## What changes were proposed in this pull request?
Update the description and tests of three external API or functions `createFunction `, `length` and `repartitionByRange `
## How was this patch tested?
N/A
Author: gatorsmile <gatorsmile@gmail.com>
Closes#20495 from gatorsmile/updateFunc.
## What changes were proposed in this pull request?
In the current test case for CombineTypedFilters, we lack the test of FilterFunction, so let's add it.
In addition, in TypedFilterOptimizationSuite's existing test cases, Let's extract a common LocalRelation.
## How was this patch tested?
add new test cases.
Author: caoxuewen <cao.xuewen@zte.com.cn>
Closes#20482 from heary-cao/TypedFilterOptimizationSuite.
## What changes were proposed in this pull request?
https://issues.apache.org/jira/browse/SPARK-23309 reported a performance regression about cached table in Spark 2.3. While the investigating is still going on, this PR adds a conf to turn off the vectorized cache reader, to unblock the 2.3 release.
## How was this patch tested?
a new test
Author: Wenchen Fan <wenchen@databricks.com>
Closes#20483 from cloud-fan/cache.
## What changes were proposed in this pull request?
SpecifiedWindowFrame.defaultWindowFrame(hasOrderSpecification, acceptWindowFrame) was designed to handle the cases when some Window functions don't support setting a window frame (e.g. rank). However this param is never used.
We may inline the whole of this function to simplify the code.
## How was this patch tested?
Existing tests.
Author: Xingbo Jiang <xingbo.jiang@databricks.com>
Closes#20463 from jiangxb1987/defaultWindowFrame.
## What changes were proposed in this pull request?
This PR include the following changes:
- Make the capacity of `VectorizedParquetRecordReader` configurable;
- Make the capacity of `OrcColumnarBatchReader` configurable;
- Update the error message when required capacity in writable columnar vector cannot be fulfilled.
## How was this patch tested?
N/A
Author: Xingbo Jiang <xingbo.jiang@databricks.com>
Closes#20361 from jiangxb1987/vectorCapacity.
## What changes were proposed in this pull request?
Here is the test snippet.
``` SQL
scala> Seq[(Integer, Integer)](
| (1, 1),
| (1, 3),
| (2, 3),
| (3, 3),
| (4, null),
| (5, null)
| ).toDF("key", "value").createOrReplaceTempView("src")
scala> sql(
| """
| |SELECT MAX(value) as value, key as col2
| |FROM src
| |GROUP BY key
| |ORDER BY value desc, key
| """.stripMargin).show
+-----+----+
|value|col2|
+-----+----+
| 3| 3|
| 3| 2|
| 3| 1|
| null| 5|
| null| 4|
+-----+----+
```SQL
Here is the explain output :
```SQL
== Parsed Logical Plan ==
'Sort ['value DESC NULLS LAST, 'key ASC NULLS FIRST], true
+- 'Aggregate ['key], ['MAX('value) AS value#9, 'key AS col2#10]
+- 'UnresolvedRelation `src`
== Analyzed Logical Plan ==
value: int, col2: int
Project [value#9, col2#10]
+- Sort [value#9 DESC NULLS LAST, col2#10 DESC NULLS LAST], true
+- Aggregate [key#5], [max(value#6) AS value#9, key#5 AS col2#10]
+- SubqueryAlias src
+- Project [_1#2 AS key#5, _2#3 AS value#6]
+- LocalRelation [_1#2, _2#3]
``` SQL
The sort direction is being wrongly changed from ASC to DSC while resolving ```Sort``` in
resolveAggregateFunctions.
The above testcase models TPCDS-Q71 and thus we have the same issue in Q71 as well.
## How was this patch tested?
A few tests are added in SQLQuerySuite.
Author: Dilip Biswal <dbiswal@us.ibm.com>
Closes#20453 from dilipbiswal/local_spark.
## What changes were proposed in this pull request?
This PR is to fix the `ReplaceExceptWithFilter` rule when the right's Filter contains the references that are not in the left output.
Before this PR, we got the error like
```
java.util.NoSuchElementException: key not found: a
at scala.collection.MapLike$class.default(MapLike.scala:228)
at scala.collection.AbstractMap.default(Map.scala:59)
at scala.collection.MapLike$class.apply(MapLike.scala:141)
at scala.collection.AbstractMap.apply(Map.scala:59)
```
After this PR, `ReplaceExceptWithFilter ` will not take an effect in this case.
## How was this patch tested?
Added tests
Author: gatorsmile <gatorsmile@gmail.com>
Closes#20444 from gatorsmile/fixReplaceExceptWithFilter.
## What changes were proposed in this pull request?
Still saw the performance regression introduced by `spark.sql.codegen.hugeMethodLimit` in our internal workloads. There are two major issues in the current solution.
- The size of the complied byte code is not identical to the bytecode size of the method. The detection is still not accurate.
- The bytecode size of a single operator (e.g., `SerializeFromObject`) could still exceed 8K limit. We saw the performance regression in such scenario.
Since it is close to the release of 2.3, we decide to increase it to 64K for avoiding the perf regression.
## How was this patch tested?
N/A
Author: gatorsmile <gatorsmile@gmail.com>
Closes#20434 from gatorsmile/revertConf.
## What changes were proposed in this pull request?
Rename the public APIs and names of pandas udfs.
- `PANDAS SCALAR UDF` -> `SCALAR PANDAS UDF`
- `PANDAS GROUP MAP UDF` -> `GROUPED MAP PANDAS UDF`
- `PANDAS GROUP AGG UDF` -> `GROUPED AGG PANDAS UDF`
## How was this patch tested?
The existing tests
Author: gatorsmile <gatorsmile@gmail.com>
Closes#20428 from gatorsmile/renamePandasUDFs.
## What changes were proposed in this pull request?
This is a followup to #19575 which added a section on setting max Arrow record batches and this will externalize the conf that was referenced in the docs.
