2558 0.0 23364.4 0.1X
numFields = 25
wide shallowly nested struct field r/w: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
2000 wide x 50 rows (write in-mem) 4237 / 4465 0.0 42367.9 0.1X
numFields = 100
wide shallowly nested struct field r/w: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
2000 wide x 50 rows (write in-mem) 10458 / 11223 0.0 104582.0 0.0X
numFields = Infinity
wide shallowly nested struct field r/w: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
[info] java.lang.OutOfMemoryError: Java heap space
```
Author: Eric Liang <ekl@databricks.com>
Author: Eric Liang <ekhliang@gmail.com>
Closes#13537 from ericl/truncated-string.
## What changes were proposed in this pull request?
The current implementations of `UnixTime` and `FromUnixTime` do not cache their parser/formatter as much as they could. This PR resolved this issue.
This PR is a take over from https://github.com/apache/spark/pull/13522 and further optimizes the re-use of the parser/formatter. It also fixes the improves handling (catching the actual exception instead of `Throwable`). All credits for this work should go to rajeshbalamohan.
This PR closes https://github.com/apache/spark/pull/13522
## How was this patch tested?
Current tests.
Author: Herman van Hovell <hvanhovell@databricks.com>
Author: Rajesh Balamohan <rbalamohan@apache.org>
Closes#13581 from hvanhovell/SPARK-14321.
## What changes were proposed in this pull request?
The help function 'toStructType' in the AttributeSeq class doesn't include the metadata when it builds the StructField, so it causes this reported problem https://issues.apache.org/jira/browse/SPARK-15804?jql=project%20%3D%20SPARK when spark writes the the dataframe with the metadata to the parquet datasource.
The code path is when spark writes the dataframe to the parquet datasource through the InsertIntoHadoopFsRelationCommand, spark will build the WriteRelation container, and it will call the help function 'toStructType' to create StructType which contains StructField, it should include the metadata there, otherwise, we will lost the user provide metadata.
## How was this patch tested?
added test case in ParquetQuerySuite.scala
(If this patch involves UI changes, please attach a screenshot; otherwise, remove this)
Author: Kevin Yu <qyu@us.ibm.com>
Closes#13555 from kevinyu98/spark-15804.
## What changes were proposed in this pull request?
The parser currently does not allow the use of some SQL keywords as table or field names. This PR adds supports for all keywords as identifier. The exception to this are table aliases, in this case most keywords are allowed except for join keywords (```anti, full, inner, left, semi, right, natural, on, join, cross```) and set-operator keywords (```union, intersect, except```).
## How was this patch tested?
I have added/move/renamed test in the catalyst `*ParserSuite`s.
Author: Herman van Hovell <hvanhovell@databricks.com>
Closes#13534 from hvanhovell/SPARK-15789.
## What changes were proposed in this pull request?
The current implementation of "CREATE TEMPORARY TABLE USING datasource..." is NOT creating any intermediate temporary data directory like temporary HDFS folder, instead, it only stores a SQL string in memory. Probably we should use "TEMPORARY VIEW" instead.
This PR assumes a temporary table has to link with some temporary intermediate data. It follows the definition of temporary table like this (from [hortonworks doc](https://docs.hortonworks.com/HDPDocuments/HDP2/HDP-2.3.0/bk_dataintegration/content/temp-tables.html)):
> A temporary table is a convenient way for an application to automatically manage intermediate data generated during a complex query
**Example**:
```
scala> spark.sql("CREATE temporary view my_tab7 (c1: String, c2: String) USING org.apache.spark.sql.execution.datasources.csv.CSVFileFormat OPTIONS (PATH '/Users/seanzhong/csv/cars.csv')")
scala> spark.sql("select c1, c2 from my_tab7").show()
+----+-----+
| c1| c2|
+----+-----+
|year| make|
|2012|Tesla|
...
```
It NOW prints a **deprecation warning** if "CREATE TEMPORARY TABLE USING..." is used.
