## What changes were proposed in this pull request?
This PR implements `EXPLAIN CODEGEN` SQL command which returns generated codes like `debugCodegen`. In `spark-shell`, we don't need to `import debug` module. In `spark-sql`, we can use this SQL command now.
**Before**
```
scala> import org.apache.spark.sql.execution.debug._
scala> sql("select 'a' as a group by 1").debugCodegen()
Found 2 WholeStageCodegen subtrees.
== Subtree 1 / 2 ==
...
Generated code:
...
== Subtree 2 / 2 ==
...
Generated code:
...
```
**After**
```
scala> sql("explain extended codegen select 'a' as a group by 1").collect().foreach(println)
[Found 2 WholeStageCodegen subtrees.]
[== Subtree 1 / 2 ==]
...
[]
[Generated code:]
...
[]
[== Subtree 2 / 2 ==]
...
[]
[Generated code:]
...
```
## How was this patch tested?
Pass the Jenkins tests (including new testcases)
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#12099 from dongjoon-hyun/SPARK-14251.
## What changes were proposed in this pull request?
This PR reduces Java byte code size of method in ```SpecificColumnarIterator``` by using a approach to make a group for lot of ```ColumnAccessor``` instantiations or method calls (more than 200) into a method
## How was this patch tested?
Added a new unit test, which includes large instantiations and method calls, to ```InMemoryColumnarQuerySuite```
Author: Kazuaki Ishizaki <ishizaki@jp.ibm.com>
Closes#12108 from kiszk/SPARK-14138-master.
## What changes were proposed in this pull request?
`SizeBasedWindowFunction.n` is a global singleton attribute created for evaluating size based aggregate window functions like `CUME_DIST`. However, this attribute gets different expression IDs when created on both driver side and executor side. This PR adds `withPartitionSize` method to `SizeBasedWindowFunction` so that we can easily rewrite `SizeBasedWindowFunction.n` on executor side.
## How was this patch tested?
A test case is added in `HiveSparkSubmitSuite`, which supports launching multi-process clusters.
Author: Cheng Lian <lian@databricks.com>
Closes#12040 from liancheng/spark-14244-fix-sized-window-function.
## What changes were proposed in this pull request?
Decision tree helper classes will be migrated to ML. This patch moves those internal classes that are not part of the public API and removes ones that are no longer used, after [SPARK-12183](https://github.com/apache/spark/pull/11855). No functional changes are made.
Details:
* Bin.scala is removed as the ML implementation does not require bins
* mllib NodeIdCache is removed. It was only used by the mllib implementation previously, which no longer exists
* mllib TreePoint is removed. It was only used by the mllib implementation previously, which no longer exists
* BaggedPoint, DTStatsAggregator, DecisionTreeMetadata, BaggedPointSuite and TimeTracker are all moved to ML.
## How was this patch tested?
No functional changes are made. Existing unit tests ensure behavior is unchanged.
Author: sethah <seth.hendrickson16@gmail.com>
Closes#12097 from sethah/cleanup_mllib_tree.
Currently, the Predictor abstraction expects the input labelCol type to be DoubleType, but we should support other numeric types. This will involve updating the PredictorParams.validateAndTransformSchema method.
Author: BenFradet <benjamin.fradet@gmail.com>
Closes#10355 from BenFradet/SPARK-7425.
## What changes were proposed in this pull request?
Adding long values for each Date in the ApplicationAttemptInfo API for easier use in code
## How was the this patch tested?
Tested with dev/run-tests
Author: Alex Bozarth <ajbozart@us.ibm.com>
Closes#11326 from ajbozarth/spark13241.
## What changes were proposed in this pull request?
Currently the Streaming tab in web UI uses records and events interchangeably; this PR tries to standardize them on "records".
"records" is chosen over "events" because:
- "records" is used extensively throughout the streaming documents, codes, and comments
- "events" is used only in Streaming UI related codes and comments
## How was this patch tested?
- existing test suites
- manually checking on the Streaming UI tab
Author: Liwei Lin <lwlin7@gmail.com>
Closes#12032 from lw-lin/streaming-events-to-records.
