## What changes were proposed in this pull request?
Instead of iterating all expressions that can return null for null inputs, we can just check `NullIntolerant`.
## How was this patch tested?
existing tests
Author: Wenchen Fan <wenchen@databricks.com>
Closes#17450 from cloud-fan/null.
…adoc
## What changes were proposed in this pull request?
Use recommended values for row boundaries in Window's scaladoc, i.e. `Window.unboundedPreceding`, `Window.unboundedFollowing`, and `Window.currentRow` (that were introduced in 2.1.0).
## How was this patch tested?
Local build
Author: Jacek Laskowski <jacek@japila.pl>
Closes#17417 from jaceklaskowski/window-expression-scaladoc.
## What changes were proposed in this pull request?
Commit 6c70a38 broke the build for scala 2.10. The commit uses some reflections which are not available in Scala 2.10. This PR fixes them.
## How was this patch tested?
Existing tests.
Author: Takuya UESHIN <ueshin@databricks.com>
Closes#17473 from ueshin/issues/SPARK-19088.
### What changes were proposed in this pull request?
`FalseLiteral` and `TrueLiteral` should have been eliminated by optimizer rule `BooleanSimplification`, but null literals might be added by optimizer rule `NullPropagation`. For safety, our filter estimation should handle all the eligible literal cases.
Our optimizer rule BooleanSimplification is unable to remove the null literal in many cases. For example, `a < 0 or null`. Thus, we need to handle null literal in filter estimation.
`Not` can be pushed down below `And` and `Or`. Then, we could see two consecutive `Not`, which need to be collapsed into one. Because of the limited expression support for filter estimation, we just need to handle the case `Not(null)` for avoiding incorrect error due to the boolean operation on null. For details, see below matrix.
```
not NULL = NULL
NULL or false = NULL
NULL or true = true
NULL or NULL = NULL
NULL and false = false
NULL and true = NULL
NULL and NULL = NULL
```
### How was this patch tested?
Added the test cases.
Author: Xiao Li <gatorsmile@gmail.com>
Closes#17446 from gatorsmile/constantFilterEstimation.
## What changes were proposed in this pull request?
This pr added `StructType.fromDDL` to convert a DDL format string into `StructType` for defining schemas in `functions.from_json`.
## How was this patch tested?
Added tests in `JsonFunctionsSuite`.
Author: Takeshi Yamamuro <yamamuro@apache.org>
Closes#17406 from maropu/SPARK-20009.
## What changes were proposed in this pull request?
When we build the deserializer expression for map type, we will use `StaticInvoke` to call `ArrayBasedMapData.toScalaMap`, and declare the return type as `scala.collection.immutable.Map`. If the map is inside an Option, we will wrap this `StaticInvoke` with `WrapOption`, which requires the input to be `scala.collect.Map`. Ideally this should be fine, as `scala.collection.immutable.Map` extends `scala.collect.Map`, but our `ObjectType` is too strict about this, this PR fixes it.
## How was this patch tested?
new regression test
Author: Wenchen Fan <wenchen@databricks.com>
Closes#17454 from cloud-fan/map.
## What changes were proposed in this pull request?
Join reorder algorithm should keep exactly the same order of output attributes in the top project.
For example, if user want to select a, b, c, after reordering, we should output a, b, c in the same order as specified by user, instead of b, a, c or other orders.
## How was this patch tested?
A new test case is added in `JoinReorderSuite`.
Author: wangzhenhua <wangzhenhua@huawei.com>
Closes#17453 from wzhfy/keepOrderInProject.
## What changes were proposed in this pull request?
TPCDS q45 fails becuase:
`ReorderJoin` collects all predicates and try to put them into join condition when creating ordered join. If a predicate with an IN subquery (`ListQuery`) is in a join condition instead of a filter condition, `RewritePredicateSubquery.rewriteExistentialExpr` would fail to convert the subquery to an `ExistenceJoin`, and thus result in error.
We should prevent push down of IN subquery to Join operator.
## How was this patch tested?
Add a new test case in `FilterPushdownSuite`.
Author: wangzhenhua <wangzhenhua@huawei.com>
Closes#17428 from wzhfy/noSubqueryInJoinCond.
## What changes were proposed in this pull request?
The current SessionState initialization code path is quite complex. A part of the creation is done in the SessionState companion objects, a part of the creation is one inside the SessionState class, and a part is done by passing functions.
This PR refactors this code path, and consolidates SessionState initialization into a builder class. This SessionState will not do any initialization and just becomes a place holder for the various Spark SQL internals. This also lays the ground work for two future improvements:
1. This provides us with a start for removing the `HiveSessionState`. Removing the `HiveSessionState` would also require us to move resource loading into a separate class, and to (re)move metadata hive.
2. This makes it easier to customize the Spark Session. Currently you will need to create a custom version of the builder. I have added hooks to facilitate this. A future step will be to create a semi stable API on top of this.
## How was this patch tested?
Existing tests.
Author: Herman van Hovell <hvanhovell@databricks.com>
Closes#17433 from hvanhovell/SPARK-20100.
## What changes were proposed in this pull request?
In current stage, we don't have advanced statistics such as sketches or histograms. As a result, some operator can't estimate `nullCount` accurately. E.g. left outer join estimation does not accurately update `nullCount` currently. So for `IsNull` and `IsNotNull` predicates, we only estimate them when the child is a leaf node, whose `nullCount` is accurate.
## How was this patch tested?
A new test case is added in `FilterEstimationSuite`.