## How was this patch tested?
NA
Author: Bryan Cutler <cutlerb@gmail.com>
Closes#20423 from BryanCutler/arrow-user-doc-externalize-maxRecordsPerBatch-SPARK-22221.
## What changes were proposed in this pull request?
It is a common pattern to apply multiple transforms to a `Dataset` (using `Dataset.withColumn` for example. This is currently quite expensive because we run `CheckAnalysis` on the full plan and create an encoder for each intermediate `Dataset`.
This PR extends the usage of the `AnalysisBarrier` to include `CheckAnalysis`. By doing this we hide the already analyzed plan from `CheckAnalysis` because barrier is a `LeafNode`. The `AnalysisBarrier` is in the `FinishAnalysis` phase of the optimizer.
We also make binding the `Dataset` encoder lazy. The bound encoder is only needed when we materialize the dataset.
## How was this patch tested?
Existing test should cover this.
Author: Herman van Hovell <hvanhovell@databricks.com>
Closes#20402 from hvanhovell/SPARK-23223.
## What changes were proposed in this pull request?
Currently, all Aggregate operations will go into RemoveRepetitionFromGroupExpressions, but there is no group expression or there is no duplicate group expression in group expression, we not need copy for logic plan.
## How was this patch tested?
the existed test case.
Author: caoxuewen <cao.xuewen@zte.com.cn>
Closes#20375 from heary-cao/RepetitionGroupExpressions.
## What changes were proposed in this pull request?
Currently we have `ReadTask` in data source v2 reader, while in writer we have `DataWriterFactory`.
To make the naming consistent and better, renaming `ReadTask` to `DataReaderFactory`.
## How was this patch tested?
Unit test
Author: Wang Gengliang <ltnwgl@gmail.com>
Closes#20397 from gengliangwang/rename.
## What changes were proposed in this pull request?
This PR proposes to expose few internal configurations found in the documentation.
Also it fixes the description for `spark.sql.execution.arrow.enabled`.
It's quite self-explanatory.
## How was this patch tested?
N/A
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#20403 from HyukjinKwon/minor-doc-arrow.
## What changes were proposed in this pull request?
Replace streaming V2 sinks with a unified StreamWriteSupport interface, with a shim to use it with microbatch execution.
Add a new SQL config to use for disabling V2 sinks, falling back to the V1 sink implementation.
## How was this patch tested?
Existing tests, which in the case of Kafka (the only existing continuous V2 sink) now use V2 for microbatch.
Author: Jose Torres <jose@databricks.com>
Closes#20369 from jose-torres/streaming-sink.
## What changes were proposed in this pull request?
Currently shuffle repartition uses RoundRobinPartitioning, the generated result is nondeterministic since the sequence of input rows are not determined.
The bug can be triggered when there is a repartition call following a shuffle (which would lead to non-deterministic row ordering), as the pattern shows below:
upstream stage -> repartition stage -> result stage
(-> indicate a shuffle)
When one of the executors process goes down, some tasks on the repartition stage will be retried and generate inconsistent ordering, and some tasks of the result stage will be retried generating different data.
The following code returns 931532, instead of 1000000:
```
import scala.sys.process._
import org.apache.spark.TaskContext
val res = spark.range(0, 1000 * 1000, 1).repartition(200).map { x =>
x
}.repartition(200).map { x =>
if (TaskContext.get.attemptNumber == 0 && TaskContext.get.partitionId < 2) {
throw new Exception("pkill -f java".!!)
}
x
}
res.distinct().count()
```
In this PR, we propose a most straight-forward way to fix this problem by performing a local sort before partitioning, after we make the input row ordering deterministic, the function from rows to partitions is fully deterministic too.
The downside of the approach is that with extra local sort inserted, the performance of repartition() will go down, so we add a new config named `spark.sql.execution.sortBeforeRepartition` to control whether this patch is applied. The patch is default enabled to be safe-by-default, but user may choose to manually turn it off to avoid performance regression.
This patch also changes the output rows ordering of repartition(), that leads to a bunch of test cases failure because they are comparing the results directly.
## How was this patch tested?
Add unit test in ExchangeSuite.
With this patch(and `spark.sql.execution.sortBeforeRepartition` set to true), the following query returns 1000000:
```
import scala.sys.process._
import org.apache.spark.TaskContext
spark.conf.set("spark.sql.execution.sortBeforeRepartition", "true")
val res = spark.range(0, 1000 * 1000, 1).repartition(200).map { x =>
x
}.repartition(200).map { x =>
if (TaskContext.get.attemptNumber == 0 && TaskContext.get.partitionId < 2) {
throw new Exception("pkill -f java".!!)
}
x
}
res.distinct().count()
res7: Long = 1000000
```
Author: Xingbo Jiang <xingbo.jiang@databricks.com>
Closes#20393 from jiangxb1987/shuffle-repartition.
## What changes were proposed in this pull request?
**Proposal**
Add a per-query ID to the codegen stages as represented by `WholeStageCodegenExec` operators. This ID will be used in
- the explain output of the physical plan, and in
- the generated class name.
Specifically, this ID will be stable within a query, counting up from 1 in depth-first post-order for all the `WholeStageCodegenExec` inserted into a plan.
The ID value 0 is reserved for "free-floating" `WholeStageCodegenExec` objects, which may have been created for one-off purposes, e.g. for fallback handling of codegen stages that failed to codegen the whole stage and wishes to codegen a subset of the children operators (as seen in `org.apache.spark.sql.execution.FileSourceScanExec#doExecute`).
