This patch adds the building blocks for codegening subexpr elimination and implements
it end to end for UnsafeProjection. The building blocks can be used to do the same thing
for other operators.
It introduces some utilities to compute common sub expressions. Expressions can be added to
this data structure. The expr and its children will be recursively matched against existing
expressions (ones previously added) and grouped into common groups. This is built using
the existing `semanticEquals`. It does not understand things like commutative or associative
expressions. This can be done as future work.
After building this data structure, the codegen process takes advantage of it by:
1. Generating a helper function in the generated class that computes the common
subexpression. This is done for all common subexpressions that have at least
two occurrences and the expression tree is sufficiently complex.
2. When generating the apply() function, if the helper function exists, call that
instead of regenerating the expression tree. Repeated calls to the helper function
shortcircuit the evaluation logic.
Author: Nong Li <nong@databricks.com>
Author: Nong Li <nongli@gmail.com>
This patch had conflicts when merged, resolved by
Committer: Michael Armbrust <michael@databricks.com>
Closes#9480 from nongli/spark-10371.
Currently the user facing api for typed aggregation has some limitations:
* the customized typed aggregation must be the first of aggregation list
* the customized typed aggregation can only use long as buffer type
* the customized typed aggregation can only use flat type as result type
This PR tries to remove these limitations.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#9599 from cloud-fan/agg.
https://issues.apache.org/jira/browse/SPARK-9830
This PR contains the following main changes.
* Removing `AggregateExpression1`.
* Removing `Aggregate` operator, which is used to evaluate `AggregateExpression1`.
* Removing planner rule used to plan `Aggregate`.
* Linking `MultipleDistinctRewriter` to analyzer.
* Renaming `AggregateExpression2` to `AggregateExpression` and `AggregateFunction2` to `AggregateFunction`.
* Updating places where we create aggregate expression. The way to create aggregate expressions is `AggregateExpression(aggregateFunction, mode, isDistinct)`.
* Changing `val`s in `DeclarativeAggregate`s that touch children of this function to `lazy val`s (when we create aggregate expression in DataFrame API, children of an aggregate function can be unresolved).
Author: Yin Huai <yhuai@databricks.com>
Closes#9556 from yhuai/removeAgg1.
This PR adds a new interface for user-defined aggregations, that can be used in `DataFrame` and `Dataset` operations to take all of the elements of a group and reduce them to a single value.
For example, the following aggregator extracts an `int` from a specific class and adds them up:
```scala
case class Data(i: Int)
val customSummer = new Aggregator[Data, Int, Int] {
def prepare(d: Data) = d.i
def reduce(l: Int, r: Int) = l + r
def present(r: Int) = r
}.toColumn()
val ds: Dataset[Data] = ...
val aggregated = ds.select(customSummer)
```
By using helper functions, users can make a generic `Aggregator` that works on any input type:
```scala
/** An `Aggregator` that adds up any numeric type returned by the given function. */
class SumOf[I, N : Numeric](f: I => N) extends Aggregator[I, N, N] with Serializable {
val numeric = implicitly[Numeric[N]]
override def zero: N = numeric.zero
override def reduce(b: N, a: I): N = numeric.plus(b, f(a))
override def present(reduction: N): N = reduction
}
def sum[I, N : Numeric : Encoder](f: I => N): TypedColumn[I, N] = new SumOf(f).toColumn
```
These aggregators can then be used alongside other built-in SQL aggregations.
```scala
val ds = Seq(("a", 10), ("a", 20), ("b", 1), ("b", 2), ("c", 1)).toDS()
ds
.groupBy(_._1)
.agg(
sum(_._2), // The aggregator defined above.
expr("sum(_2)").as[Int], // A built-in dynatically typed aggregation.
count("*")) // A built-in statically typed aggregation.
.collect()
res0: ("a", 30, 30, 2L), ("b", 3, 3, 2L), ("c", 1, 1, 1L)
```
The current implementation focuses on integrating this into the typed API, but currently only supports running aggregations that return a single long value as explained in `TypedAggregateExpression`. This will be improved in a followup PR.
Author: Michael Armbrust <michael@databricks.com>
Closes#9555 from marmbrus/dataset-useragg.
Actually this was resolved by https://github.com/apache/spark/pull/8275.
But I found the JIRA issue for this is not marked as resolved since the PR above was made for another issue but the PR above resolved both.
I commented that this is resolved by the PR above; however, I opened this PR as I would like to just add
a little bit of corrections.
In the previous PR, I refactored the test by not reducing just collecting filters; however, this would not test properly `And` filter (which is not given to the tests). I unintentionally changed this from the original way (before being refactored).
In this PR, I just followed the original way to collect filters by reducing.
I would like to close this if this PR is inappropriate and somebody would like this deal with it in the separate PR related with this.
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#9554 from HyukjinKwon/SPARK-9557.
The reason is that:
1. For partitioned hive table, we will move the partitioned columns after data columns. (e.g. `<a: Int, b: Int>` partition by `a` will become `<b: Int, a: Int>`)
2. When append data to table, we use position to figure out how to match input columns to table's columns.
So when we append data to partitioned table, we will match wrong columns between input and table. A solution is reordering the input columns before match by position, like what we did for [`InsertIntoHadoopFsRelation`](https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/InsertIntoHadoopFsRelation.scala#L101-L105)
Author: Wenchen Fan <wenchen@databricks.com>
Closes#9408 from cloud-fan/append.
A few changes:
1. Removed fold, since it can be confusing for distributed collections.
2. Created specific interfaces for each Dataset function (e.g. MapFunction, ReduceFunction, MapPartitionsFunction)
3. Added more documentation and test cases.
The other thing I'm considering doing is to have a "collector" interface for FlatMapFunction and MapPartitionsFunction, similar to MapReduce's map function.
Author: Reynold Xin <rxin@databricks.com>
Closes#9531 from rxin/SPARK-11564.
This PR adds support for multiple column in a single count distinct aggregate to the new aggregation path.
cc yhuai
Author: Herman van Hovell <hvanhovell@questtec.nl>
Closes#9409 from hvanhovell/SPARK-11451.
This PR enables the Expand operator to process and produce Unsafe Rows.