Author: wangzhenhua <wangzhenhua@huawei.com>
Closes#17438 from wzhfy/nullEstimation.
## What changes were proposed in this pull request?
The `CollapseWindow` is currently to aggressive when collapsing adjacent windows. It also collapses windows in the which the parent produces a column that is consumed by the child; this creates an invalid window which will fail at runtime.
This PR fixes this by adding a check for dependent adjacent windows to the `CollapseWindow` rule.
## How was this patch tested?
Added a new test case to `CollapseWindowSuite`
Author: Herman van Hovell <hvanhovell@databricks.com>
Closes#17432 from hvanhovell/SPARK-20086.
## What changes were proposed in this pull request?
The `FailureSafeParser` is only used in sql core, it doesn't make sense to put it in catalyst module.
## How was this patch tested?
N/A
Author: Wenchen Fan <wenchen@databricks.com>
Closes#17408 from cloud-fan/minor.
## What changes were proposed in this pull request?
Constraint propagation can be computation expensive and block the driver execution for long time. For example, the below benchmark needs 30mins.
Compared with previous PRs #16998, #16785, this is a much simpler option: add a flag to disable constraint propagation.
### Benchmark
Run the following codes locally.
import org.apache.spark.ml.{Pipeline, PipelineStage}
import org.apache.spark.ml.feature.{OneHotEncoder, StringIndexer, VectorAssembler}
import org.apache.spark.sql.internal.SQLConf
spark.conf.set(SQLConf.CONSTRAINT_PROPAGATION_ENABLED.key, false)
val df = (1 to 40).foldLeft(Seq((1, "foo"), (2, "bar"), (3, "baz")).toDF("id", "x0"))((df, i) => df.withColumn(s"x$i", $"x0"))
val indexers = df.columns.tail.map(c => new StringIndexer()
.setInputCol(c)
.setOutputCol(s"${c}_indexed")
.setHandleInvalid("skip"))
val encoders = indexers.map(indexer => new OneHotEncoder()
.setInputCol(indexer.getOutputCol)
.setOutputCol(s"${indexer.getOutputCol}_encoded")
.setDropLast(true))
val stages: Array[PipelineStage] = indexers ++ encoders
val pipeline = new Pipeline().setStages(stages)
val startTime = System.nanoTime
pipeline.fit(df).transform(df).show
val runningTime = System.nanoTime - startTime
Before this patch: 1786001 ms ~= 30 mins
After this patch: 26392 ms = less than half of a minute
Related PRs: #16998, #16785.
## How was this patch tested?
Jenkins tests.
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes#17186 from viirya/add-flag-disable-constraint-propagation.
## What changes were proposed in this pull request?
Fix for typo in Analyzer
## How was this patch tested?
local build
Author: Jacek Laskowski <jacek@japila.pl>
Closes#17409 from jaceklaskowski/analyzer-typo.
## What changes were proposed in this pull request?
Since the state is tied a "group" in the "mapGroupsWithState" operations, its better to call the state "GroupState" instead of a key. This would make it more general if you extends this operation to RelationGroupedDataset and python APIs.
## How was this patch tested?
Existing unit tests.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#17385 from tdas/SPARK-20057.
## What changes were proposed in this pull request?
Currently, when we perform count with timestamp types, it prints the internal representation as the column name as below:
```scala
Seq(new java.sql.Timestamp(1)).toDF("a").groupBy("a").pivot("a").count().show()
```
```
+--------------------+----+
| a|1000|
+--------------------+----+
|1969-12-31 16:00:...| 1|
+--------------------+----+
```
This PR proposes to use external Scala value instead of the internal representation in the column names as below:
```
+--------------------+-----------------------+
| a|1969-12-31 16:00:00.001|
+--------------------+-----------------------+
|1969-12-31 16:00:...| 1|
+--------------------+-----------------------+
```
## How was this patch tested?
Unit test in `DataFramePivotSuite` and manual tests.
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#17348 from HyukjinKwon/SPARK-20018.
## What changes were proposed in this pull request?
This PR proposes to make `mode` options in both CSV and JSON to use `cass object` and fix some related comments related previous fix.
Also, this PR modifies some tests related parse modes.
## How was this patch tested?
Modified unit tests in both `CSVSuite.scala` and `JsonSuite.scala`.
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#17377 from HyukjinKwon/SPARK-19949.
## What changes were proposed in this pull request?
Adding event time based timeout. The user sets the timeout timestamp directly using `KeyedState.setTimeoutTimestamp`. The keys times out when the watermark crosses the timeout timestamp.
## How was this patch tested?
Unit tests
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#17361 from tdas/SPARK-20030.
## What changes were proposed in this pull request?
Change the nullability of function `StringToMap` from `false` to `true`.
Author: zhaorongsheng <334362872@qq.com>
Closes#17350 from zhaorongsheng/bug-fix_strToMap_NPE.
## What changes were proposed in this pull request?
Support` ALTER TABLE ADD COLUMNS (...) `syntax for Hive serde and some datasource tables.
In this PR, we consider a few aspects:
1. View is not supported for `ALTER ADD COLUMNS`
2. Since tables created in SparkSQL with Hive DDL syntax will populate table properties with schema information, we need make sure the consistency of the schema before and after ALTER operation in order for future use.
3. For embedded-schema type of format, such as `parquet`, we need to make sure that the predicate on the newly-added columns can be evaluated properly, or pushed down properly. In case of the data file does not have the columns for the newly-added columns, such predicates should return as if the column values are NULLs.