```
scala> spark.sql("CREATE temporary table my_tab7 (c1: String, c2: String) USING org.apache.spark.sql.execution.datasources.csv.CSVFileFormat OPTIONS (PATH '/Users/seanzhong/csv/cars.csv')")
16/05/31 10:39:27 WARN SparkStrategies$DDLStrategy: CREATE TEMPORARY TABLE tableName USING... is deprecated, please use CREATE TEMPORARY VIEW viewName USING... instead
```
## How was this patch tested?
Unit test.
Author: Sean Zhong <seanzhong@databricks.com>
Closes#13414 from clockfly/create_temp_view_using.
## What changes were proposed in this pull request?
This PR allows customization of verbosity in explain output. After change, `dataframe.explain()` and `dataframe.explain(true)` has different verbosity output for physical plan.
Currently, this PR only enables verbosity string for operator `HashAggregateExec` and `SortAggregateExec`. We will gradually enable verbosity string for more operators in future.
**Less verbose mode:** dataframe.explain(extended = false)
`output=[count(a)#85L]` is **NOT** displayed for HashAggregate.
```
scala> Seq((1,2,3)).toDF("a", "b", "c").createTempView("df2")
scala> spark.sql("select count(a) from df2").explain()
== Physical Plan ==
*HashAggregate(key=[], functions=[count(1)])
+- Exchange SinglePartition
+- *HashAggregate(key=[], functions=[partial_count(1)])
+- LocalTableScan
```
**Verbose mode:** dataframe.explain(extended = true)
`output=[count(a)#85L]` is displayed for HashAggregate.
```
scala> spark.sql("select count(a) from df2").explain(true) // "output=[count(a)#85L]" is added
...
== Physical Plan ==
*HashAggregate(key=[], functions=[count(1)], output=[count(a)#85L])
+- Exchange SinglePartition
+- *HashAggregate(key=[], functions=[partial_count(1)], output=[count#87L])
+- LocalTableScan
```
## How was this patch tested?
Manual test.
Author: Sean Zhong <seanzhong@databricks.com>
Closes#13535 from clockfly/verbose_breakdown_2.
## What changes were proposed in this pull request?
This PR makes sure the typed Filter doesn't change the Dataset schema.
**Before the change:**
```
scala> val df = spark.range(0,9)
scala> df.schema
res12: org.apache.spark.sql.types.StructType = StructType(StructField(id,LongType,false))
scala> val afterFilter = df.filter(_=>true)
scala> afterFilter.schema // !!! schema is CHANGED!!! Column name is changed from id to value, nullable is changed from false to true.
res13: org.apache.spark.sql.types.StructType = StructType(StructField(value,LongType,true))
```
SerializeFromObject and DeserializeToObject are inserted to wrap the Filter, and these two can possibly change the schema of Dataset.
**After the change:**
```
scala> afterFilter.schema // schema is NOT changed.
res47: org.apache.spark.sql.types.StructType = StructType(StructField(id,LongType,false))
```
## How was this patch tested?
Unit test.
Author: Sean Zhong <seanzhong@databricks.com>
Closes#13529 from clockfly/spark-15632.
BindReferences contains a n^2 loop which causes performance issues when operating over large schemas: to determine the ordinal of an attribute reference, we perform a linear scan over the `input` array. Because input can sometimes be a `List`, the call to `input(ordinal).nullable` can also be O(n).
Instead of performing a linear scan, we can convert the input into an array and build a hash map to map from expression ids to ordinals. The greater up-front cost of the map construction is offset by the fact that an expression can contain multiple attribute references, so the cost of the map construction is amortized across a number of lookups.
Perf. benchmarks to follow. /cc ericl
Author: Josh Rosen <joshrosen@databricks.com>
Closes#13505 from JoshRosen/bind-references-improvement.
## What changes were proposed in this pull request?
`an -> a`
Use cmds like `find . -name '*.R' | xargs -i sh -c "grep -in ' an [^aeiou]' {} && echo {}"` to generate candidates, and review them one by one.
## How was this patch tested?
manual tests
Author: Zheng RuiFeng <ruifengz@foxmail.com>
Closes#13515 from zhengruifeng/an_a.
## What changes were proposed in this pull request?
This PR improves the error handling of `RowEncoder`. When we create a `RowEncoder` with a given schema, we should validate the data type of input object. e.g. we should throw an exception when a field is boolean but is declared as a string column.