Example: for the following query:
```scala
scala> spark.conf.set("spark.sql.autoBroadcastJoinThreshold", 1)
scala> val df1 = spark.range(10).select('id as 'x, 'id + 1 as 'y).orderBy('x).select('x + 1 as 'z, 'y)
df1: org.apache.spark.sql.DataFrame = [z: bigint, y: bigint]
scala> val df2 = spark.range(5)
df2: org.apache.spark.sql.Dataset[Long] = [id: bigint]
scala> val query = df1.join(df2, 'z === 'id)
query: org.apache.spark.sql.DataFrame = [z: bigint, y: bigint ... 1 more field]
```
The explain output before the change is:
```scala
scala> query.explain
== Physical Plan ==
*SortMergeJoin [z#9L], [id#13L], Inner
:- *Sort [z#9L ASC NULLS FIRST], false, 0
: +- Exchange hashpartitioning(z#9L, 200)
: +- *Project [(x#3L + 1) AS z#9L, y#4L]
: +- *Sort [x#3L ASC NULLS FIRST], true, 0
: +- Exchange rangepartitioning(x#3L ASC NULLS FIRST, 200)
: +- *Project [id#0L AS x#3L, (id#0L + 1) AS y#4L]
: +- *Range (0, 10, step=1, splits=8)
+- *Sort [id#13L ASC NULLS FIRST], false, 0
+- Exchange hashpartitioning(id#13L, 200)
+- *Range (0, 5, step=1, splits=8)
```
Note how codegen'd operators are annotated with a prefix `"*"`. See how the `SortMergeJoin` operator and its direct children `Sort` operators are adjacent and all annotated with the `"*"`, so it's hard to tell they're actually in separate codegen stages.
and after this change it'll be:
```scala
scala> query.explain
== Physical Plan ==
*(6) SortMergeJoin [z#9L], [id#13L], Inner
:- *(3) Sort [z#9L ASC NULLS FIRST], false, 0
: +- Exchange hashpartitioning(z#9L, 200)
: +- *(2) Project [(x#3L + 1) AS z#9L, y#4L]
: +- *(2) Sort [x#3L ASC NULLS FIRST], true, 0
: +- Exchange rangepartitioning(x#3L ASC NULLS FIRST, 200)
: +- *(1) Project [id#0L AS x#3L, (id#0L + 1) AS y#4L]
: +- *(1) Range (0, 10, step=1, splits=8)
+- *(5) Sort [id#13L ASC NULLS FIRST], false, 0
+- Exchange hashpartitioning(id#13L, 200)
+- *(4) Range (0, 5, step=1, splits=8)
```
Note that the annotated prefix becomes `"*(id) "`. See how the `SortMergeJoin` operator and its direct children `Sort` operators have different codegen stage IDs.
It'll also show up in the name of the generated class, as a suffix in the format of `GeneratedClass$GeneratedIterator$id`.
For example, note how `GeneratedClass$GeneratedIteratorForCodegenStage3` and `GeneratedClass$GeneratedIteratorForCodegenStage6` in the following stack trace corresponds to the IDs shown in the explain output above:
```
"Executor task launch worker for task 42412957" daemon prio=5 tid=0x58 nid=NA runnable
java.lang.Thread.State: RUNNABLE
at org.apache.spark.sql.execution.UnsafeExternalRowSorter.insertRow(UnsafeExternalRowSorter.java:109)
at org.apache.spark.sql.catalyst.expressions.GeneratedClass$GeneratedIteratorForCodegenStage3.sort_addToSorter$(generated.java:32)
at org.apache.spark.sql.catalyst.expressions.GeneratedClass$GeneratedIteratorForCodegenStage3.processNext(generated.java:41)
at org.apache.spark.sql.execution.BufferedRowIterator.hasNext(BufferedRowIterator.java:43)
at org.apache.spark.sql.execution.WholeStageCodegenExec$$anonfun$9$$anon$1.hasNext(WholeStageCodegenExec.scala:494)
at org.apache.spark.sql.catalyst.expressions.GeneratedClass$GeneratedIteratorForCodegenStage6.findNextInnerJoinRows$(generated.java:42)
at org.apache.spark.sql.catalyst.expressions.GeneratedClass$GeneratedIteratorForCodegenStage6.processNext(generated.java:101)
at org.apache.spark.sql.execution.BufferedRowIterator.hasNext(BufferedRowIterator.java:43)
at org.apache.spark.sql.execution.WholeStageCodegenExec$$anonfun$11$$anon$2.hasNext(WholeStageCodegenExec.scala:513)
at org.apache.spark.sql.execution.SparkPlan$$anonfun$2.apply(SparkPlan.scala:253)
at org.apache.spark.sql.execution.SparkPlan$$anonfun$2.apply(SparkPlan.scala:247)
at org.apache.spark.rdd.RDD$$anonfun$mapPartitionsInternal$1$$anonfun$apply$25.apply(RDD.scala:828)
at org.apache.spark.rdd.RDD$$anonfun$mapPartitionsInternal$1$$anonfun$apply$25.apply(RDD.scala:828)
at org.apache.spark.rdd.MapPartitionsRDD.compute(MapPartitionsRDD.scala:38)
at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:324)
at org.apache.spark.rdd.RDD.iterator(RDD.scala:288)
at org.apache.spark.scheduler.ResultTask.runTask(ResultTask.scala:87)
at org.apache.spark.scheduler.Task.run(Task.scala:109)
at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:345)
at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
at java.lang.Thread.run(Thread.java:748)
```
**Rationale**
Right now, the codegen from Spark SQL lacks the means to differentiate between a couple of things:
1. It's hard to tell which physical operators are in the same WholeStageCodegen stage. Note that this "stage" is a separate notion from Spark's RDD execution stages; this one is only to delineate codegen units.
There can be adjacent physical operators that are both codegen'd but are in separate codegen stages. Some of this is due to hacky implementation details, such as the case with `SortMergeJoin` and its `Sort` inputs -- they're hard coded to be split into separate stages although both are codegen'd.
When printing out the explain output of the physical plan, you'd only see the codegen'd physical operators annotated with a preceding star (`'*'`) but would have no way to figure out if they're in the same stage.
2. Performance/error diagnosis
The generated code has class/method names that are hard to differentiate between queries or even between codegen stages within the same query. If we use a Java-level profiler to collect profiles, or if we encounter a Java-level exception with a stack trace in it, it's really hard to tell which part of a query it's at.