4. For datasource table, this feature does not support the following:
4.1 TEXT format, since there is only one default column `value` is inferred for text format data.
4.2 ORC format, since SparkSQL native ORC reader does not support the difference between user-specified-schema and inferred schema from ORC files.
4.3 Third party datasource types that implements RelationProvider, including the built-in JDBC format, since different implementations by the vendors may have different ways to dealing with schema.
4.4 Other datasource types, such as `parquet`, `json`, `csv`, `hive` are supported.
5. Column names being added can not be duplicate of any existing data column or partition column names. Case sensitivity is taken into consideration according to the sql configuration.
6. This feature also supports In-Memory catalog, while Hive support is turned off.
## How was this patch tested?
Add new test cases
Author: Xin Wu <xinwu@us.ibm.com>
Closes#16626 from xwu0226/alter_add_columns.
## What changes were proposed in this pull request?
1. Improve documentation for class `Cost` and `JoinReorderDP` and method `buildJoin()`.
2. Change code structure of `buildJoin()` to make the logic clearer.
3. Add a debug-level log to record information for join reordering, including time cost, the number of items and the number of plans in memo.
## How was this patch tested?
Not related.
Author: wangzhenhua <wangzhenhua@huawei.com>
Closes#17353 from wzhfy/reorderFollow.
### What changes were proposed in this pull request?
SessionCatalog API setCurrentDatabase does not set the current database of the underlying ExternalCatalog. Thus, weird errors could come in the test suites after we call reset. We need to fix it.
So far, have not found the direct impact in the other code paths because we expect all the SessionCatalog APIs should always use the current database value we managed, unless some of code paths skip it. Thus, we fix it in the test-only function reset().
### How was this patch tested?
Multiple test case failures are observed in mvn and add a test case in SessionCatalogSuite.
Author: Xiao Li <gatorsmile@gmail.com>
Closes#17354 from gatorsmile/useDB.
## What changes were proposed in this pull request?
Currently JSON and CSV have exactly the same logic about handling bad records, this PR tries to abstract it and put it in a upper level to reduce code duplication.
The overall idea is, we make the JSON and CSV parser to throw a BadRecordException, then the upper level, FailureSafeParser, handles bad records according to the parse mode.
Behavior changes:
1. with PERMISSIVE mode, if the number of tokens doesn't match the schema, previously CSV parser will treat it as a legal record and parse as many tokens as possible. After this PR, we treat it as an illegal record, and put the raw record string in a special column, but we still parse as many tokens as possible.
2. all logging is removed as they are not very useful in practice.
## How was this patch tested?
existing tests
Author: Wenchen Fan <wenchen@databricks.com>
Author: hyukjinkwon <gurwls223@gmail.com>
Author: Wenchen Fan <cloud0fan@gmail.com>
Closes#17315 from cloud-fan/bad-record2.
## What changes were proposed in this pull request?
A Bean serializer in `ExpressionEncoder` could change values when Beans having NULL. A concrete example is as follows;
```
scala> :paste
class Outer extends Serializable {
private var cls: Inner = _
def setCls(c: Inner): Unit = cls = c
def getCls(): Inner = cls
}
class Inner extends Serializable {
private var str: String = _
def setStr(s: String): Unit = str = str
def getStr(): String = str
}
scala> Seq("""{"cls":null}""", """{"cls": {"str":null}}""").toDF().write.text("data")
scala> val encoder = Encoders.bean(classOf[Outer])
scala> val schema = encoder.schema
scala> val df = spark.read.schema(schema).json("data").as[Outer](encoder)
scala> df.show
+------+
| cls|
+------+
|[null]|
| null|
+------+
scala> df.map(x => x)(encoder).show()
+------+
| cls|
+------+
|[null]|
|[null]| // <-- Value changed
+------+
```
This is because the Bean serializer does not have the NULL-check expressions that the serializer of Scala's product types has. Actually, this value change does not happen in Scala's product types;
```
scala> :paste
case class Outer(cls: Inner)
case class Inner(str: String)
scala> val encoder = Encoders.product[Outer]
scala> val schema = encoder.schema
scala> val df = spark.read.schema(schema).json("data").as[Outer](encoder)
scala> df.show
+------+
| cls|
+------+
|[null]|
| null|
+------+
scala> df.map(x => x)(encoder).show()
+------+
| cls|
+------+
|[null]|
| null|
+------+
```
This pr added the NULL-check expressions in Bean serializer along with the serializer of Scala's product types.
## How was this patch tested?
Added tests in `JavaDatasetSuite`.
Author: Takeshi Yamamuro <yamamuro@apache.org>
Closes#17347 from maropu/SPARK-19980.
## What changes were proposed in this pull request?
After sort merge join for inner join, now we only keep left key ordering. However, after inner join, right key has the same value and order as left key. So if we need another smj on right key, we will unnecessarily add a sort which causes additional cost.
As a more complicated example, A join B on A.key = B.key join C on B.key = C.key join D on A.key = D.key. We will unnecessarily add a sort on B.key when join {A, B} and C, and add a sort on A.key when join {A, B, C} and D.
To fix this, we need to propagate all sorted information (equivalent expressions) from bottom up through `outputOrdering` and `SortOrder`.
## How was this patch tested?
Test cases are added.
Author: wangzhenhua <wangzhenhua@huawei.com>
Closes#17339 from wzhfy/sortEnhance.