This PR also removes the support to use `Product` as a valid external type of struct type. This support is added at https://github.com/apache/spark/pull/9712, but is incomplete, e.g. nested product, product in array are both not working. However, we never officially support this feature and I think it's ok to ban it.
## How was this patch tested?
new tests in `RowEncoderSuite`.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#13401 from cloud-fan/bug.
## What changes were proposed in this pull request?
In forType function of object RandomDataGenerator, the code following:
if (maybeSqlTypeGenerator.isDefined){
....
Some(generator)
} else{
None
}
will be changed. Instead, maybeSqlTypeGenerator.map will be used.
## How was this patch tested?
All of the current unit tests passed.
Author: Weiqing Yang <yangweiqing001@gmail.com>
Closes#13448 from Sherry302/master.
We should cache `Metadata.hashCode` and use a singleton for `Metadata.empty` because calculating metadata hashCodes appears to be a bottleneck for certain workloads.
We should also cache `StructType.hashCode`.
In an optimizer stress-test benchmark run by ericl, these `hashCode` calls accounted for roughly 40% of the total CPU time and this bottleneck was completely eliminated by the caching added by this patch.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#13504 from JoshRosen/metadata-fix.
## What changes were proposed in this pull request?
For input object of non-flat type, we can't encode it to row if it's null, as Spark SQL doesn't allow row to be null, only its columns can be null.
This PR explicitly add this constraint and throw exception if users break it.
## How was this patch tested?
several new tests
Author: Wenchen Fan <wenchen@databricks.com>
Closes#13469 from cloud-fan/null-object.
## What changes were proposed in this pull request?
There are 2 kinds of `GetStructField`:
1. resolved from `UnresolvedExtractValue`, and it will have a `name` property.
2. created when we build deserializer expression for nested tuple, no `name` property.
When we want to validate the ordinals of nested tuple, we should only catch `GetStructField` without the name property.
## How was this patch tested?
new test in `EncoderResolutionSuite`
Author: Wenchen Fan <wenchen@databricks.com>
Closes#13474 from cloud-fan/ordinal-check.
#### What changes were proposed in this pull request?
Before this PR, the output of EXPLAIN of following SQL is like
```SQL
CREATE EXTERNAL TABLE extTable_with_partitions (key INT, value STRING)
PARTITIONED BY (ds STRING, hr STRING)
LOCATION '/private/var/folders/4b/sgmfldk15js406vk7lw5llzw0000gn/T/spark-b39a6185-8981-403b-a4aa-36fb2f4ca8a9'
```
``ExecutedCommand CreateTableCommand CatalogTable(`extTable_with_partitions`,CatalogTableType(EXTERNAL),CatalogStorageFormat(Some(/private/var/folders/4b/sgmfldk15js406vk7lw5llzw0000gn/T/spark-dd234718-e85d-4c5a-8353-8f1834ac0323),Some(org.apache.hadoop.mapred.TextInputFormat),Some(org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat),None,false,Map()),List(CatalogColumn(key,int,true,None), CatalogColumn(value,string,true,None), CatalogColumn(ds,string,true,None), CatalogColumn(hr,string,true,None)),List(ds, hr),List(),List(),-1,,1463026413544,-1,Map(),None,None,None), false``
After this PR, the output is like
```
ExecutedCommand
: +- CreateTableCommand CatalogTable(
Table:`extTable_with_partitions`
Created:Thu Jun 02 21:30:54 PDT 2016
Last Access:Wed Dec 31 15:59:59 PST 1969
Type:EXTERNAL
Schema:[`key` int, `value` string, `ds` string, `hr` string]
Partition Columns:[`ds`, `hr`]
Storage(Location:/private/var/folders/4b/sgmfldk15js406vk7lw5llzw0000gn/T/spark-a06083b8-8e88-4d07-9ff0-d6bd8d943ad3, InputFormat:org.apache.hadoop.mapred.TextInputFormat, OutputFormat:org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat)), false
```
This is also applicable to `DESC EXTENDED`. However, this does not have special handling for Data Source Tables. If needed, we need to move the logics of `DDLUtil`. Let me know if we should do it in this PR. Thanks! rxin liancheng
#### How was this patch tested?
Manual testing
Author: gatorsmile <gatorsmile@gmail.com>
Closes#13070 from gatorsmile/betterExplainCatalogTable.