Author: Herman van Hovell <hvanhovell@questtec.nl>
Closes#9414 from hvanhovell/SPARK-11450.
https://issues.apache.org/jira/browse/SPARK-10116
This is really trivial, just happened to notice it -- if `XORShiftRandom.hashSeed` is really supposed to have random bits throughout (as the comment implies), it needs to do something for the conversion to `long`.
mengxr mkolod
Author: Imran Rashid <irashid@cloudera.com>
Closes#8314 from squito/SPARK-10116.
This PR adds test cases that test various column pruning and filter push-down cases.
Author: Cheng Lian <lian@databricks.com>
Closes#9468 from liancheng/spark-10978.follow-up.
JIRA: https://issues.apache.org/jira/browse/SPARK-9162
Currently ScalaUDF extends CodegenFallback and doesn't provide code generation implementation. This path implements code generation for ScalaUDF.
Author: Liang-Chi Hsieh <viirya@appier.com>
Closes#9270 from viirya/scalaudf-codegen.
This PR adds the ability to do typed SQL aggregations. We will likely also want to provide an interface to allow users to do aggregations on objects, but this is deferred to another PR.
```scala
val ds = Seq(("a", 10), ("a", 20), ("b", 1), ("b", 2), ("c", 1)).toDS()
ds.groupBy(_._1).agg(sum("_2").as[Int]).collect()
res0: Array(("a", 30), ("b", 3), ("c", 1))
```
Author: Michael Armbrust <michael@databricks.com>
Closes#9499 from marmbrus/dataset-agg.
the main problem is: we interpret column name with special handling of `.` for DataFrame. This enables us to write something like `df("a.b")` to get the field `b` of `a`. However, we don't need this feature in `DataFrame.apply("*")` or `DataFrame.withColumnRenamed`. In these 2 cases, the column name is the final name already, we don't need extra process to interpret it.
The solution is simple, use `queryExecution.analyzed.output` to get resolved column directly, instead of using `DataFrame.resolve`.
close https://github.com/apache/spark/pull/8811
Author: Wenchen Fan <wenchen@databricks.com>
Closes#9462 from cloud-fan/special-chars.
After aggregation, the dataset could be smaller than inputs, so it's better to do hash based aggregation for all inputs, then using sort based aggregation to merge them.
Author: Davies Liu <davies@databricks.com>
Closes#9383 from davies/fix_switch.
We have some aggregate function tests in both DataFrameAggregateSuite and SQLQuerySuite. The two have almost the same coverage and we should just remove the SQL one.
Author: Reynold Xin <rxin@databricks.com>
Closes#9475 from rxin/SPARK-11510.
1. Renamed localSort -> sortWithinPartitions to avoid ambiguity in "local"
2. distributeBy -> repartition to match the existing repartition.
Author: Reynold Xin <rxin@databricks.com>
Closes#9470 from rxin/SPARK-11504.
stddev is an alias for stddev_samp. variance should be consistent with stddev.
Also took the chance to remove internal Stddev and Variance, and only kept StddevSamp/StddevPop and VarianceSamp/VariancePop.
Author: Reynold Xin <rxin@databricks.com>
Closes#9449 from rxin/SPARK-11490.
depend on `caseSensitive` to do column name equality check, instead of just `==`
Author: Wenchen Fan <wenchen@databricks.com>
Closes#9410 from cloud-fan/partition.
We added a bunch of higher order statistics such as skewness and kurtosis to GroupedData. I don't think they are common enough to justify being listed, since users can always use the normal statistics aggregate functions.
That is to say, after this change, we won't support
```scala
df.groupBy("key").kurtosis("colA", "colB")
```
However, we will still support
```scala
df.groupBy("key").agg(kurtosis(col("colA")), kurtosis(col("colB")))
```
Author: Reynold Xin <rxin@databricks.com>
Closes#9446 from rxin/SPARK-11489.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#9434 from cloud-fan/rdd2ds and squashes the following commits:
0892d72 [Wenchen Fan] support create Dataset from RDD
This PR adds a new method `unhandledFilters` to `BaseRelation`. Data sources which implement this method properly may avoid the overhead of defensive filtering done by Spark SQL.
Author: Cheng Lian <lian@databricks.com>
Closes#9399 from liancheng/spark-10978.unhandled-filters.
JIRA: https://issues.apache.org/jira/browse/SPARK-10304
This patch detects if the structure of partition directories is not valid.
The test cases are from #8547. Thanks zhzhan.
cc liancheng
Author: Liang-Chi Hsieh <viirya@appier.com>
Closes#8840 from viirya/detect_invalid_part_dir.
This PR adds a new method `groupBy(cols: Column*)` to `Dataset` that allows users to group using column expressions instead of a lambda function. Since the return type of these expressions is not known at compile time, we just set the key type as a generic `Row`. If the user would like to work the key in a type-safe way, they can call `grouped.asKey[Type]`, which is also added in this PR.
```scala
val ds = Seq(("a", 10), ("a", 20), ("b", 1), ("b", 2), ("c", 1)).toDS()
val grouped = ds.groupBy($"_1").asKey[String]
val agged = grouped.mapGroups { case (g, iter) =>
Iterator((g, iter.map(_._2).sum))
}
agged.collect()
res0: Array(("a", 30), ("b", 3), ("c", 1))
```
Author: Michael Armbrust <michael@databricks.com>
Closes#9359 from marmbrus/columnGroupBy and squashes the following commits:
bbcb03b [Michael Armbrust] Update DatasetSuite.scala
8fd2908 [Michael Armbrust] Update DatasetSuite.scala
0b0e2f8 [Michael Armbrust] [SPARK-11404] [SQL] Support for groupBy using column expressions
When we join 2 datasets, we will combine 2 encoders into a tupled one, and use it as the encoder for the jioned dataset. Assume both of the 2 encoders are flat, their `constructExpression`s both reference to the first element of input row. However, when we combine 2 encoders, the schema of input row changed, now the right encoder should reference to second element of input row. So we should rebind right encoder to let it know the new schema of input row before combine it.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#9391 from cloud-fan/join and squashes the following commits:
846d3ab [Wenchen Fan] rebind right encoder when join 2 datasets
1. Supporting expanding structs in Projections. i.e.
"SELECT s.*" where s is a struct type.