## What changes were proposed in this pull request?
update `StatFunctions.multipleApproxQuantiles` to handle NaN/null
## How was this patch tested?
existing tests and added tests
Author: Zheng RuiFeng <ruifengz@foxmail.com>
Closes#16971 from zhengruifeng/quantiles_nan.
## What changes were proposed in this pull request?
Star schema consists of one or more fact tables referencing a number of dimension tables. In general, queries against star schema are expected to run fast because of the established RI constraints among the tables. This design proposes a join reordering based on natural, generally accepted heuristics for star schema queries:
- Finds the star join with the largest fact table and places it on the driving arm of the left-deep join. This plan avoids large tables on the inner, and thus favors hash joins.
- Applies the most selective dimensions early in the plan to reduce the amount of data flow.
The design document was included in SPARK-17791.
Link to the google doc: [StarSchemaDetection](https://docs.google.com/document/d/1UAfwbm_A6wo7goHlVZfYK99pqDMEZUumi7pubJXETEA/edit?usp=sharing)
## How was this patch tested?
A new test suite StarJoinSuite.scala was implemented.
Author: Ioana Delaney <ioanamdelaney@gmail.com>
Closes#15363 from ioana-delaney/starJoinReord2.
## What changes were proposed in this pull request?
This PR proposes to support an array of struct type in `to_json` as below:
```scala
import org.apache.spark.sql.functions._
val df = Seq(Tuple1(Tuple1(1) :: Nil)).toDF("a")
df.select(to_json($"a").as("json")).show()
```
```
+----------+
| json|
+----------+
|[{"_1":1}]|
+----------+
```
Currently, it throws an exception as below (a newline manually inserted for readability):
```
org.apache.spark.sql.AnalysisException: cannot resolve 'structtojson(`array`)' due to data type
mismatch: structtojson requires that the expression is a struct expression.;;
```
This allows the roundtrip with `from_json` as below:
```scala
import org.apache.spark.sql.functions._
import org.apache.spark.sql.types._
val schema = ArrayType(StructType(StructField("a", IntegerType) :: Nil))
val df = Seq("""[{"a":1}, {"a":2}]""").toDF("json").select(from_json($"json", schema).as("array"))
df.show()
// Read back.
df.select(to_json($"array").as("json")).show()
```
```
+----------+
| array|
+----------+
|[[1], [2]]|
+----------+
+-----------------+
| json|
+-----------------+
|[{"a":1},{"a":2}]|
+-----------------+
```
Also, this PR proposes to rename from `StructToJson` to `StructsToJson ` and `JsonToStruct` to `JsonToStructs`.
## How was this patch tested?
Unit tests in `JsonFunctionsSuite` and `JsonExpressionsSuite` for Scala, doctest for Python and test in `test_sparkSQL.R` for R.
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#17192 from HyukjinKwon/SPARK-19849.
## What changes were proposed in this pull request?
When a key does not get any new data in `mapGroupsWithState`, the mapping function is never called on it. So we need a timeout feature that calls the function again in such cases, so that the user can decide whether to continue waiting or clean up (remove state, save stuff externally, etc.).
Timeouts can be either based on processing time or event time. This JIRA is for processing time, but defines the high level API design for both. The usage would look like this.
```
def stateFunction(key: K, value: Iterator[V], state: KeyedState[S]): U = {
...
state.setTimeoutDuration(10000)
...
}
dataset // type is Dataset[T]
.groupByKey[K](keyingFunc) // generates KeyValueGroupedDataset[K, T]
.mapGroupsWithState[S, U](
func = stateFunction,
timeout = KeyedStateTimeout.withProcessingTime) // returns Dataset[U]
```
Note the following design aspects.
- The timeout type is provided as a param in mapGroupsWithState as a parameter global to all the keys. This is so that the planner knows this at planning time, and accordingly optimize the execution based on whether to saves extra info in state or not (e.g. timeout durations or timestamps).
- The exact timeout duration is provided inside the function call so that it can be customized on a per key basis.
- When the timeout occurs for a key, the function is called with no values, and KeyedState.isTimingOut() set to true.
- The timeout is reset for key every time the function is called on the key, that is, when the key has new data, or the key has timed out. So the user has to set the timeout duration everytime the function is called, otherwise there will not be any timeout set.
Guarantees provided on timeout of key, when timeout duration is D ms:
- Timeout will never be called before real clock time has advanced by D ms
- Timeout will be called eventually when there is a trigger with any data in it (i.e. after D ms). So there is a no strict upper bound on when the timeout would occur. For example, if there is no data in the stream (for any key) for a while, then the timeout will not be hit.
Implementation details:
- Added new param to `mapGroupsWithState` for timeout
- Added new method to `StateStore` to filter data based on timeout timestamp
- Changed the internal map type of `HDFSBackedStateStore` from Java's `HashMap` to `ConcurrentHashMap` as the latter allows weakly-consistent fail-safe iterators on the map data. See comments in code for more details.
- Refactored logic of `MapGroupsWithStateExec` to
- Save timeout info to state store for each key that has data.
- Then, filter states that should be timed out based on the current batch processing timestamp.
- Moved KeyedState for `o.a.s.sql` to `o.a.s.sql.streaming`. I remember that this was a feedback in the MapGroupsWithState PR that I had forgotten to address.
## How was this patch tested?
New unit tests in
- MapGroupsWithStateSuite for timeouts.
- StateStoreSuite for new APIs in StateStore.
Author: Tathagata Das <tathagata.das1565@gmail.com>
Closes#17179 from tdas/mapgroupwithstate-timeout.