In Catalyst's TreeNode transform methods we end up calling `productIterator.map(...).toArray` in a number of places, which is slightly inefficient because it needs to allocate an `ArrayBuilder` and grow a temporary array. Since we already know the size of the final output (`productArity`), we can simply allocate an array up-front and use a while loop to consume the iterator and populate the array.
For most workloads, this performance difference is negligible but it does make a measurable difference in optimizer performance for queries that operate over very wide schemas (such as the benchmark queries in #13456).
### Perf results (from #13456 benchmarks)
**Before**
```
Java HotSpot(TM) 64-Bit Server VM 1.8.0_66-b17 on Mac OS X 10.10.5
Intel(R) Core(TM) i7-4960HQ CPU 2.60GHz
parsing large select: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
1 select expressions 19 / 22 0.0 19119858.0 1.0X
10 select expressions 23 / 25 0.0 23208774.0 0.8X
100 select expressions 55 / 73 0.0 54768402.0 0.3X
1000 select expressions 229 / 259 0.0 228606373.0 0.1X
2500 select expressions 530 / 554 0.0 529938178.0 0.0X
```
**After**
```
parsing large select: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
1 select expressions 15 / 21 0.0 14978203.0 1.0X
10 select expressions 22 / 27 0.0 22492262.0 0.7X
100 select expressions 48 / 64 0.0 48449834.0 0.3X
1000 select expressions 189 / 208 0.0 189346428.0 0.1X
2500 select expressions 429 / 449 0.0 428943897.0 0.0X
```
###
Author: Josh Rosen <joshrosen@databricks.com>
Closes#13484 from JoshRosen/treenode-productiterator-map.
## What changes were proposed in this pull request?
Queries with scalar sub-query in the SELECT list run against a local, in-memory relation throw
UnsupportedOperationException exception.
Problem repro:
```SQL
scala> Seq((1, 1), (2, 2)).toDF("c1", "c2").createOrReplaceTempView("t1")
scala> Seq((1, 1), (2, 2)).toDF("c1", "c2").createOrReplaceTempView("t2")
scala> sql("select (select min(c1) from t2) from t1").show()
java.lang.UnsupportedOperationException: Cannot evaluate expression: scalar-subquery#62 []
at org.apache.spark.sql.catalyst.expressions.Unevaluable$class.eval(Expression.scala:215)
at org.apache.spark.sql.catalyst.expressions.ScalarSubquery.eval(subquery.scala:62)
at org.apache.spark.sql.catalyst.expressions.Alias.eval(namedExpressions.scala:142)
at org.apache.spark.sql.catalyst.expressions.InterpretedProjection.apply(Projection.scala:45)
at org.apache.spark.sql.catalyst.expressions.InterpretedProjection.apply(Projection.scala:29)
at scala.collection.TraversableLike$$anonfun$map$1.apply(TraversableLike.scala:234)
at scala.collection.TraversableLike$$anonfun$map$1.apply(TraversableLike.scala:234)
at scala.collection.immutable.List.foreach(List.scala:381)
at scala.collection.TraversableLike$class.map(TraversableLike.scala:234)
at scala.collection.immutable.List.map(List.scala:285)
at org.apache.spark.sql.catalyst.optimizer.ConvertToLocalRelation$$anonfun$apply$37.applyOrElse(Optimizer.scala:1473)
```
The problem is specific to local, in memory relations. It is caused by rule ConvertToLocalRelation, which attempts to push down
a scalar-subquery expression to the local tables.
The solution prevents the rule to apply if Project references scalar subqueries.
## How was this patch tested?
Added regression tests to SubquerySuite.scala
Author: Ioana Delaney <ioanamdelaney@gmail.com>
Closes#13418 from ioana-delaney/scalarSubV2.
## What changes were proposed in this pull request?
Our encoder framework has been evolved a lot, this PR tries to clean up the code to make it more readable and emphasise the concept that encoder should be used as a container of serde expressions.