## What changes were proposed in this pull request?
Upgrade to 2.11.8 (from the current 2.11.7)
## How was this patch tested?
A manual build
Author: Jacek Laskowski <jacek@japila.pl>
Closes#11681 from jaceklaskowski/SPARK-13825-scala-2_11_8.
This PR adds the ability to perform aggregations inside of a `ContinuousQuery`. In order to implement this feature, the planning of aggregation has augmented with a new `StatefulAggregationStrategy`. Unlike batch aggregation, stateful-aggregation uses the `StateStore` (introduced in #11645) to persist the results of partial aggregation across different invocations. The resulting physical plan performs the aggregation using the following progression:
- Partial Aggregation
- Shuffle
- Partial Merge (now there is at most 1 tuple per group)
- StateStoreRestore (now there is 1 tuple from this batch + optionally one from the previous)
- Partial Merge (now there is at most 1 tuple per group)
- StateStoreSave (saves the tuple for the next batch)
- Complete (output the current result of the aggregation)
The following refactoring was also performed to allow us to plug into existing code:
- The get/put implementation is taken from #12013
- The logic for breaking down and de-duping the physical execution of aggregation has been move into a new pattern `PhysicalAggregation`
- The `AttributeReference` used to identify the result of an `AggregateFunction` as been moved into the `AggregateExpression` container. This change moves the reference into the same object as the other intermediate references used in aggregation and eliminates the need to pass around a `Map[(AggregateFunction, Boolean), Attribute]`. Further clean up (using a different aggregation container for logical/physical plans) is deferred to a followup.
- Some planning logic is moved from the `SessionState` into the `QueryExecution` to make it easier to override in the streaming case.
- The ability to write a `StreamTest` that checks only the output of the last batch has been added to simulate the future addition of output modes.
Author: Michael Armbrust <michael@databricks.com>
Closes#12048 from marmbrus/statefulAgg.
## What changes were proposed in this pull request?
RpcEndpoint is not thread safe and allows multiple messages to be processed at the same time. StateStoreCoordinator should use ThreadSafeRpcEndpoint.
## How was this patch tested?
Existing unit tests.
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#12100 from zsxwing/fix-StateStoreCoordinator.
This patch adds support for caching blocks in the executor processes using direct / off-heap memory.
## User-facing changes
**Updated semantics of `OFF_HEAP` storage level**: In Spark 1.x, the `OFF_HEAP` storage level indicated that an RDD should be cached in Tachyon. Spark 2.x removed the external block store API that Tachyon caching was based on (see #10752 / SPARK-12667), so `OFF_HEAP` became an alias for `MEMORY_ONLY_SER`. As of this patch, `OFF_HEAP` means "serialized and cached in off-heap memory or on disk". Via the `StorageLevel` constructor, `useOffHeap` can be set if `serialized == true` and can be used to construct custom storage levels which support replication.
**Storage UI reporting**: the storage UI will now report whether in-memory blocks are stored on- or off-heap.
**Only supported by UnifiedMemoryManager**: for simplicity, this feature is only supported when the default UnifiedMemoryManager is used; applications which use the legacy memory manager (`spark.memory.useLegacyMode=true`) are not currently able to allocate off-heap storage memory, so using off-heap caching will fail with an error when legacy memory management is enabled. Given that we plan to eventually remove the legacy memory manager, this is not a significant restriction.
**Memory management policies:** the policies for dividing available memory between execution and storage are the same for both on- and off-heap memory. For off-heap memory, the total amount of memory available for use by Spark is controlled by `spark.memory.offHeap.size`, which is an absolute size. Off-heap storage memory obeys `spark.memory.storageFraction` in order to control the amount of unevictable storage memory. For example, if `spark.memory.offHeap.size` is 1 gigabyte and Spark uses the default `storageFraction` of 0.5, then up to 500 megabytes of off-heap cached blocks will be protected from eviction due to execution memory pressure. If necessary, we can split `spark.memory.storageFraction` into separate on- and off-heap configurations, but this doesn't seem necessary now and can be done later without any breaking changes.