By introducing a per-query codegen stage ID, we'd at least be able to know which codegen stage (and in turn, which group of physical operators) was a profile tick or an exception happened.
The reason why this proposal uses a per-query ID is because it's stable within a query, so that multiple runs of the same query will see the same resulting IDs. This both benefits understandability for users, and also it plays well with the codegen cache in Spark SQL which uses the generated source code as the key.
The downside to using per-query IDs as opposed to a per-session or globally incrementing ID is of course we can't tell apart different query runs with this ID alone. But for now I believe this is a good enough tradeoff.
## How was this patch tested?
Existing tests. This PR does not involve any runtime behavior changes other than some name changes.
The SQL query test suites that compares explain outputs have been updates to ignore the newly added `codegenStageId`.
Author: Kris Mok <kris.mok@databricks.com>
Closes#20224 from rednaxelafx/wsc-codegenstageid.
## What changes were proposed in this pull request?
It has been observed in SPARK-21603 that whole-stage codegen suffers performance degradation, if the generated functions are too long to be optimized by JIT.
We basically produce a single function to incorporate generated codes from all physical operators in whole-stage. Thus, it is possibly to grow the size of generated function over a threshold that we can't have JIT optimization for it anymore.
This patch is trying to decouple the logic of consuming rows in physical operators to avoid a giant function processing rows.
## How was this patch tested?
Added tests.
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes#18931 from viirya/SPARK-21717.
## What changes were proposed in this pull request?
The `GenArrayData.genCodeToCreateArrayData` produces illegal java code when code splitting is enabled. This is used in `CreateArray` and `CreateMap` expressions for complex object arrays.
This issue is caused by a typo.
## How was this patch tested?
Added a regression test in `complexTypesSuite`.
Author: Herman van Hovell <hvanhovell@databricks.com>
Closes#20391 from hvanhovell/SPARK-23208.
## What changes were proposed in this pull request?
This PR is repaired as follows
1、update y -> x in "left outer join" test case ,maybe is mistake.
2、add a new test case:"left outer join and left sides are limited"
3、add a new test case:"left outer join and right sides are limited"
4、add a new test case: "right outer join and right sides are limited"
5、add a new test case: "right outer join and left sides are limited"
6、Remove annotations without code implementation
## How was this patch tested?
add new unit test case.
Author: caoxuewen <cao.xuewen@zte.com.cn>
Closes#20381 from heary-cao/LimitPushdownSuite.
## What changes were proposed in this pull request?
We need to override the prettyName for bit_length and octet_length for getting the expected auto-generated alias name.
## How was this patch tested?
The existing tests
Author: gatorsmile <gatorsmile@gmail.com>
Closes#20358 from gatorsmile/test2.3More.
## What changes were proposed in this pull request?
Add support for using pandas UDFs with groupby().agg().
This PR introduces a new type of pandas UDF - group aggregate pandas UDF. This type of UDF defines a transformation of multiple pandas Series -> a scalar value. Group aggregate pandas UDFs can be used with groupby().agg(). Note group aggregate pandas UDF doesn't support partial aggregation, i.e., a full shuffle is required.
This PR doesn't support group aggregate pandas UDFs that return ArrayType, StructType or MapType. Support for these types is left for future PR.
## How was this patch tested?
GroupbyAggPandasUDFTests
Author: Li Jin <ice.xelloss@gmail.com>
Closes#19872 from icexelloss/SPARK-22274-groupby-agg.
## What changes were proposed in this pull request?
a new interface which allows data source to report partitioning and avoid shuffle at Spark side.
The design is pretty like the internal distribution/partitioing framework. Spark defines a `Distribution` interfaces and several concrete implementations, and ask the data source to report a `Partitioning`, the `Partitioning` should tell Spark if it can satisfy a `Distribution` or not.
## How was this patch tested?
new test
Author: Wenchen Fan <wenchen@databricks.com>
Closes#20201 from cloud-fan/partition-reporting.
## What changes were proposed in this pull request?
Typo fixes
## How was this patch tested?
Local build / Doc-only changes
Author: Jacek Laskowski <jacek@japila.pl>
Closes#20344 from jaceklaskowski/typo-fixes.
## What changes were proposed in this pull request?
Several improvements:
* provide a default implementation for the batch get methods
* rename `getChildColumn` to `getChild`, which is more concise
* remove `getStruct(int, int)`, it's only used to simplify the codegen, which is an internal thing, we should not add a public API for this purpose.
## How was this patch tested?
existing tests
Author: Wenchen Fan <wenchen@databricks.com>
Closes#20277 from cloud-fan/column-vector.
## What changes were proposed in this pull request?
CheckCartesianProduct raises an AnalysisException also when the join condition is always false/null. In this case, we shouldn't raise it, since the result will not be a cartesian product.
## How was this patch tested?
added UT
Author: Marco Gaido <marcogaido91@gmail.com>
Closes#20333 from mgaido91/SPARK-23087.
## What changes were proposed in this pull request?
This is a follow-up of #20246.
If a UDT in Python doesn't have its corresponding Scala UDT, cast to string will be the raw string of the internal value, e.g. `"org.apache.spark.sql.catalyst.expressions.UnsafeArrayDataxxxxxxxx"` if the internal type is `ArrayType`.
This pr fixes it by using its `sqlType` casting.
## How was this patch tested?
Added a test and existing tests.
Author: Takuya UESHIN <ueshin@databricks.com>
Closes#20306 from ueshin/issues/SPARK-23054/fup1.
## What changes were proposed in this pull request?
There were two related fixes regarding `from_json`, `get_json_object` and `json_tuple` ([Fix#1](c8803c0685),
[Fix#2](86174ea89b)), but they weren't comprehensive it seems. I wanted to extend those fixes to all the parsers, and add tests for each case.
## How was this patch tested?
Regression tests
Author: Burak Yavuz <brkyvz@gmail.com>
Closes#20302 from brkyvz/json-invfix.