This is fixed by allowing the expand function to handle structs in addition to tables.
2. Supporting expanding * inside aggregate functions of structs.
"SELECT max(struct(col1, structCol.*))"
This requires recursively expanding the expressions. In this case, it it the aggregate
expression "max(...)" and we need to recursively expand its children inputs.
Author: Nong Li <nongli@gmail.com>
Closes#9343 from nongli/spark-11329.
…ering.
For cached tables, we can just maintain the partitioning and ordering from the
source relation.
Author: Nong Li <nongli@gmail.com>
Closes#9404 from nongli/spark-5354.
DISTRIBUTE BY allows the user to hash partition the data by specified exprs. It also allows for
optioning sorting within each resulting partition. There is no required relationship between the
exprs for partitioning and sorting (i.e. one does not need to be a prefix of the other).
This patch adds to APIs to DataFrames which can be used together to provide this functionality:
1. distributeBy() which partitions the data frame into a specified number of partitions using the
partitioning exprs.
2. localSort() which sorts each partition using the provided sorting exprs.
To get the DISTRIBUTE BY functionality, the user simply does: df.distributeBy(...).localSort(...)
Author: Nong Li <nongli@gmail.com>
Closes#9364 from nongli/spark-11410.
This PR fixes two issues:
1. `PhysicalRDD.outputsUnsafeRows` is always `false`
Thus a `ConvertToUnsafe` operator is often required even if the underlying data source relation does output `UnsafeRow`.
1. Internal/external row conversion for `HadoopFsRelation` is kinda messy
Currently we're using `HadoopFsRelation.needConversion` and [dirty type erasure hacks][1] to indicate whether the relation outputs external row or internal row and apply external-to-internal conversion when necessary. Basically, all builtin `HadoopFsRelation` data sources, i.e. Parquet, JSON, ORC, and Text output `InternalRow`, while typical external `HadoopFsRelation` data sources, e.g. spark-avro and spark-csv, output `Row`.
This PR adds a `private[sql]` interface method `HadoopFsRelation.buildInternalScan`, which by default invokes `HadoopFsRelation.buildScan` and converts `Row`s to `UnsafeRow`s (which are also `InternalRow`s). All builtin `HadoopFsRelation` data sources override this method and directly output `UnsafeRow`s. In this way, now `HadoopFsRelation` always produces `UnsafeRow`s. Thus `PhysicalRDD.outputsUnsafeRows` can be properly set by checking whether the underlying data source is a `HadoopFsRelation`.
A remaining question is that, can we assume that all non-builtin `HadoopFsRelation` data sources output external rows? At least all well known ones do so. However it's possible that some users implemented their own `HadoopFsRelation` data sources that leverages `InternalRow` and thus all those unstable internal data representations. If this assumption is safe, we can deprecate `HadoopFsRelation.needConversion` and cleanup some more conversion code (like [here][2] and [here][3]).
This PR supersedes #9125.
Follow-ups:
1. Makes JSON and ORC data sources output `UnsafeRow` directly
1. Makes `HiveTableScan` output `UnsafeRow` directly
This is related to 1 since ORC data source shares the same `Writable` unwrapping code with `HiveTableScan`.
[1]: https://github.com/apache/spark/blob/v1.5.1/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/parquet/ParquetRelation.scala#L353
[2]: https://github.com/apache/spark/blob/v1.5.1/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/DataSourceStrategy.scala#L331-L335
[3]: https://github.com/apache/spark/blob/v1.5.1/sql/core/src/main/scala/org/apache/spark/sql/sources/interfaces.scala#L630-L669
Author: Cheng Lian <lian@databricks.com>
Closes#9305 from liancheng/spark-11345.unsafe-hadoop-fs-relation.
Currently the empty line in json file will be parsed into Row with all null field values. But in json, "{}" represents a json object, empty line is supposed to be skipped.
Make a trivial change for this.
Author: Jeff Zhang <zjffdu@apache.org>
Closes#9211 from zjffdu/SPARK-11226.
When we cogroup 2 `GroupedIterator`s in `CoGroupedIterator`, if the right side is smaller, we will consume right data and keep the left data unchanged. Then we call `hasNext` which will call `left.hasNext`. This will make `GroupedIterator` generate an extra group as the previous one has not been comsumed yet.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#9346 from cloud-fan/cogroup and squashes the following commits:
9be67c8 [Wenchen Fan] SPARK-11393
When enabling mergedSchema and predicate filter, this fails since Parquet does not accept filters pushed down when the columns of the filters do not exist in the schema.
This is related with Parquet issue (https://issues.apache.org/jira/browse/PARQUET-389).
For now, it just simply disables predicate push down when using merged schema in this PR.
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#9327 from HyukjinKwon/SPARK-11103.
This PR introduce a mechanism to call spill() on those SQL operators that support spilling (for example, BytesToBytesMap, UnsafeExternalSorter and ShuffleExternalSorter) if there is not enough memory for execution. The preserved first page is needed anymore, so removed.
Other Spillable objects in Spark core (ExternalSorter and AppendOnlyMap) are not included in this PR, but those could benefit from this (trigger others' spilling).
The PrepareRDD may be not needed anymore, could be removed in follow up PR.
The following script will fail with OOM before this PR, finished in 150 seconds with 2G heap (also works in 1.5 branch, with similar duration).
```python
sqlContext.setConf("spark.sql.shuffle.partitions", "1")
df = sqlContext.range(1<<25).selectExpr("id", "repeat(id, 2) as s")
df2 = df.select(df.id.alias('id2'), df.s.alias('s2'))
j = df.join(df2, df.id==df2.id2).groupBy(df.id).max("id", "id2")
j.explain()
print j.count()
```
For thread-safety, here what I'm got:
1) Without calling spill(), the operators should only be used by single thread, no safety problems.
2) spill() could be triggered in two cases, triggered by itself, or by other operators. we can check trigger == this in spill(), so it's still in the same thread, so safety problems.
3) if it's triggered by other operators (right now cache will not trigger spill()), we only spill the data into disk when it's in scanning stage (building is finished), so the in-memory sorter or memory pages are read-only, we only need to synchronize the iterator and change it.