1. move validation logic to analyzer instead of encoder
2. only have a `resolveAndBind` method in encoder instead of `resolve` and `bind`, as we don't have the encoder life cycle concept anymore.
3. `Dataset` don't need to keep a resolved encoder, as there is no such concept anymore. bound encoder is still needed to do serialization outside of query framework.
4. Using `BoundReference` to represent an unresolved field in deserializer expression is kind of weird, this PR adds a `GetColumnByOrdinal` for this purpose. (serializer expression still use `BoundReference`, we can replace it with `GetColumnByOrdinal` in follow-ups)
## How was this patch tested?
existing test
Author: Wenchen Fan <wenchen@databricks.com>
Author: Cheng Lian <lian@databricks.com>
Closes#13269 from cloud-fan/clean-encoder.
## What changes were proposed in this pull request?
This PR makes the explain output less verbose by hiding some verbose output like `None`, `null`, empty List `[]`, empty set `{}`, and etc.
**Before change**:
```
== Physical Plan ==
ExecutedCommand
: +- ShowTablesCommand None, None
```
**After change**:
```
== Physical Plan ==
ExecutedCommand
: +- ShowTablesCommand
```
## How was this patch tested?
Manual test.
Author: Sean Zhong <seanzhong@databricks.com>
Closes#13470 from clockfly/verbose_breakdown_4.
## What changes were proposed in this pull request?
When users create a case class and use java reserved keyword as field name, spark sql will generate illegal java code and throw exception at runtime.
This PR checks the field names when building the encoder, and if illegal field names are used, throw exception immediately with a good error message.
## How was this patch tested?
new test in DatasetSuite
Author: Wenchen Fan <wenchen@databricks.com>
Closes#13485 from cloud-fan/java.
## What changes were proposed in this pull request?
This command didn't work for Hive tables. Now it does:
```
ALTER TABLE boxes PARTITION (width=3)
SET SERDE 'com.sparkbricks.serde.ColumnarSerDe'
WITH SERDEPROPERTIES ('compress'='true')
```
## How was this patch tested?
`HiveExternalCatalogSuite`
Author: Andrew Or <andrew@databricks.com>
Closes#13453 from andrewor14/alter-partition-storage.
## What changes were proposed in this pull request?
This patch fixes a number of `com.esotericsoftware.kryo.KryoException: java.lang.NullPointerException` exceptions reported in [SPARK-15604], [SPARK-14752] etc. (while executing sparkSQL queries with the kryo serializer) by explicitly implementing `KryoSerialization` for `LazilyGenerateOrdering`.
## How was this patch tested?
1. Modified `OrderingSuite` so that all tests in the suite also test kryo serialization (for both interpreted and generated ordering).
2. Manually verified TPC-DS q1.
Author: Sameer Agarwal <sameer@databricks.com>
Closes#13466 from sameeragarwal/kryo.
## What changes were proposed in this pull request?
This PR add a rule at the end of analyzer to correct nullable fields of attributes in a logical plan by using nullable fields of the corresponding attributes in its children logical plans (these plans generate the input rows).
This is another approach for addressing SPARK-13484 (the first approach is https://github.com/apache/spark/pull/11371).
Close#113711
Author: Takeshi YAMAMURO <linguin.m.s@gmail.com>
Author: Yin Huai <yhuai@databricks.com>
Closes#13290 from yhuai/SPARK-13484.
## What changes were proposed in this pull request?
Join on transformed dataset has attributes conflicts, which make query execution failure, for example:
```
val dataset = Seq(1, 2, 3).toDs
val mappedDs = dataset.map(_ + 1)
mappedDs.as("t1").joinWith(mappedDs.as("t2"), $"t1.value" === $"t2.value").show()
```
will throw exception:
```
org.apache.spark.sql.AnalysisException: cannot resolve '`t1.value`' given input columns: [value];
at org.apache.spark.sql.catalyst.analysis.package$AnalysisErrorAt.failAnalysis(package.scala:42)
at org.apache.spark.sql.catalyst.analysis.CheckAnalysis$$anonfun$checkAnalysis$1$$anonfun$apply$2.applyOrElse(CheckAnalysis.scala:62)
at org.apache.spark.sql.catalyst.analysis.CheckAnalysis$$anonfun$checkAnalysis$1$$anonfun$apply$2.applyOrElse(CheckAnalysis.scala:59)
at org.apache.spark.sql.catalyst.trees.TreeNode$$anonfun$transformUp$1.apply(TreeNode.scala:287)
at org.apache.spark.sql.catalyst.trees.TreeNode$$anonfun$transformUp$1.apply(TreeNode.scala:287)
```
## How was this patch tested?
Unit test.