## What changes were proposed in this pull request?
When there is an operation between Decimals and the result is a number which is not representable exactly with the result's precision and scale, Spark is returning `NULL`. This was done to reflect Hive's behavior, but it is against SQL ANSI 2011, which states that "If the result cannot be represented exactly in the result type, then whether it is rounded or truncated is implementation-defined". Moreover, Hive now changed its behavior in order to respect the standard, thanks to HIVE-15331.
Therefore, the PR propose to:
- update the rules to determine the result precision and scale according to the new Hive's ones introduces in HIVE-15331;
- round the result of the operations, when it is not representable exactly with the result's precision and scale, instead of returning `NULL`
- introduce a new config `spark.sql.decimalOperations.allowPrecisionLoss` which default to `true` (ie. the new behavior) in order to allow users to switch back to the previous one.
Hive behavior reflects SQLServer's one. The only difference is that the precision and scale are adjusted for all the arithmetic operations in Hive, while SQL Server is said to do so only for multiplications and divisions in the documentation. This PR follows Hive's behavior.
A more detailed explanation is available here: https://mail-archives.apache.org/mod_mbox/spark-dev/201712.mbox/%3CCAEorWNAJ4TxJR9NBcgSFMD_VxTg8qVxusjP%2BAJP-x%2BJV9zH-yA%40mail.gmail.com%3E.
## How was this patch tested?
modified and added UTs. Comparisons with results of Hive and SQLServer.
Author: Marco Gaido <marcogaido91@gmail.com>
Closes#20023 from mgaido91/SPARK-22036.
## What changes were proposed in this pull request?
Previously, PR #19201 fix the problem of non-converging constraints.
After that PR #19149 improve the loop and constraints is inferred only once.
So the problem of non-converging constraints is gone.
However, the case below will fail.
```
spark.range(5).write.saveAsTable("t")
val t = spark.read.table("t")
val left = t.withColumn("xid", $"id" + lit(1)).as("x")
val right = t.withColumnRenamed("id", "xid").as("y")
val df = left.join(right, "xid").filter("id = 3").toDF()
checkAnswer(df, Row(4, 3))
```
Because `aliasMap` replace all the aliased child. See the test case in PR for details.
This PR is to fix this bug by removing useless code for preventing non-converging constraints.
It can be also fixed with #20270, but this is much simpler and clean up the code.
## How was this patch tested?
Unit test
Author: Wang Gengliang <ltnwgl@gmail.com>
Closes#20278 from gengliangwang/FixConstraintSimple.
## What changes were proposed in this pull request?
ORC filter push-down is disabled by default from the beginning, [SPARK-2883](aa31e431fc (diff-41ef65b9ef5b518f77e2a03559893f4dR149)
).
Now, Apache Spark starts to depend on Apache ORC 1.4.1. For Apache Spark 2.3, this PR turns on ORC filter push-down by default like Parquet ([SPARK-9207](https://issues.apache.org/jira/browse/SPARK-21783)) as a part of [SPARK-20901](https://issues.apache.org/jira/browse/SPARK-20901), "Feature parity for ORC with Parquet".
## How was this patch tested?
Pass the existing tests.
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#20265 from dongjoon-hyun/SPARK-21783.
## What changes were proposed in this pull request?
There are already quite a few integration tests using window frames, but the unit tests coverage is not ideal.
In this PR the already existing tests are reorganized, extended and where gaps found additional cases added.
## How was this patch tested?
Automated: Pass the Jenkins.
Author: Gabor Somogyi <gabor.g.somogyi@gmail.com>
Closes#20019 from gaborgsomogyi/SPARK-22361.
## What changes were proposed in this pull request?
The following SQL involving scalar correlated query returns a map exception.
``` SQL
SELECT t1a
FROM t1
WHERE t1a = (SELECT count(*)
FROM t2
WHERE t2c = t1c
HAVING count(*) >= 1)
```
``` SQL
key not found: ExprId(278,786682bb-41f9-4bd5-a397-928272cc8e4e)
java.util.NoSuchElementException: key not found: ExprId(278,786682bb-41f9-4bd5-a397-928272cc8e4e)
at scala.collection.MapLike$class.default(MapLike.scala:228)
at scala.collection.AbstractMap.default(Map.scala:59)
at scala.collection.MapLike$class.apply(MapLike.scala:141)
at scala.collection.AbstractMap.apply(Map.scala:59)
at org.apache.spark.sql.catalyst.optimizer.RewriteCorrelatedScalarSubquery$.org$apache$spark$sql$catalyst$optimizer$RewriteCorrelatedScalarSubquery$$evalSubqueryOnZeroTups(subquery.scala:378)
at org.apache.spark.sql.catalyst.optimizer.RewriteCorrelatedScalarSubquery$$anonfun$org$apache$spark$sql$catalyst$optimizer$RewriteCorrelatedScalarSubquery$$constructLeftJoins$1.apply(subquery.scala:430)
at org.apache.spark.sql.catalyst.optimizer.RewriteCorrelatedScalarSubquery$$anonfun$org$apache$spark$sql$catalyst$optimizer$RewriteCorrelatedScalarSubquery$$constructLeftJoins$1.apply(subquery.scala:426)
```
In this case, after evaluating the HAVING clause "count(*) > 1" statically
against the binding of aggregtation result on empty input, we determine
that this query will not have a the count bug. We should simply return
the evalSubqueryOnZeroTups with empty value.
(Please fill in changes proposed in this fix)
## How was this patch tested?
A new test was added in the Subquery bucket.
Author: Dilip Biswal <dbiswal@us.ibm.com>
Closes#20283 from dilipbiswal/scalar-count-defect.
## What changes were proposed in this pull request?
When a user puts the wrong number of parameters in a function, an AnalysisException is thrown. If the function is a UDF, he user is told how many parameters the function expected and how many he/she put. If the function, instead, is a built-in one, no information about the number of parameters expected and the actual one is provided. This can help in some cases, to debug the errors (eg. bad quotes escaping may lead to a different number of parameters than expected, etc. etc.)