4) During scanning, the iterator will only use one record in one page, we can't free this page, because the downstream is currently using it (used by UnsafeRow or other objects). In BytesToBytesMap, we just skip the current page, and dump all others into disk. In UnsafeExternalSorter, we keep the page that is used by current record (having the same baseObject), free it when loading the next record. In ShuffleExternalSorter, the spill() will not trigger during scanning.
5) In order to avoid deadlock, we didn't call acquireMemory during spill (so we reused the pointer array in InMemorySorter).
Author: Davies Liu <davies@databricks.com>
Closes#9241 from davies/force_spill.
Before this PR, user has to consume the iterator of one group before process next group, or we will get into infinite loops.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#9330 from cloud-fan/group.
In some cases, we can broadcast the smaller relation in cartesian join, which improve the performance significantly.
Author: Cheng Hao <hao.cheng@intel.com>
Closes#8652 from chenghao-intel/cartesian.
This PR adds a new operation `joinWith` to a `Dataset`, which returns a `Tuple` for each pair where a given `condition` evaluates to true.
```scala
case class ClassData(a: String, b: Int)
val ds1 = Seq(ClassData("a", 1), ClassData("b", 2)).toDS()
val ds2 = Seq(("a", 1), ("b", 2)).toDS()
> ds1.joinWith(ds2, $"_1" === $"a").collect()
res0: Array((ClassData("a", 1), ("a", 1)), (ClassData("b", 2), ("b", 2)))
```
This operation is similar to the relation `join` function with one important difference in the result schema. Since `joinWith` preserves objects present on either side of the join, the result schema is similarly nested into a tuple under the column names `_1` and `_2`.
This type of join can be useful both for preserving type-safety with the original object types as well as working with relational data where either side of the join has column names in common.
## Required Changes to Encoders
In the process of working on this patch, several deficiencies to the way that we were handling encoders were discovered. Specifically, it turned out to be very difficult to `rebind` the non-expression based encoders to extract the nested objects from the results of joins (and also typed selects that return tuples).
As a result the following changes were made.
- `ClassEncoder` has been renamed to `ExpressionEncoder` and has been improved to also handle primitive types. Additionally, it is now possible to take arbitrary expression encoders and rewrite them into a single encoder that returns a tuple.
- All internal operations on `Dataset`s now require an `ExpressionEncoder`. If the users tries to pass a non-`ExpressionEncoder` in, an error will be thrown. We can relax this requirement in the future by constructing a wrapper class that uses expressions to project the row to the expected schema, shielding the users code from the required remapping. This will give us a nice balance where we don't force user encoders to understand attribute references and binding, but still allow our native encoder to leverage runtime code generation to construct specific encoders for a given schema that avoid an extra remapping step.
- Additionally, the semantics for different types of objects are now better defined. As stated in the `ExpressionEncoder` scaladoc:
- Classes will have their sub fields extracted by name using `UnresolvedAttribute` expressions
and `UnresolvedExtractValue` expressions.
- Tuples will have their subfields extracted by position using `BoundReference` expressions.
- Primitives will have their values extracted from the first ordinal with a schema that defaults
to the name `value`.
- Finally, the binding lifecycle for `Encoders` has now been unified across the codebase. Encoders are now `resolved` to the appropriate schema in the constructor of `Dataset`. This process replaces an unresolved expressions with concrete `AttributeReference` expressions. Binding then happens on demand, when an encoder is going to be used to construct an object. This closely mirrors the lifecycle for standard expressions when executing normal SQL or `DataFrame` queries.
Author: Michael Armbrust <michael@databricks.com>
Closes#9300 from marmbrus/datasets-tuples.
When sampling and then filtering DataFrame, the SQL Optimizer will push down filter into sample and produce wrong result. This is due to the sampler is calculated based on the original scope rather than the scope after filtering.
Author: Yanbo Liang <ybliang8@gmail.com>
Closes#9294 from yanboliang/spark-11303.
Currently, when a schema is inferred from a JSON file using sqlContext.read.json, the primitive object types are inferred as string, long, boolean, etc.
However, if the inferred type is too specific (JSON obviously does not enforce types itself), this can cause issues with merging dataframe schemas.
This pull request adds the option "primitivesAsString" to the JSON DataFrameReader which when true (defaults to false if not set) will infer all primitives as strings.
Below is an example usage of this new functionality.
```
val jsonDf = sqlContext.read.option("primitivesAsString", "true").json(sampleJsonFile)
scala> jsonDf.printSchema()
root
|-- bigInteger: string (nullable = true)
|-- boolean: string (nullable = true)
|-- double: string (nullable = true)
|-- integer: string (nullable = true)
|-- long: string (nullable = true)
|-- null: string (nullable = true)
|-- string: string (nullable = true)
```
Author: Stephen De Gennaro <stepheng@realitymine.com>
Closes#9249 from stephend-realitymine/stephend-primitives.
This patch refactors the MemoryManager class structure. After #9000, Spark had the following classes:
- MemoryManager
- StaticMemoryManager
- ExecutorMemoryManager
- TaskMemoryManager
- ShuffleMemoryManager
This is fairly confusing. To simplify things, this patch consolidates several of these classes:
- ShuffleMemoryManager and ExecutorMemoryManager were merged into MemoryManager.
- TaskMemoryManager is moved into Spark Core.
**Key changes and tasks**:
- [x] Merge ExecutorMemoryManager into MemoryManager.
- [x] Move pooling logic into Allocator.
- [x] Move TaskMemoryManager from `spark-unsafe` to `spark-core`.
- [x] Refactor the existing Tungsten TaskMemoryManager interactions so Tungsten code use only this and not both this and ShuffleMemoryManager.
- [x] Refactor non-Tungsten code to use the TaskMemoryManager instead of ShuffleMemoryManager.
- [x] Merge ShuffleMemoryManager into MemoryManager.
- [x] Move code
- [x] ~~Simplify 1/n calculation.~~ **Will defer to followup, since this needs more work.**
- [x] Port ShuffleMemoryManagerSuite tests.
- [x] Move classes from `unsafe` package to `memory` package.