Author: jerryshao <sshao@hortonworks.com>
Closes#13399 from jerryshao/SPARK-15620.
## What changes were proposed in this pull request?
Improves the explain output of several physical plans by displaying embedded logical plan in tree style
Some physical plan contains a embedded logical plan, for example, `cache tableName query` maps to:
```
case class CacheTableCommand(
tableName: String,
plan: Option[LogicalPlan],
isLazy: Boolean)
extends RunnableCommand
```
It is easier to read the explain output if we can display the `plan` in tree style.
**Before change:**
Everything is messed in one line.
```
scala> Seq((1,2)).toDF().createOrReplaceTempView("testView")
scala> spark.sql("cache table testView2 select * from testView").explain()
== Physical Plan ==
ExecutedCommand CacheTableCommand testView2, Some('Project [*]
+- 'UnresolvedRelation `testView`, None
), false
```
**After change:**
```
scala> spark.sql("cache table testView2 select * from testView").explain()
== Physical Plan ==
ExecutedCommand
: +- CacheTableCommand testView2, false
: : +- 'Project [*]
: : +- 'UnresolvedRelation `testView`, None
```
## How was this patch tested?
Manual test.
Author: Sean Zhong <seanzhong@databricks.com>
Closes#13433 from clockfly/verbose_breakdown_3_2.
## What changes were proposed in this pull request?
Currently we can't encode top level null object into internal row, as Spark SQL doesn't allow row to be null, only its columns can be null.
This is not a problem before, as we assume the input object is never null. However, for outer join, we do need the semantics of null object.
This PR fixes this problem by making both join sides produce a single column, i.e. nest the logical plan output(by `CreateStruct`), so that we have an extra level to represent top level null obejct.
## How was this patch tested?
new test in `DatasetSuite`
Author: Wenchen Fan <wenchen@databricks.com>
Closes#13425 from cloud-fan/outer-join2.
This PR is an alternative to #13120 authored by xwu0226.
## What changes were proposed in this pull request?
When creating an external Spark SQL data source table and persisting its metadata to Hive metastore, we don't use the standard Hive `Table.dataLocation` field because Hive only allows directory paths as data locations while Spark SQL also allows file paths. However, if we don't set `Table.dataLocation`, Hive always creates an unexpected empty table directory under database location, but doesn't remove it while dropping the table (because the table is external).
This PR works around this issue by explicitly setting `Table.dataLocation` and then manullay removing the created directory after creating the external table.
Please refer to [this JIRA comment][1] for more details about why we chose this approach as a workaround.
[1]: https://issues.apache.org/jira/browse/SPARK-15269?focusedCommentId=15297408&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-15297408
## How was this patch tested?
1. A new test case is added in `HiveQuerySuite` for this case
2. Updated `ShowCreateTableSuite` to use the same table name in all test cases. (This is how I hit this issue at the first place.)
Author: Cheng Lian <lian@databricks.com>
Closes#13270 from liancheng/spark-15269-unpleasant-fix.
## What changes were proposed in this pull request?
This patch moves all user-facing structured streaming classes into sql.streaming. As part of this, I also added some since version annotation to methods and classes that don't have them.
## How was this patch tested?
Updated tests to reflect the moves.
Author: Reynold Xin <rxin@databricks.com>
Closes#13429 from rxin/SPARK-15686.
## What changes were proposed in this pull request?
Currently `spark.sql.warehouse.dir` is pointed to local dir by default, which will throw exception when HADOOP_CONF_DIR is configured and default FS is hdfs.