The PR adds the information about the number of parameters passed and the expected one, analogously to what happens for UDF.
## How was this patch tested?
modified existing UT + manual test
Author: Marco Gaido <marcogaido91@gmail.com>
Closes#20271 from mgaido91/SPARK-23080.
## What changes were proposed in this pull request?
Problem: it throw TempTableAlreadyExistsException and output "Temporary table '$table' already exists" when we create temp view by using org.apache.spark.sql.catalyst.catalog.GlobalTempViewManager#create, it's improper.
So fix improper information about TempTableAlreadyExistsException when create temp view:
change "Temporary table" to "Temporary view"
## How was this patch tested?
test("rename temporary view - destination table already exists, with: CREATE TEMPORARY view")
test("rename temporary view - destination table with database name,with:CREATE TEMPORARY view")
Author: xubo245 <601450868@qq.com>
Closes#20227 from xubo245/fixDeprecated.
## What changes were proposed in this pull request?
Remove `MaxPermSize` for `sql` module
## How was this patch tested?
Manually tested.
Author: Yuming Wang <yumwang@ebay.com>
Closes#20268 from wangyum/SPARK-19550-MaxPermSize.
## What changes were proposed in this pull request?
This pr fixed the issue when casting `UserDefinedType`s into strings;
```
>>> from pyspark.ml.classification import MultilayerPerceptronClassifier
>>> from pyspark.ml.linalg import Vectors
>>> df = spark.createDataFrame([(0.0, Vectors.dense([0.0, 0.0])), (1.0, Vectors.dense([0.0, 1.0]))], ["label", "features"])
>>> df.selectExpr("CAST(features AS STRING)").show(truncate = False)
+-------------------------------------------+
|features |
+-------------------------------------------+
|[6,1,0,0,2800000020,2,0,0,0] |
|[6,1,0,0,2800000020,2,0,0,3ff0000000000000]|
+-------------------------------------------+
```
The root cause is that `Cast` handles input data as `UserDefinedType.sqlType`(this is underlying storage type), so we should pass data into `UserDefinedType.deserialize` then `toString`.
This pr modified the result into;
```
+---------+
|features |
+---------+
|[0.0,0.0]|
|[0.0,1.0]|
+---------+
```
## How was this patch tested?
Added tests in `UserDefinedTypeSuite `.
Author: Takeshi Yamamuro <yamamuro@apache.org>
Closes#20246 from maropu/SPARK-23054.
## What changes were proposed in this pull request?
SHOW DATABASES (LIKE pattern = STRING)? Can be like the back increase?
When using this command, LIKE keyword can be removed.
You can refer to the SHOW TABLES command, SHOW TABLES 'test *' and SHOW TABELS like 'test *' can be used.
Similarly SHOW DATABASES 'test *' and SHOW DATABASES like 'test *' can be used.
## How was this patch tested?
unit tests manual tests
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: guoxiaolong <guo.xiaolong1@zte.com.cn>
Closes#20194 from guoxiaolongzte/SPARK-22999.
## What changes were proposed in this pull request?
`AnalysisBarrier` in the current master cuts off explain results for parsed logical plans;
```
scala> Seq((1, 1)).toDF("a", "b").groupBy("a").count().sample(0.1).explain(true)
== Parsed Logical Plan ==
Sample 0.0, 0.1, false, -7661439431999668039
+- AnalysisBarrier Aggregate [a#5], [a#5, count(1) AS count#14L]
```
To fix this, `AnalysisBarrier` needs to override `innerChildren` and this pr changed the output to;
```
== Parsed Logical Plan ==
Sample 0.0, 0.1, false, -5086223488015741426
+- AnalysisBarrier
+- Aggregate [a#5], [a#5, count(1) AS count#14L]
+- Project [_1#2 AS a#5, _2#3 AS b#6]
+- LocalRelation [_1#2, _2#3]
```
## How was this patch tested?
Added tests in `DataFrameSuite`.
Author: Takeshi Yamamuro <yamamuro@apache.org>
Closes#20247 from maropu/SPARK-23021-2.
## What changes were proposed in this pull request?
as per discussion in https://github.com/apache/spark/pull/19864#discussion_r156847927
the current HadoopFsRelation is purely based on the underlying file size which is not accurate and makes the execution vulnerable to errors like OOM
Users can enable CBO with the functionalities in https://github.com/apache/spark/pull/19864 to avoid this issue
This JIRA proposes to add a configurable factor to sizeInBytes method in HadoopFsRelation class so that users can mitigate this problem without CBO
## How was this patch tested?
Existing tests
Author: CodingCat <zhunansjtu@gmail.com>
Author: Nan Zhu <nanzhu@uber.com>
Closes#20072 from CodingCat/SPARK-22790.
## What changes were proposed in this pull request?
This patch bumps the master branch version to `2.4.0-SNAPSHOT`.
## How was this patch tested?
N/A
Author: gatorsmile <gatorsmile@gmail.com>
Closes#20222 from gatorsmile/bump24.
## What changes were proposed in this pull request?
Add support for `Null` type in the `schemaFor` method for Scala reflection.
## How was this patch tested?
Added UT
Author: Marco Gaido <marcogaido91@gmail.com>
Closes#20219 from mgaido91/SPARK-23025.
## What changes were proposed in this pull request?
(courtesy of liancheng)
Spark SQL supports both global aggregation and grouping aggregation. Global aggregation always return a single row with the initial aggregation state as the output, even there are zero input rows. Spark implements this by simply checking the number of grouping keys and treats an aggregation as a global aggregation if it has zero grouping keys.
However, this simple principle drops the ball in the following case:
```scala
spark.emptyDataFrame.dropDuplicates().agg(count($"*") as "c").show()
// +---+
// | c |
// +---+
// | 1 |
// +---+
```
The reason is that:
1. `df.dropDuplicates()` is roughly translated into something equivalent to:
```scala
val allColumns = df.columns.map { col }
df.groupBy(allColumns: _*).agg(allColumns.head, allColumns.tail: _*)
```
This translation is implemented in the rule `ReplaceDeduplicateWithAggregate`.