## What changes were proposed in this pull request?
If case classes have circular references below, it throws StackOverflowError;
```
scala> :pasge
case class classA(i: Int, cls: classB)
case class classB(cls: classA)
scala> Seq(classA(0, null)).toDS()
java.lang.StackOverflowError
at scala.reflect.internal.Symbols$Symbol.info(Symbols.scala:1494)
at scala.reflect.runtime.JavaMirrors$JavaMirror$$anon$1.scala$reflect$runtime$SynchronizedSymbols$SynchronizedSymbol$$super$info(JavaMirrors.scala:66)
at scala.reflect.runtime.SynchronizedSymbols$SynchronizedSymbol$$anonfun$info$1.apply(SynchronizedSymbols.scala:127)
at scala.reflect.runtime.SynchronizedSymbols$SynchronizedSymbol$$anonfun$info$1.apply(SynchronizedSymbols.scala:127)
at scala.reflect.runtime.Gil$class.gilSynchronized(Gil.scala:19)
at scala.reflect.runtime.JavaUniverse.gilSynchronized(JavaUniverse.scala:16)
at scala.reflect.runtime.SynchronizedSymbols$SynchronizedSymbol$class.gilSynchronizedIfNotThreadsafe(SynchronizedSymbols.scala:123)
at scala.reflect.runtime.JavaMirrors$JavaMirror$$anon$1.gilSynchronizedIfNotThreadsafe(JavaMirrors.scala:66)
at scala.reflect.runtime.SynchronizedSymbols$SynchronizedSymbol$class.info(SynchronizedSymbols.scala:127)
at scala.reflect.runtime.JavaMirrors$JavaMirror$$anon$1.info(JavaMirrors.scala:66)
at scala.reflect.internal.Mirrors$RootsBase.getModuleOrClass(Mirrors.scala:48)
at scala.reflect.internal.Mirrors$RootsBase.getModuleOrClass(Mirrors.scala:45)
at scala.reflect.internal.Mirrors$RootsBase.getModuleOrClass(Mirrors.scala:45)
at scala.reflect.internal.Mirrors$RootsBase.getModuleOrClass(Mirrors.scala:45)
at scala.reflect.internal.Mirrors$RootsBase.getModuleOrClass(Mirrors.scala:45)
```
This pr added code to throw UnsupportedOperationException in that case as follows;
```
scala> :paste
case class A(cls: B)
case class B(cls: A)
scala> Seq(A(null)).toDS()
java.lang.UnsupportedOperationException: cannot have circular references in class, but got the circular reference of class B
at org.apache.spark.sql.catalyst.ScalaReflection$.org$apache$spark$sql$catalyst$ScalaReflection$$serializerFor(ScalaReflection.scala:627)
at org.apache.spark.sql.catalyst.ScalaReflection$$anonfun$9.apply(ScalaReflection.scala:644)
at org.apache.spark.sql.catalyst.ScalaReflection$$anonfun$9.apply(ScalaReflection.scala:632)
at scala.collection.TraversableLike$$anonfun$flatMap$1.apply(TraversableLike.scala:241)
at scala.collection.TraversableLike$$anonfun$flatMap$1.apply(TraversableLike.scala:241)
at scala.collection.immutable.List.foreach(List.scala:381)
at scala.collection.TraversableLike$class.flatMap(TraversableLike.scala:241)
```
## How was this patch tested?
Added tests in `DatasetSuite`.
Author: Takeshi Yamamuro <yamamuro@apache.org>
Closes#17318 from maropu/SPARK-19896.
## What changes were proposed in this pull request?
We have some concerns about removing size in the cost model [in the previous pr](https://github.com/apache/spark/pull/17240). It's a tradeoff between code structure and algorithm completeness. I tend to keep the size and thus create this new pr without changing cost model.
What this pr does:
1. We only consider consecutive inner joinable items, thus excluding cartesian products in reordering procedure. This significantly reduces the search space and memory overhead of memo. Otherwise every combination of items will exist in the memo.
2. This pr also includes a bug fix: if a leaf item is a project(_, child), current solution will miss the project.
## How was this patch tested?
Added test cases.
Author: wangzhenhua <wangzhenhua@huawei.com>
Closes#17286 from wzhfy/joinReorder3.
## What changes were proposed in this pull request?
This pr added entries in `FunctionRegistry` and supported `from_json` in SQL.
## How was this patch tested?
Added tests in `JsonFunctionsSuite` and `SQLQueryTestSuite`.
Author: Takeshi Yamamuro <yamamuro@apache.org>
Closes#17320 from maropu/SPARK-19967.
## What changes were proposed in this pull request?
Allows null values of the pivot column to be included in the pivot values list without throwing NPE
Note this PR was made as an alternative to #17224 but preserves the two phase aggregate operation that is needed for good performance.
## How was this patch tested?
Additional unit test
Author: Andrew Ray <ray.andrew@gmail.com>
Closes#17226 from aray/pivot-null.
## What changes were proposed in this pull request?
Currently `SessionCatalogSuite` is only for `InMemoryCatalog`, there is no suite for `HiveExternalCatalog`.
And there are some ddl function is not proper to test in `ExternalCatalogSuite`, because some logic are not full implement in `ExternalCatalog`, these ddl functions are full implement in `SessionCatalog`(e.g. merge the same logic from `ExternalCatalog` up to `SessionCatalog` ).
It is better to test it in `SessionCatalogSuite` for this situation.