**Use of off-heap memory does not imply use of off-heap execution (or vice-versa)**: for now, the settings controlling the use of off-heap execution memory (`spark.memory.offHeap.enabled`) and off-heap caching are completely independent, so Spark SQL can be configured to use off-heap memory for execution while continuing to cache blocks on-heap. If desired, we can change this in a followup patch so that `spark.memory.offHeap.enabled` affect the default storage level for cached SQL tables.
## Internal changes
- Rename `ByteArrayChunkOutputStream` to `ChunkedByteBufferOutputStream`
- It now returns a `ChunkedByteBuffer` instead of an array of byte arrays.
- Its constructor now accept an `allocator` function which is called to allocate `ByteBuffer`s. This allows us to control whether it allocates regular ByteBuffers or off-heap DirectByteBuffers.
- Because block serialization is now performed during the unroll process, a `ChunkedByteBufferOutputStream` which is configured with a `DirectByteBuffer` allocator will use off-heap memory for both unroll and storage memory.
- The `MemoryStore`'s MemoryEntries now tracks whether blocks are stored on- or off-heap.
- `evictBlocksToFreeSpace()` now accepts a `MemoryMode` parameter so that we don't try to evict off-heap blocks in response to on-heap memory pressure (or vice-versa).
- Make sure that off-heap buffers are properly de-allocated during MemoryStore eviction.
- The JVM limits the total size of allocated direct byte buffers using the `-XX:MaxDirectMemorySize` flag and the default tends to be fairly low (< 512 megabytes in some JVMs). To work around this limitation, this patch adds a custom DirectByteBuffer allocator which ignores this memory limit.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#11805 from JoshRosen/off-heap-caching.
Currently, when max number of executor failures reached the `maxNumExecutorFailures`, `ApplicationMaster` will be killed and re-register another one.This time, `YarnAllocator` will be created a new instance.
But, the value of property `executorIdCounter` in `YarnAllocator` will reset to `0`. Then the Id of new executor will starting from `1`. This will confuse with the executor has already created before, which will cause FetchFailedException.
This situation is just in yarn client mode, so this is an issue in yarn client mode. For more details, [link to jira issues SPARK-12864](https://issues.apache.org/jira/browse/SPARK-12864)
This PR introduce a mechanism to initialize `executorIdCounter` after `ApplicationMaster` killed.
Author: zhonghaihua <793507405@qq.com>
Closes#10794 from zhonghaihua/initExecutorIdCounterAfterAMKilled.
JIRA: https://issues.apache.org/jira/browse/SPARK-13674
## What changes were proposed in this pull request?
Sample operator doesn't support wholestage codegen now. This pr is to add support to it.
## How was this patch tested?
A test is added into `BenchmarkWholeStageCodegen`. Besides, all tests should be passed.
Author: Liang-Chi Hsieh <simonh@tw.ibm.com>
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes#11517 from viirya/add-wholestage-sample.
## What changes were proposed in this pull request?
This PR adds the function `window` as a column expression.
`window` can be used to bucket rows into time windows given a time column. With this expression, performing time series analysis on batch data, as well as streaming data should become much more simpler.
### Usage
Assume the following schema:
`sensor_id, measurement, timestamp`
To average 5 minute data every 1 minute (window length of 5 minutes, slide duration of 1 minute), we will use:
```scala
df.groupBy(window("timestamp", “5 minutes”, “1 minute”), "sensor_id")
.agg(mean("measurement").as("avg_meas"))
```
This will generate windows such as:
```
09:00:00-09:05:00
09:01:00-09:06:00
09:02:00-09:07:00 ...
```
Intervals will start at every `slideDuration` starting at the unix epoch (1970-01-01 00:00:00 UTC).
To start intervals at a different point of time, e.g. 30 seconds after a minute, the `startTime` parameter can be used.