2. `spark.emptyDataFrame` contains zero columns and zero rows.
Therefore, rule `ReplaceDeduplicateWithAggregate` makes a confusing transformation roughly equivalent to the following one:
```scala
spark.emptyDataFrame.dropDuplicates()
=> spark.emptyDataFrame.groupBy().agg(Map.empty[String, String])
```
The above transformation is confusing because the resulting aggregate operator contains no grouping keys (because `emptyDataFrame` contains no columns), and gets recognized as a global aggregation. As a result, Spark SQL allocates a single row filled by the initial aggregation state and uses it as the output, and returns a wrong result.
To fix this issue, this PR tweaks `ReplaceDeduplicateWithAggregate` by appending a literal `1` to the grouping key list of the resulting `Aggregate` operator when the input plan contains zero output columns. In this way, `spark.emptyDataFrame.dropDuplicates()` is now translated into a grouping aggregation, roughly depicted as:
```scala
spark.emptyDataFrame.dropDuplicates()
=> spark.emptyDataFrame.groupBy(lit(1)).agg(Map.empty[String, String])
```
Which is now properly treated as a grouping aggregation and returns the correct answer.
## How was this patch tested?
New unit tests added
Author: Feng Liu <fengliu@databricks.com>
Closes#20174 from liufengdb/fix-duplicate.
## What changes were proposed in this pull request?
This is mostly from https://github.com/apache/spark/pull/13775
The wrapper solution is pretty good for string/binary type, as the ORC column vector doesn't keep bytes in a continuous memory region, and has a significant overhead when copying the data to Spark columnar batch. For other cases, the wrapper solution is almost same with the current solution.
I think we can treat the wrapper solution as a baseline and keep improving the writing to Spark solution.
## How was this patch tested?
existing tests.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#20205 from cloud-fan/orc.
## What changes were proposed in this pull request?
This pr fixed the issue when casting structs into strings;
```
scala> val df = Seq(((1, "a"), 0), ((2, "b"), 0)).toDF("a", "b")
scala> df.write.saveAsTable("t")
scala> sql("SELECT CAST(a AS STRING) FROM t").show
+-------------------+
| a|
+-------------------+
|[0,1,1800000001,61]|
|[0,2,1800000001,62]|
+-------------------+
```
This pr modified the result into;
```
+------+
| a|
+------+
|[1, a]|
|[2, b]|
+------+
```
## How was this patch tested?
Added tests in `CastSuite`.
Author: Takeshi Yamamuro <yamamuro@apache.org>
Closes#20176 from maropu/SPARK-22981.
## What changes were proposed in this pull request?
This PR adds an ORC columnar-batch reader to native `OrcFileFormat`. Since both Spark `ColumnarBatch` and ORC `RowBatch` are used together, it is faster than the current Spark implementation. This replaces the prior PR, #17924.
Also, this PR adds `OrcReadBenchmark` to show the performance improvement.
## How was this patch tested?
Pass the existing test cases.
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#19943 from dongjoon-hyun/SPARK-16060.
## What changes were proposed in this pull request?
This PR fixes a longstanding correctness bug in `GenerateUnsafeRowJoiner`. This class was introduced in https://github.com/apache/spark/pull/7821 (July 2015 / Spark 1.5.0+) and is used to combine pairs of UnsafeRows in TungstenAggregationIterator, CartesianProductExec, and AppendColumns.
### Bugs fixed by this patch
1. **Incorrect combining of null-tracking bitmaps**: when concatenating two UnsafeRows, the implementation "Concatenate the two bitsets together into a single one, taking padding into account". If one row has no columns then it has a bitset size of 0, but the code was incorrectly assuming that if the left row had a non-zero number of fields then the right row would also have at least one field, so it was copying invalid bytes and and treating them as part of the bitset. I'm not sure whether this bug was also present in the original implementation or whether it was introduced in https://github.com/apache/spark/pull/7892 (which fixed another bug in this code).
2. **Incorrect updating of data offsets for null variable-length fields**: after updating the bitsets and copying fixed-length and variable-length data, we need to perform adjustments to the offsets pointing the start of variable length fields's data. The existing code was _conditionally_ adding a fixed offset to correct for the new length of the combined row, but it is unsafe to do this if the variable-length field has a null value: we always represent nulls by storing `0` in the fixed-length slot, but this code was incorrectly incrementing those values. This bug was present since the original version of `GenerateUnsafeRowJoiner`.
### Why this bug remained latent for so long
The PR which introduced `GenerateUnsafeRowJoiner` features several randomized tests, including tests of the cases where one side of the join has no fields and where string-valued fields are null. However, the existing assertions were too weak to uncover this bug:
- If a null field has a non-zero value in its fixed-length data slot then this will not cause problems for field accesses because the null-tracking bitmap should still be correct and we will not try to use the incorrect offset for anything.
- If the null tracking bitmap is corrupted by joining against a row with no fields then the corruption occurs in field numbers past the actual field numbers contained in the row. Thus valid `isNullAt()` calls will not read the incorrectly-set bits.
The existing `GenerateUnsafeRowJoinerSuite` tests only exercised `.get()` and `isNullAt()`, but didn't actually check the UnsafeRows for bit-for-bit equality, preventing these bugs from failing assertions. It turns out that there was even a [GenerateUnsafeRowJoinerBitsetSuite](03377d2522/sql/catalyst/src/test/scala/org/apache/spark/sql/catalyst/expressions/codegen/GenerateUnsafeRowJoinerBitsetSuite.scala) but it looks like it also didn't catch this problem because it only tested the bitsets in an end-to-end fashion by accessing them through the `UnsafeRow` interface instead of actually comparing the bitsets' bytes.