So we should add a test suite for `SessionCatalog` with `HiveExternalCatalog`
The main change is that in `SessionCatalogSuite` add two functions:
`withBasicCatalog` and `withEmptyCatalog`
And replace the code like `val catalog = new SessionCatalog(newBasicCatalog)` with above two functions
## How was this patch tested?
add `HiveExternalSessionCatalogSuite`
Author: windpiger <songjun@outlook.com>
Closes#17287 from windpiger/sessioncatalogsuit.
### What changes were proposed in this pull request?
Specifying the table schema in DDL formats is needed for different scenarios. For example,
- [specifying the schema in SQL function `from_json` using DDL formats](https://issues.apache.org/jira/browse/SPARK-19637), which is suggested by marmbrus ,
- [specifying the customized JDBC data types](https://github.com/apache/spark/pull/16209).
These two PRs need users to use the JSON format to specify the table schema. This is not user friendly.
This PR is to provide a `parseTableSchema` API in `ParserInterface`.
### How was this patch tested?
Added a test suite `TableSchemaParserSuite`
Author: Xiao Li <gatorsmile@gmail.com>
Closes#17171 from gatorsmile/parseDDLStmt.
## What changes were proposed in this pull request?
The current master throws `StackOverflowError` in `createDataFrame`/`createDataset` if bean has one's own class in fields;
```
public class SelfClassInFieldBean implements Serializable {
private SelfClassInFieldBean child;
...
}
```
This pr added code to throw `UnsupportedOperationException` in that case as soon as possible.
## How was this patch tested?
Added tests in `JavaDataFrameSuite` and `JavaDatasetSuite`.
Author: Takeshi Yamamuro <yamamuro@apache.org>
Closes#17188 from maropu/SPARK-19751.
## What changes were proposed in this pull request?
unify a exception erro msg for dropdatabase when the database still have some tables for HiveExternalCatalog and InMemoryCatalog
## How was this patch tested?
N/A
Author: windpiger <songjun@outlook.com>
Closes#17305 from windpiger/unifyErromsg.
## What issue does this PR address ?
Jira: https://issues.apache.org/jira/browse/SPARK-13450
In `SortMergeJoinExec`, rows of the right relation having the same value for a join key are buffered in-memory. In case of skew, this causes OOMs (see comments in SPARK-13450 for more details). Heap dump from a failed job confirms this : https://issues.apache.org/jira/secure/attachment/12846382/heap-dump-analysis.png . While its possible to increase the heap size to workaround, Spark should be resilient to such issues as skews can happen arbitrarily.
## Change proposed in this pull request
- Introduces `ExternalAppendOnlyUnsafeRowArray`
- It holds `UnsafeRow`s in-memory upto a certain threshold.
- After the threshold is hit, it switches to `UnsafeExternalSorter` which enables spilling of the rows to disk. It does NOT sort the data.
- Allows iterating the array multiple times. However, any alteration to the array (using `add` or `clear`) will invalidate the existing iterator(s)
- `WindowExec` was already using `UnsafeExternalSorter` to support spilling. Changed it to use the new array
- Changed `SortMergeJoinExec` to use the new array implementation
- NOTE: I have not changed FULL OUTER JOIN to use this new array implementation. Changing that will need more surgery and I will rather put up a separate PR for that once this gets in.
- Changed `CartesianProductExec` to use the new array implementation
#### Note for reviewers
The diff can be divided into 3 parts. My motive behind having all the changes in a single PR was to demonstrate that the API is sane and supports 2 use cases. If reviewing as 3 separate PRs would help, I am happy to make the split.
## How was this patch tested ?
#### Unit testing
- Added unit tests `ExternalAppendOnlyUnsafeRowArray` to validate all its APIs and access patterns
- Added unit test for `SortMergeExec`
- with and without spill for inner join, left outer join, right outer join to confirm that the spill threshold config behaves as expected and output is as expected.
- This PR touches the scanning logic in `SortMergeExec` for _all_ joins (except FULL OUTER JOIN). However, I expect existing test cases to cover that there is no regression in correctness.
- Added unit test for `WindowExec` to check behavior of spilling and correctness of results.
#### Stress testing
- Confirmed that OOM is gone by running against a production job which used to OOM
- Since I cannot share details about prod workload externally, created synthetic data to mimic the issue. Ran before and after the fix to demonstrate the issue and query success with this PR
Generating the synthetic data
```
./bin/spark-shell --driver-memory=6G
import org.apache.spark.sql._
val hc = SparkSession.builder.master("local").getOrCreate()
hc.sql("DROP TABLE IF EXISTS spark_13450_large_table").collect
hc.sql("DROP TABLE IF EXISTS spark_13450_one_row_table").collect
val df1 = (0 until 1).map(i => ("10", "100", i.toString, (i * 2).toString)).toDF("i", "j", "str1", "str2")
df1.write.format("org.apache.spark.sql.hive.orc.OrcFileFormat").bucketBy(100, "i", "j").sortBy("i", "j").saveAsTable("spark_13450_one_row_table")
val df2 = (0 until 3000000).map(i => ("10", "100", i.toString, (i * 2).toString)).toDF("i", "j", "str1", "str2")
df2.write.format("org.apache.spark.sql.hive.orc.OrcFileFormat").bucketBy(100, "i", "j").sortBy("i", "j").saveAsTable("spark_13450_large_table")
```
Ran this against trunk VS local build with this PR. OOM repros with trunk and with the fix this query runs fine.