```
java.lang.IllegalArgumentException: Wrong FS: file:/Users/sshao/projects/apache-spark/spark-warehouse, expected: hdfs://localhost:8020
```
So we should always get the `FileSystem` from `Path` to avoid wrong FS problem.
## How was this patch tested?
Local test.
Author: jerryshao <sshao@hortonworks.com>
Closes#13405 from jerryshao/SPARK-15659.
## What changes were proposed in this pull request?
In benchmarks involving tables with very wide and complex schemas (thousands of columns, deep nesting), I noticed that significant amounts of time (order of tens of seconds per task) were being spent generating comments during the code generation phase.
The root cause of the performance problem stems from the fact that calling toString() on a complex expression can involve thousands of string concatenations, resulting in huge amounts (tens of gigabytes) of character array allocation and copying.
In the long term, we can avoid this problem by passing StringBuilders down the tree and using them to accumulate output. As a short-term workaround, this patch guards comment generation behind a flag and disables comments by default (for wide tables / complex queries, these comments were being truncated prior to display and thus were not very useful).
## How was this patch tested?
This was tested manually by running a Spark SQL query over an empty table with a very wide schema obtained from a real workload. Disabling comments brought the per-task time down from about 16 seconds to 600 milliseconds.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#13421 from JoshRosen/disable-line-comments-in-codegen.
## What changes were proposed in this pull request?
Currently structured streaming only supports append output mode. This PR adds the following.
- Added support for Complete output mode in the internal state store, analyzer and planner.
- Added public API in Scala and Python for users to specify output mode
- Added checks for unsupported combinations of output mode and DF operations
- Plans with no aggregation should support only Append mode
- Plans with aggregation should support only Update and Complete modes
- Default output mode is Append mode (**Question: should we change this to automatically set to Complete mode when there is aggregation?**)
- Added support for Complete output mode in Memory Sink. So Memory Sink internally supports append and complete, update. But from public API only Complete and Append output modes are supported.
## How was this patch tested?
Unit tests in various test suites
- StreamingAggregationSuite: tests for complete mode
- MemorySinkSuite: tests for checking behavior in Append and Complete modes.
- UnsupportedOperationSuite: tests for checking unsupported combinations of DF ops and output modes
- DataFrameReaderWriterSuite: tests for checking that output mode cannot be called on static DFs
- Python doc test and existing unit tests modified to call write.outputMode.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#13286 from tdas/complete-mode.
In this case, the result type of the expression becomes DECIMAL(38, 36) as we promote the individual string literals to DECIMAL(38, 18) when we handle string promotions for `BinaryArthmaticExpression`.
I think we need to cast the string literals to Double type instead. I looked at the history and found that this was changed to use decimal instead of double to avoid potential loss of precision when we cast decimal to double.
To double check i ran the query against hive, mysql. This query returns non NULL result for both the databases and both promote the expression to use double.
Here is the output.
- Hive
```SQL
hive> create table l2 as select (cast(99 as decimal(19,6)) + '2') from l1;
OK
hive> describe l2;
OK
_c0 double
```
- MySQL
```SQL
mysql> create table foo2 as select (cast(99 as decimal(19,6)) + '2') from test;
Query OK, 1 row affected (0.01 sec)
Records: 1 Duplicates: 0 Warnings: 0
mysql> describe foo2;
+-----------------------------------+--------+------+-----+---------+-------+
| Field | Type | Null | Key | Default | Extra |
+-----------------------------------+--------+------+-----+---------+-------+
| (cast(99 as decimal(19,6)) + '2') | double | NO | | 0 | |
+-----------------------------------+--------+------+-----+---------+-------+
```
## How was this patch tested?
Added a new test in SQLQuerySuite
Author: Dilip Biswal <dbiswal@us.ibm.com>
Closes#13368 from dilipbiswal/spark-15557.
## What changes were proposed in this pull request?
Right now, we will split the code for expressions into multiple functions when it exceed 64k, which requires that the the expressions are using Row object, but this is not true for whole-state codegen, it will fail to compile after splitted.
This PR will not split the code in whole-stage codegen.
## How was this patch tested?
Added regression tests.
Author: Davies Liu <davies@databricks.com>
Closes#13235 from davies/fix_nested_codegen.