- [ ] Figure out how to handle the hacky use of the memory managers in HashedRelation's broadcast variable construction.
- [x] Test porting and cleanup: several tests relied on mock functionality (such as `TestShuffleMemoryManager.markAsOutOfMemory`) which has been changed or broken during the memory manager consolidation
- [x] AbstractBytesToBytesMapSuite
- [x] UnsafeExternalSorterSuite
- [x] UnsafeFixedWidthAggregationMapSuite
- [x] UnsafeKVExternalSorterSuite
**Compatiblity notes**:
- This patch introduces breaking changes in `ExternalAppendOnlyMap`, which is marked as `DevloperAPI` (likely for legacy reasons): this class now cannot be used outside of a task.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9127 from JoshRosen/SPARK-10984.
This adds API for reading and writing text files, similar to SparkContext.textFile and RDD.saveAsTextFile.
```
SQLContext.read.text("/path/to/something.txt")
DataFrame.write.text("/path/to/write.txt")
```
Using the new Dataset API, this also supports
```
val ds: Dataset[String] = SQLContext.read.text("/path/to/something.txt").as[String]
```
Author: Reynold Xin <rxin@databricks.com>
Closes#9240 from rxin/SPARK-11274.
*This PR adds a new experimental API to Spark, tentitively named Datasets.*
A `Dataset` is a strongly-typed collection of objects that can be transformed in parallel using functional or relational operations. Example usage is as follows:
### Functional
```scala
> val ds: Dataset[Int] = Seq(1, 2, 3).toDS()
> ds.filter(_ % 1 == 0).collect()
res1: Array[Int] = Array(1, 2, 3)
```
### Relational
```scala
scala> ds.toDF().show()
+-----+
|value|
+-----+
| 1|
| 2|
| 3|
+-----+
> ds.select(expr("value + 1").as[Int]).collect()
res11: Array[Int] = Array(2, 3, 4)
```
## Comparison to RDDs
A `Dataset` differs from an `RDD` in the following ways:
- The creation of a `Dataset` requires the presence of an explicit `Encoder` that can be
used to serialize the object into a binary format. Encoders are also capable of mapping the
schema of a given object to the Spark SQL type system. In contrast, RDDs rely on runtime
reflection based serialization.
- Internally, a `Dataset` is represented by a Catalyst logical plan and the data is stored
in the encoded form. This representation allows for additional logical operations and
enables many operations (sorting, shuffling, etc.) to be performed without deserializing to
an object.
A `Dataset` can be converted to an `RDD` by calling the `.rdd` method.
## Comparison to DataFrames
A `Dataset` can be thought of as a specialized DataFrame, where the elements map to a specific
JVM object type, instead of to a generic `Row` container. A DataFrame can be transformed into
specific Dataset by calling `df.as[ElementType]`. Similarly you can transform a strongly-typed
`Dataset` to a generic DataFrame by calling `ds.toDF()`.
## Implementation Status and TODOs
This is a rough cut at the least controversial parts of the API. The primary purpose here is to get something committed so that we can better parallelize further work and get early feedback on the API. The following is being deferred to future PRs:
- Joins and Aggregations (prototype here f11f91e6f0)
- Support for Java
Additionally, the responsibility for binding an encoder to a given schema is currently done in a fairly ad-hoc fashion. This is an internal detail, and what we are doing today works for the cases we care about. However, as we add more APIs we'll probably need to do this in a more principled way (i.e. separate resolution from binding as we do in DataFrames).
## COMPATIBILITY NOTE
Long term we plan to make `DataFrame` extend `Dataset[Row]`. However,
making this change to che class hierarchy would break the function signatures for the existing
function operations (map, flatMap, etc). As such, this class should be considered a preview
of the final API. Changes will be made to the interface after Spark 1.6.
Author: Michael Armbrust <michael@databricks.com>
Closes#9190 from marmbrus/dataset-infra.
There's a lot of duplication between SortShuffleManager and UnsafeShuffleManager. Given that these now provide the same set of functionality, now that UnsafeShuffleManager supports large records, I think that we should replace SortShuffleManager's serialized shuffle implementation with UnsafeShuffleManager's and should merge the two managers together.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8829 from JoshRosen/consolidate-sort-shuffle-implementations.
This PR introduce a new feature to run SQL directly on files without create a table, for example:
```
select id from json.`path/to/json/files` as j
```
Author: Davies Liu <davies@databricks.com>
Closes#9173 from davies/source.
Due to PARQUET-251, `BINARY` columns in existing Parquet files may be written with corrupted statistics information. This information is used by filter push-down optimization. Since Spark 1.5 turns on Parquet filter push-down by default, we may end up with wrong query results. PARQUET-251 has been fixed in parquet-mr 1.8.1, but Spark 1.5 is still using 1.7.0.
This affects all Spark SQL data types that can be mapped to Parquet {{BINARY}}, namely:
- `StringType`
- `BinaryType`
- `DecimalType`
(But Spark SQL doesn't support pushing down filters involving `DecimalType` columns for now.)
To avoid wrong query results, we should disable filter push-down for columns of `StringType` and `BinaryType` until we upgrade to parquet-mr 1.8.
Author: Cheng Lian <lian@databricks.com>
Closes#9152 from liancheng/spark-11153.workaround-parquet-251.
(cherry picked from commit 0887e5e878)
Signed-off-by: Cheng Lian <lian@databricks.com>
Currently, we use CartesianProduct for join with null-safe-equal condition.
```
scala> sqlContext.sql("select * from t a join t b on (a.i <=> b.i)").explain
== Physical Plan ==
TungstenProject [i#2,j#3,i#7,j#8]
Filter (i#2 <=> i#7)
CartesianProduct
LocalTableScan [i#2,j#3], [[1,1]]
LocalTableScan [i#7,j#8], [[1,1]]
```
Actually, we can have an equal-join condition as `coalesce(i, default) = coalesce(b.i, default)`, then an partitioned join algorithm could be used.