### Impact of these bugs
- This bug will cause `equals()` and `hashCode()` to be incorrect for these rows, which will be problematic in case`GenerateUnsafeRowJoiner`'s results are used as join or grouping keys.
- Chained / repeated invocations of `GenerateUnsafeRowJoiner` may result in reads from invalid null bitmap positions causing fields to incorrectly become NULL (see the end-to-end example below).
- It looks like this generally only happens in `CartesianProductExec`, which our query optimizer often avoids executing (usually we try to plan a `BroadcastNestedLoopJoin` instead).
### End-to-end test case demonstrating the problem
The following query demonstrates how this bug may result in incorrect query results:
```sql
set spark.sql.autoBroadcastJoinThreshold=-1; -- Needed to trigger CartesianProductExec
create table a as select * from values 1;
create table b as select * from values 2;
SELECT
t3.col1,
t1.col1
FROM a t1
CROSS JOIN b t2
CROSS JOIN b t3
```
This should return `(2, 1)` but instead was returning `(null, 1)`.
Column pruning ends up trimming off all columns from `t2`, so when `t2` joins with another table this triggers the bitmap-copying bug. This incorrect bitmap is subsequently copied again when performing the final join, causing the final output to have an incorrectly-set null bit for the first field.
## How was this patch tested?
Strengthened the assertions in existing tests in GenerateUnsafeRowJoinerSuite. Also verified that the end-to-end test case which uncovered this now passes.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#20181 from JoshRosen/SPARK-22984-fix-generate-unsaferow-joiner-bitmap-bugs.
## What changes were proposed in this pull request?
**The current shuffle planning logic**
1. Each operator specifies the distribution requirements for its children, via the `Distribution` interface.
2. Each operator specifies its output partitioning, via the `Partitioning` interface.
3. `Partitioning.satisfy` determines whether a `Partitioning` can satisfy a `Distribution`.
4. For each operator, check each child of it, add a shuffle node above the child if the child partitioning can not satisfy the required distribution.
5. For each operator, check if its children's output partitionings are compatible with each other, via the `Partitioning.compatibleWith`.
6. If the check in 5 failed, add a shuffle above each child.
7. try to eliminate the shuffles added in 6, via `Partitioning.guarantees`.
This design has a major problem with the definition of "compatible".
`Partitioning.compatibleWith` is not well defined, ideally a `Partitioning` can't know if it's compatible with other `Partitioning`, without more information from the operator. For example, `t1 join t2 on t1.a = t2.b`, `HashPartitioning(a, 10)` should be compatible with `HashPartitioning(b, 10)` under this case, but the partitioning itself doesn't know it.
As a result, currently `Partitioning.compatibleWith` always return false except for literals, which make it almost useless. This also means, if an operator has distribution requirements for multiple children, Spark always add shuffle nodes to all the children(although some of them can be eliminated). However, there is no guarantee that the children's output partitionings are compatible with each other after adding these shuffles, we just assume that the operator will only specify `ClusteredDistribution` for multiple children.
I think it's very hard to guarantee children co-partition for all kinds of operators, and we can not even give a clear definition about co-partition between distributions like `ClusteredDistribution(a,b)` and `ClusteredDistribution(c)`.
I think we should drop the "compatible" concept in the distribution model, and let the operator achieve the co-partition requirement by special distribution requirements.
**Proposed shuffle planning logic after this PR**
(The first 4 are same as before)
1. Each operator specifies the distribution requirements for its children, via the `Distribution` interface.
2. Each operator specifies its output partitioning, via the `Partitioning` interface.
3. `Partitioning.satisfy` determines whether a `Partitioning` can satisfy a `Distribution`.
4. For each operator, check each child of it, add a shuffle node above the child if the child partitioning can not satisfy the required distribution.
5. For each operator, check if its children's output partitionings have the same number of partitions.
6. If the check in 5 failed, pick the max number of partitions from children's output partitionings, and add shuffle to child whose number of partitions doesn't equal to the max one.
The new distribution model is very simple, we only have one kind of relationship, which is `Partitioning.satisfy`. For multiple children, Spark only guarantees they have the same number of partitions, and it's the operator's responsibility to leverage this guarantee to achieve more complicated requirements. For example, non-broadcast joins can use the newly added `HashPartitionedDistribution` to achieve co-partition.
## How was this patch tested?
existing tests.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#19080 from cloud-fan/exchange.
## What changes were proposed in this pull request?
The following SQL query should return zero rows, but in Spark it actually returns one row:
```
SELECT 1 from (
SELECT 1 AS z,
MIN(a.x)
FROM (select 1 as x) a
WHERE false
) b
where b.z != b.z
```
The problem stems from the `PushDownPredicate` rule: when this rule encounters a filter on top of an Aggregate operator, e.g. `Filter(Agg(...))`, it removes the original filter and adds a new filter onto Aggregate's child, e.g. `Agg(Filter(...))`. This is sometimes okay, but the case above is a counterexample: because there is no explicit `GROUP BY`, we are implicitly computing a global aggregate over the entire table so the original filter was not acting like a `HAVING` clause filtering the number of groups: if we push this filter then it fails to actually reduce the cardinality of the Aggregate output, leading to the wrong answer.
In 2016 I fixed a similar problem involving invalid pushdowns of data-independent filters (filters which reference no columns of the filtered relation). There was additional discussion after my fix was merged which pointed out that my patch was an incomplete fix (see #15289), but it looks I must have either misunderstood the comment or forgot to follow up on the additional points raised there.
This patch fixes the problem by choosing to never push down filters in cases where there are no grouping expressions. Since there are no grouping keys, the only columns are aggregate columns and we can't push filters defined over aggregate results, so this change won't cause us to miss out on any legitimate pushdown opportunities.
## How was this patch tested?
New regression tests in `SQLQueryTestSuite` and `FilterPushdownSuite`.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#20180 from JoshRosen/SPARK-22983-dont-push-filters-beneath-aggs-with-empty-grouping-expressions.