```
./bin/spark-shell --driver-java-options="-XX:+HeapDumpOnOutOfMemoryError -XX:HeapDumpPath=/tmp/spark.driver.heapdump.hprof"
import org.apache.spark.sql._
val hc = SparkSession.builder.master("local").getOrCreate()
hc.sql("SET spark.sql.autoBroadcastJoinThreshold=1")
hc.sql("SET spark.sql.sortMergeJoinExec.buffer.spill.threshold=10000")
hc.sql("DROP TABLE IF EXISTS spark_13450_result").collect
hc.sql("""
CREATE TABLE spark_13450_result
AS
SELECT
a.i AS a_i, a.j AS a_j, a.str1 AS a_str1, a.str2 AS a_str2,
b.i AS b_i, b.j AS b_j, b.str1 AS b_str1, b.str2 AS b_str2
FROM
spark_13450_one_row_table a
JOIN
spark_13450_large_table b
ON
a.i=b.i AND
a.j=b.j
""")
```
## Performance comparison
### Macro-benchmark
I ran a SMB join query over two real world tables (2 trillion rows (40 TB) and 6 million rows (120 GB)). Note that this dataset does not have skew so no spill happened. I saw improvement in CPU time by 2-4% over version without this PR. This did not add up as I was expected some regression. I think allocating array of capacity of 128 at the start (instead of starting with default size 16) is the sole reason for the perf. gain : https://github.com/tejasapatil/spark/blob/SPARK-13450_smb_buffer_oom/sql/core/src/main/scala/org/apache/spark/sql/execution/ExternalAppendOnlyUnsafeRowArray.scala#L43 . I could remove that and rerun, but effectively the change will be deployed in this form and I wanted to see the effect of it over large workload.
### Micro-benchmark
Two types of benchmarking can be found in `ExternalAppendOnlyUnsafeRowArrayBenchmark`:
[A] Comparing `ExternalAppendOnlyUnsafeRowArray` against raw `ArrayBuffer` when all rows fit in-memory and there is no spill
```
Array with 1000 rows: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
ArrayBuffer 7821 / 7941 33.5 29.8 1.0X
ExternalAppendOnlyUnsafeRowArray 8798 / 8819 29.8 33.6 0.9X
Array with 30000 rows: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
ArrayBuffer 19200 / 19206 25.6 39.1 1.0X
ExternalAppendOnlyUnsafeRowArray 19558 / 19562 25.1 39.8 1.0X
Array with 100000 rows: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
ArrayBuffer 5949 / 6028 17.2 58.1 1.0X
ExternalAppendOnlyUnsafeRowArray 6078 / 6138 16.8 59.4 1.0X
```
[B] Comparing `ExternalAppendOnlyUnsafeRowArray` against raw `UnsafeExternalSorter` when there is spilling of data
```
Spilling with 1000 rows: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
UnsafeExternalSorter 9239 / 9470 28.4 35.2 1.0X
ExternalAppendOnlyUnsafeRowArray 8857 / 8909 29.6 33.8 1.0X
Spilling with 10000 rows: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
UnsafeExternalSorter 4 / 5 39.3 25.5 1.0X
ExternalAppendOnlyUnsafeRowArray 5 / 6 29.8 33.5 0.8X
```
Author: Tejas Patil <tejasp@fb.com>
Closes#16909 from tejasapatil/SPARK-13450_smb_buffer_oom.
## What changes were proposed in this pull request?
We should restrict the nested level of a view, to avoid stack overflow exception during the view resolution.
## How was this patch tested?
Add new test case in `SQLViewSuite`.
Author: jiangxingbo <jiangxb1987@gmail.com>
Closes#17241 from jiangxb1987/view-depth.
## What changes were proposed in this pull request?
When dynamic partition value is null or empty string, we should write the data to a directory like `a=__HIVE_DEFAULT_PARTITION__`, when we read the data back, we should respect this special directory name and treat it as null.
This is the same behavior of impala, see https://issues.apache.org/jira/browse/IMPALA-252
## How was this patch tested?
new regression test
Author: Wenchen Fan <wenchen@databricks.com>
Closes#17277 from cloud-fan/partition.
## What changes were proposed in this pull request?
As timezone setting can also affect partition values, it works for all formats, we should make it clear.
## How was this patch tested?
Existing tests.
Author: Takuya UESHIN <ueshin@databricks.com>
Closes#17281 from ueshin/issues/SPARK-19817.
## What changes were proposed in this pull request?
This PR fixes the following problem:
````
Seq((1, 2)).toDF("a1", "a2").createOrReplaceTempView("a")
Seq[(java.lang.Integer, java.lang.Integer)]((1, null)).toDF("b1", "b2").createOrReplaceTempView("b")
// The expected result is 1 row of (1,2) as shown in the next statement.