After this PR, the plan will become:
```
>>> sqlContext.sql("select * from a join b ON a.id <=> b.id").explain()
TungstenProject [id#0L,id#1L]
Filter (id#0L <=> id#1L)
SortMergeJoin [coalesce(id#0L,0)], [coalesce(id#1L,0)]
TungstenSort [coalesce(id#0L,0) ASC], false, 0
TungstenExchange hashpartitioning(coalesce(id#0L,0),200)
ConvertToUnsafe
Scan PhysicalRDD[id#0L]
TungstenSort [coalesce(id#1L,0) ASC], false, 0
TungstenExchange hashpartitioning(coalesce(id#1L,0),200)
ConvertToUnsafe
Scan PhysicalRDD[id#1L]
```
Author: Davies Liu <davies@databricks.com>
Closes#9120 from davies/null_safe.
The purpose of this PR is to keep the unsafe format detail only inside the unsafe class itself, so when we use them(like use unsafe array in unsafe map, use unsafe array and map in columnar cache), we don't need to understand the format before use them.
change list:
* unsafe array's 4-bytes numElements header is now required(was optional), and become a part of unsafe array format.
* w.r.t the previous changing, the `sizeInBytes` of unsafe array now counts the 4-bytes header.
* unsafe map's format was `[numElements] [key array numBytes] [key array content(without numElements header)] [value array content(without numElements header)]` before, which is a little hacky as it makes unsafe array's header optional. I think saving 4 bytes is not a big deal, so the format is now: `[key array numBytes] [unsafe key array] [unsafe value array]`.
* w.r.t the previous changing, the `sizeInBytes` of unsafe map now counts both map's header and array's header.
Author: Wenchen Fan <wenchen@databricks.com>
Closes#9131 from cloud-fan/unsafe.
The unit test added in #9132 is flaky. This is a follow up PR to add `listenerBus.waitUntilEmpty` to fix it.
Author: zsxwing <zsxwing@gmail.com>
Closes#9163 from zsxwing/SPARK-11126-follow-up.
SQLListener adds all stage infos to `_stageIdToStageMetrics`, but only removes stage infos belonging to SQL executions. This PR fixed it by ignoring stages that don't belong to SQL executions.
Reported by Terry Hoo in https://www.mail-archive.com/userspark.apache.org/msg38810.html
Author: zsxwing <zsxwing@gmail.com>
Closes#9132 from zsxwing/SPARK-11126.
Make sure comma-separated paths get processed correcly in ResolvedDataSource for a HadoopFsRelationProvider
Author: Koert Kuipers <koert@tresata.com>
Closes#8416 from koertkuipers/feat-sql-comma-separated-paths.
In Spark SQL, the Exchange planner tries to avoid unnecessary sorts in cases where the data has already been sorted by a superset of the requested sorting columns. For instance, let's say that a query calls for an operator's input to be sorted by `a.asc` and the input happens to already be sorted by `[a.asc, b.asc]`. In this case, we do not need to re-sort the input. The converse, however, is not true: if the query calls for `[a.asc, b.asc]`, then `a.asc` alone will not satisfy the ordering requirements, requiring an additional sort to be planned by Exchange.
However, the current Exchange code gets this wrong and incorrectly skips sorting when the existing output ordering is a subset of the required ordering. This is simple to fix, however.
This bug was introduced in https://github.com/apache/spark/pull/7458, so it affects 1.5.0+.
This patch fixes the bug and significantly improves the unit test coverage of Exchange's sort-planning logic.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9140 from JoshRosen/SPARK-11135.
#9084 uncovered that many tests that test spilling don't actually spill. This is a follow-up patch to fix that to ensure our unit tests actually catch potential bugs in spilling. The size of this patch is inflated by the refactoring of `ExternalSorterSuite`, which had a lot of duplicate code and logic.
Author: Andrew Or <andrew@databricks.com>
Closes#9124 from andrewor14/spilling-tests.
This patch extends TungstenAggregate to support ImperativeAggregate functions. The existing TungstenAggregate operator only supported DeclarativeAggregate functions, which are defined in terms of Catalyst expressions and can be evaluated via generated projections. ImperativeAggregate functions, on the other hand, are evaluated by calling their `initialize`, `update`, `merge`, and `eval` methods.
The basic strategy here is similar to how SortBasedAggregate evaluates both types of aggregate functions: use a generated projection to evaluate the expression-based declarative aggregates with dummy placeholder expressions inserted in place of the imperative aggregate function output, then invoke the imperative aggregate functions and target them against the aggregation buffer. The bulk of the diff here consists of code that was copied and adapted from SortBasedAggregate, with some key changes to handle TungstenAggregate's sort fallback path.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9038 from JoshRosen/support-interpreted-in-tungsten-agg-final.
```scala
withSQLConf(SQLConf.PARQUET_FILTER_PUSHDOWN_ENABLED.key -> "true") {
withTempPath { dir =>
val path = s"${dir.getCanonicalPath}/part=1"
(1 to 3).map(i => (i, i.toString)).toDF("a", "b").write.parquet(path)
// If the "part = 1" filter gets pushed down, this query will throw an exception since
// "part" is not a valid column in the actual Parquet file
checkAnswer(
sqlContext.read.parquet(path).filter("a > 0 and (part = 0 or a > 1)"),
(2 to 3).map(i => Row(i, i.toString, 1)))
}
}
```
We expect the result to be:
```
2,1
3,1
```
But got
```
1,1
2,1
3,1
```
Author: Cheng Hao <hao.cheng@intel.com>
Closes#8916 from chenghao-intel/partition_filter.
Right now, we have QualifiedTableName, TableIdentifier, and Seq[String] to represent table identifiers. We should only have one form and TableIdentifier is the best one because it provides methods to get table name, database name, return unquoted string, and return quoted string.
Author: Wenchen Fan <wenchen@databricks.com>
Author: Wenchen Fan <cloud0fan@163.com>
Closes#8453 from cloud-fan/table-name.
With this feature, we can track the query plan, time cost, exception during query execution for spark users.
Author: Wenchen Fan <cloud0fan@163.com>
Closes#9078 from cloud-fan/callback.
We should not stop resolving having when the having condtion is resolved, or something like `count(1)` will crash.
Author: Wenchen Fan <cloud0fan@163.com>
Closes#9105 from cloud-fan/having.
This patch unifies the memory management of the storage and execution regions such that either side can borrow memory from each other. When memory pressure arises, storage will be evicted in favor of execution. To avoid regressions in cases where storage is crucial, we dynamically allocate a fraction of space for storage that execution cannot evict. Several configurations are introduced:
- **spark.memory.fraction (default 0.75)**: fraction of the heap space used for execution and storage. The lower this is, the more frequently spills and cached data eviction occur. The purpose of this config is to set aside memory for internal metadata, user data structures, and imprecise size estimation in the case of sparse, unusually large records.
- **spark.memory.storageFraction (default 0.5)**: size of the storage region within the space set aside by `spark.memory.fraction`. Cached data may only be evicted if total storage exceeds this region.
- **spark.memory.useLegacyMode (default false)**: whether to use the memory management that existed in Spark 1.5 and before. This is mainly for backward compatibility.
For a detailed description of the design, see [SPARK-10000](https://issues.apache.org/jira/browse/SPARK-10000). This patch builds on top of the `MemoryManager` interface introduced in #9000.
Author: Andrew Or <andrew@databricks.com>
Closes#9084 from andrewor14/unified-memory-manager.
The SQLTab will be shared by multiple sessions.
If we create multiple independent SQLContexts (not using newSession()), will still see multiple SQLTabs in the Spark UI.
Author: Davies Liu <davies@databricks.com>
Closes#9048 from davies/sqlui.
This PR improve the unrolling and read of complex types in columnar cache:
1) Using UnsafeProjection to do serialization of complex types, so they will not be serialized three times (two for actualSize)
2) Copy the bytes from UnsafeRow/UnsafeArrayData to ByteBuffer directly, avoiding the immediate byte[]
3) Using the underlying array in ByteBuffer to create UTF8String/UnsafeRow/UnsafeArrayData without copy.
Combine these optimizations, we can reduce the unrolling time from 25s to 21s (20% less), reduce the scanning time from 3.5s to 2.5s (28% less).
```
df = sqlContext.read.parquet(path)
t = time.time()
df.cache()
df.count()
print 'unrolling', time.time() - t
for i in range(10):
t = time.time()
print df.select("*")._jdf.queryExecution().toRdd().count()
print time.time() - t
```
The schema is
```
root
|-- a: struct (nullable = true)
| |-- b: long (nullable = true)
| |-- c: string (nullable = true)
|-- d: array (nullable = true)
| |-- element: long (containsNull = true)
|-- e: map (nullable = true)
| |-- key: long
| |-- value: string (valueContainsNull = true)
```
Now the columnar cache depends on that UnsafeProjection support all the data types (including UDT), this PR also fix that.
Author: Davies Liu <davies@databricks.com>
Closes#9016 from davies/complex2.
For Parquet decimal columns that are encoded using plain-dictionary encoding, we can make the upper level converter aware of the dictionary, so that we can pre-instantiate all the decimals to avoid duplicated instantiation.
Note that plain-dictionary encoding isn't available for `FIXED_LEN_BYTE_ARRAY` for Parquet writer version `PARQUET_1_0`. So currently only decimals written as `INT32` and `INT64` can benefit from this optimization.
Author: Cheng Lian <lian@databricks.com>
Closes#9040 from liancheng/spark-11007.decimal-converter-dict-support.
marmbrus
rxin
This patch adds a JdbcDialect class, which customizes the datatype mappings for Derby backends. The patch also adds unit tests for the new dialect, corresponding to the existing tests for other JDBC dialects.
JDBCSuite runs cleanly for me with this patch. So does JDBCWriteSuite, although it produces noise as described here: https://issues.apache.org/jira/browse/SPARK-10890
This patch is my original work, which I license to the ASF. I am a Derby contributor, so my ICLA is on file under SVN id "rhillegas": http://people.apache.org/committer-index.html
Touches the following files:
---------------------------------
org.apache.spark.sql.jdbc.JdbcDialects
Adds a DerbyDialect.
---------------------------------
org.apache.spark.sql.jdbc.JDBCSuite
Adds unit tests for the new DerbyDialect.
Author: Rick Hillegas <rhilleg@us.ibm.com>
Closes#8982 from rick-ibm/b_10855.
This patch introduces a `MemoryManager` that is the central arbiter of how much memory to grant to storage and execution. This patch is primarily concerned only with refactoring while preserving the existing behavior as much as possible.
This is the first step away from the existing rigid separation of storage and execution memory, which has several major drawbacks discussed on the [issue](https://issues.apache.org/jira/browse/SPARK-10956). It is the precursor of a series of patches that will attempt to address those drawbacks.
Author: Andrew Or <andrew@databricks.com>
Author: Josh Rosen <joshrosen@databricks.com>
Author: andrewor14 <andrew@databricks.com>
Closes#9000 from andrewor14/memory-manager.
This PR improve the sessions management by replacing the thread-local based to one SQLContext per session approach, introduce separated temporary tables and UDFs/UDAFs for each session.
A new session of SQLContext could be created by:
1) create an new SQLContext
2) call newSession() on existing SQLContext
For HiveContext, in order to reduce the cost for each session, the classloader and Hive client are shared across multiple sessions (created by newSession).
CacheManager is also shared by multiple sessions, so cache a table multiple times in different sessions will not cause multiple copies of in-memory cache.
Added jars are still shared by all the sessions, because SparkContext does not support sessions.
cc marmbrus yhuai rxin
Author: Davies Liu <davies@databricks.com>
Closes#8909 from davies/sessions.
UnsafeRow contains 3 pieces of information when pointing to some data in memory (an object, a base offset, and length). When the row is serialized with Java/Kryo serialization, the object layout in memory can change if two machines have different pointer width (Oops in JVM).
To reproduce, launch Spark using
MASTER=local-cluster[2,1,1024] bin/spark-shell --conf "spark.executor.extraJavaOptions=-XX:-UseCompressedOops"
And then run the following
scala> sql("select 1 xx").collect()
Author: Reynold Xin <rxin@databricks.com>
Closes#9030 from rxin/SPARK-10914.
This PR refactors Parquet write path to follow parquet-format spec. It's a successor of PR #7679, but with less non-essential changes.
Major changes include:
1. Replaces `RowWriteSupport` and `MutableRowWriteSupport` with `CatalystWriteSupport`
- Writes Parquet data using standard layout defined in parquet-format
Specifically, we are now writing ...