sql("select * from a where a1 not in (select b1 from b where b2 = a2)").show
+---+---+
| a1| a2|
+---+---+
+---+---+
sql("select * from a where a1 not in (select b1 from b where b2 = 2)").show
+---+---+
| a1| a2|
+---+---+
| 1| 2|
+---+---+
````
There are a number of scenarios to consider:
1. When the correlated predicate yields a match (i.e., B.B2 = A.A2)
1.1. When the NOT IN expression yields a match (i.e., A.A1 = B.B1)
1.2. When the NOT IN expression yields no match (i.e., A.A1 = B.B1 returns false)
1.3. When A.A1 is null
1.4. When B.B1 is null
1.4.1. When A.A1 is not null
1.4.2. When A.A1 is null
2. When the correlated predicate yields no match (i.e.,B.B2 = A.A2 is false or unknown)
2.1. When B.B2 is null and A.A2 is null
2.2. When B.B2 is null and A.A2 is not null
2.3. When the value of A.A2 does not match any of B.B2
````
A.A1 A.A2 B.B1 B.B2
----- ----- ----- -----
1 1 1 1 (1.1)
2 1 (1.2)
null 1 (1.3)
1 3 null 3 (1.4.1)
null 3 (1.4.2)
1 null 1 null (2.1)
null 2 (2.2 & 2.3)
````
We can divide the evaluation of the above correlated NOT IN subquery into 2 groups:-
Group 1: The rows in A when there is a match from the correlated predicate (A.A1 = B.B1)
In this case, the result of the subquery is not empty and the semantics of the NOT IN depends solely on the evaluation of the equality comparison of the columns of NOT IN, i.e., A1 = B1, which says
- If A.A1 is null, the row is filtered (1.3 and 1.4.2)
- If A.A1 = B.B1, the row is filtered (1.1)
- If B.B1 is null, any rows of A in the same group (A.A2 = B.B2) is filtered (1.4.1 & 1.4.2)
- Otherwise, the row is qualified.
Hence, in this group, the result is the row from (1.2).
Group 2: The rows in A when there is no match from the correlated predicate (A.A2 = B.B2)
In this case, all the rows in A, including the rows where A.A1, are qualified because the subquery returns an empty set and by the semantics of the NOT IN, all rows from the parent side qualifies as the result set, that is, the rows from (2.1, 2.2 and 2.3).
In conclusion, the correct result set of the above query is
````
A.A1 A.A2
----- -----
2 1 (1.2)
1 null (2.1)
null 2 (2.2 & 2.3)
````
## How was this patch tested?
unit tests, regression tests, and new test cases focusing on the problem being fixed.
Author: Nattavut Sutyanyong <nsy.can@gmail.com>
Closes#17294 from nsyca/18966.
## What changes were proposed in this pull request?
The `RemoveRedundantAlias` rule can change the output attributes (the expression id's to be precise) of a query by eliminating the redundant alias producing them. This is no problem for a regular query, but can cause problems for correlated subqueries: The attributes produced by the subquery are used in the parent plan; changing them will break the parent plan.
This PR fixes this by wrapping a subquery in a `Subquery` top level node when it gets optimized. The `RemoveRedundantAlias` rule now recognizes `Subquery` and makes sure that the output attributes of the `Subquery` node are retained.
## How was this patch tested?
Added a test case to `RemoveRedundantAliasAndProjectSuite` and added a regression test to `SubquerySuite`.
Author: Herman van Hovell <hvanhovell@databricks.com>
Closes#17278 from hvanhovell/SPARK-19933.
## What changes were proposed in this pull request?
Commit 4ce970d714 in accidentally broke the 2.10 build for Spark. This PR fixes this by simplifying the offending pattern match.
## How was this patch tested?
Existing tests.
Author: Herman van Hovell <hvanhovell@databricks.com>
Closes#17288 from hvanhovell/SPARK-18874.
## What changes were proposed in this pull request?
We currently cannot use aliases in SQL function calls. This is inconvenient when you try to create a struct. This SQL query for example `select struct(1, 2) st`, will create a struct with column names `col1` and `col2`. This is even more problematic when we want to append a field to an existing struct. For example if we want to a field to struct `st` we would issue the following SQL query `select struct(st.*, 1) as st from src`, the result will be struct `st` with an a column with a non descriptive name `col3` (if `st` itself has 2 fields).
This PR proposes to change this by allowing the use of aliased expression in function parameters. For example `select struct(1 as a, 2 as b) st`, will create a struct with columns `a` & `b`.
## How was this patch tested?
Added a test to `ExpressionParserSuite` and added a test file for `SQLQueryTestSuite`.
Author: Herman van Hovell <hvanhovell@databricks.com>
Closes#17245 from hvanhovell/SPARK-19850.
## What changes were proposed in this pull request?
This patch moves SQLConf from sql/core to sql/catalyst. To minimize the changes, the patch used type alias to still keep CatalystConf (as a type alias) and SimpleCatalystConf (as a concrete class that extends SQLConf).
Motivation for the change is that it is pretty weird to have SQLConf only in sql/core and then we have to duplicate config options that impact optimizer/analyzer in sql/catalyst using CatalystConf.
## How was this patch tested?
N/A
Author: Reynold Xin <rxin@databricks.com>
Closes#17285 from rxin/SPARK-19944.
## What changes were proposed in this pull request?
- Timestamp hashing is done as per [TimestampWritable.hashCode()](ff67cdda1c/serde/src/java/org/apache/hadoop/hive/serde2/io/TimestampWritable.java (L406)) in Hive
- Interval hashing is done as per [HiveIntervalDayTime.hashCode()](ff67cdda1c/storage-api/src/java/org/apache/hadoop/hive/common/type/HiveIntervalDayTime.java (L178)). Note that there are inherent differences in how Hive and Spark store intervals under the hood which limits the ability to be in completely sync with hive's hashing function. I have explained this in the method doc.
- Date type was already supported. This PR adds test for that.
## How was this patch tested?
Added unit tests
Author: Tejas Patil <tejasp@fb.com>
Closes#17062 from tejasapatil/SPARK-17495_time_related_types.