- ... arrays and maps in standard 3-level structure with proper annotations and field names
- ... decimals as `INT32` and `INT64` whenever possible, and taking `FIXED_LEN_BYTE_ARRAY` as the final fallback
- Supports legacy mode which is compatible with Spark 1.4 and prior versions
The legacy mode is by default off, and can be turned on by flipping SQL option `spark.sql.parquet.writeLegacyFormat` to `true`.
- Eliminates per value data type dispatching costs via prebuilt composed writer functions
1. Cleans up the last pieces of old Parquet support code
As pointed out by rxin previously, we probably want to rename all those `Catalyst*` Parquet classes to `Parquet*` for clarity. But I'd like to do this in a follow-up PR to minimize code review noises in this one.
Author: Cheng Lian <lian@databricks.com>
Closes#8988 from liancheng/spark-8848/standard-parquet-write-path.
In `aggregate/utils.scala`, there is a substantial amount of duplication in the expression-rewriting logic. As a prerequisite to supporting imperative aggregate functions in `TungstenAggregate`, this patch refactors this file so that the same expression-rewriting logic is used for both `SortAggregate` and `TungstenAggregate`.
In order to allow both operators to use the same rewriting logic, `TungstenAggregationIterator. generateResultProjection()` has been updated so that it first evaluates all declarative aggregate functions' `evaluateExpression`s and writes the results into a temporary buffer, and then uses this temporary buffer and the grouping expressions to evaluate the final resultExpressions. This matches the logic in SortAggregateIterator, where this two-pass approach is necessary in order to support imperative aggregates. If this change turns out to cause performance regressions, then we can look into re-implementing the single-pass evaluation in a cleaner way as part of a followup patch.
Since the rewriting logic is now shared across both operators, this patch also extracts that logic and places it in `SparkStrategies`. This makes the rewriting logic a bit easier to follow, I think.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#9015 from JoshRosen/SPARK-10988.
This PR refactors `HashJoinNode` to take a existing `HashedRelation`. So, we can reuse this node for both `ShuffledHashJoin` and `BroadcastHashJoin`.
https://issues.apache.org/jira/browse/SPARK-10887
Author: Yin Huai <yhuai@databricks.com>
Closes#8953 from yhuai/SPARK-10887.
This PR addresses [SPARK-7869](https://issues.apache.org/jira/browse/SPARK-7869)
Before the patch, attempt to load the table from Postgres with JSON/JSONb datatype caused error `java.sql.SQLException: Unsupported type 1111`
Postgres data types JSON and JSONb are now mapped to String on Spark side thus they can be loaded into DF and processed on Spark side
Example
Postgres:
```
create table test_json (id int, value json);
create table test_jsonb (id int, value jsonb);
insert into test_json (id, value) values
(1, '{"field1":"value1","field2":"value2","field3":[1,2,3]}'::json),
(2, '{"field1":"value3","field2":"value4","field3":[4,5,6]}'::json),
(3, '{"field3":"value5","field4":"value6","field3":[7,8,9]}'::json);
insert into test_jsonb (id, value) values
(4, '{"field1":"value1","field2":"value2","field3":[1,2,3]}'::jsonb),
(5, '{"field1":"value3","field2":"value4","field3":[4,5,6]}'::jsonb),
(6, '{"field3":"value5","field4":"value6","field3":[7,8,9]}'::jsonb);
```
PySpark:
```
>>> import json
>>> df1 = sqlContext.read.jdbc("jdbc:postgresql://127.0.0.1:5432/test?user=testuser", "test_json")
>>> df1.map(lambda x: (x.id, json.loads(x.value))).map(lambda (id, value): (id, value.get('field3'))).collect()
[(1, [1, 2, 3]), (2, [4, 5, 6]), (3, [7, 8, 9])]
>>> df2 = sqlContext.read.jdbc("jdbc:postgresql://127.0.0.1:5432/test?user=testuser", "test_jsonb")
>>> df2.map(lambda x: (x.id, json.loads(x.value))).map(lambda (id, value): (id, value.get('field1'))).collect()
[(4, u'value1'), (5, u'value3'), (6, None)]
```
Author: 0x0FFF <programmerag@gmail.com>
Closes#8948 from 0x0FFF/SPARK-7869.
This PR improve the performance of complex types in columnar cache by using UnsafeProjection instead of KryoSerializer.
A simple benchmark show that this PR could improve the performance of scanning a cached table with complex columns by 15x (comparing to Spark 1.5).
Here is the code used to benchmark:
```
df = sc.range(1<<23).map(lambda i: Row(a=Row(b=i, c=str(i)), d=range(10), e=dict(zip(range(10), [str(i) for i in range(10)])))).toDF()
df.write.parquet("table")
```
```
df = sqlContext.read.parquet("table")
df.cache()
df.count()
t = time.time()
print df.select("*")._jdf.queryExecution().toRdd().count()
print time.time() - t
```
Author: Davies Liu <davies@databricks.com>
Closes#8971 from davies/complex.
This patch refactors several of the Aggregate2 interfaces in order to improve code clarity.
The biggest change is a refactoring of the `AggregateFunction2` class hierarchy. In the old code, we had a class named `AlgebraicAggregate` that inherited from `AggregateFunction2`, added a new set of methods, then banned the use of the inherited methods. I found this to be fairly confusing because.
If you look carefully at the existing code, you'll see that subclasses of `AggregateFunction2` fall into two disjoint categories: imperative aggregation functions which directly extended `AggregateFunction2` and declarative, expression-based aggregate functions which extended `AlgebraicAggregate`. In order to make this more explicit, this patch refactors things so that `AggregateFunction2` is a sealed abstract class with two subclasses, `ImperativeAggregateFunction` and `ExpressionAggregateFunction`. The superclass, `AggregateFunction2`, now only contains methods and fields that are common to both subclasses.
After making this change, I updated the various AggregationIterator classes to comply with this new naming scheme. I also performed several small renamings in the aggregate interfaces themselves in order to improve clarity and rewrote or expanded a number of comments.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#8973 from JoshRosen/tungsten-agg-comments.