```scala
df.groupBy(window("timestamp", “5 minutes”, “1 minute”, "30 second"), "sensor_id")
.agg(mean("measurement").as("avg_meas"))
```
This will generate windows such as:
```
09:00:30-09:05:30
09:01:30-09:06:30
09:02:30-09:07:30 ...
```
Support for Python will be made in a follow up PR after this.
## How was this patch tested?
This patch has some basic unit tests for the `TimeWindow` expression testing that the parameters pass validation, and it also has some unit/integration tests testing the correctness of the windowing and usability in complex operations (multi-column grouping, multi-column projections, joins).
Author: Burak Yavuz <brkyvz@gmail.com>
Author: Michael Armbrust <michael@databricks.com>
Closes#12008 from brkyvz/df-time-window.
## What changes were proposed in this pull request?
This patch enables use of OrcRelation for SQL queries which read data from Hive tables. Changes in this patch:
- Added a new rule `OrcConversions` which would alter the plan to use `OrcRelation`. In this diff, the conversion is done only for reads.
- Added a new config `spark.sql.hive.convertMetastoreOrc` to control the conversion
BEFORE
```
scala> hqlContext.sql("SELECT * FROM orc_table").explain(true)
== Parsed Logical Plan ==
'Project [unresolvedalias(*, None)]
+- 'UnresolvedRelation `orc_table`, None
== Analyzed Logical Plan ==
key: string, value: string
Project [key#171,value#172]
+- MetastoreRelation default, orc_table, None
== Optimized Logical Plan ==
MetastoreRelation default, orc_table, None
== Physical Plan ==
HiveTableScan [key#171,value#172], MetastoreRelation default, orc_table, None
```
AFTER
```
scala> hqlContext.sql("SELECT * FROM orc_table").explain(true)
== Parsed Logical Plan ==
'Project [unresolvedalias(*, None)]
+- 'UnresolvedRelation `orc_table`, None
== Analyzed Logical Plan ==
key: string, value: string
Project [key#76,value#77]
+- SubqueryAlias orc_table
+- Relation[key#76,value#77] ORC part: struct<>, data: struct<key:string,value:string>
== Optimized Logical Plan ==
Relation[key#76,value#77] ORC part: struct<>, data: struct<key:string,value:string>
== Physical Plan ==
WholeStageCodegen
: +- Scan ORC part: struct<>, data: struct<key:string,value:string>[key#76,value#77] InputPaths: file:/user/hive/warehouse/orc_table
```
## How was this patch tested?
- Added a new unit test. Ran existing unit tests
- Ran with production like data
## Performance gains
Ran on a production table in Facebook (note that the data was in DWRF file format which is similar to ORC)
Best case : when there was no matching rows for the predicate in the query (everything is filtered out)
```
CPU time Wall time Total wall time across all tasks
================================================================
Without the change 541_515 sec 25.0 mins 165.8 hours
With change 407 sec 1.5 mins 15 mins
```
Average case: A subset of rows in the data match the query predicate
```
CPU time Wall time Total wall time across all tasks
================================================================
Without the change 624_630 sec 31.0 mins 199.0 h
With change 14_769 sec 5.3 mins 7.7 h
```
Author: Tejas Patil <tejasp@fb.com>
Closes#11891 from tejasapatil/orc_ppd.
`Expand` operator now uses its child plan's constraints as its valid constraints (i.e., the base of constraints). This is not correct because `Expand` will set its group by attributes to null values. So the nullability of these attributes should be true.
E.g., for an `Expand` operator like:
val input = LocalRelation('a.int, 'b.int, 'c.int).where('c.attr > 10 && 'a.attr < 5 && 'b.attr > 2)
Expand(
Seq(
Seq('c, Literal.create(null, StringType), 1),
Seq('c, 'a, 2)),
Seq('c, 'a, 'gid.int),
Project(Seq('a, 'c), input))
The `Project` operator has the constraints `IsNotNull('a)`, `IsNotNull('b)` and `IsNotNull('c)`. But the `Expand` should not have `IsNotNull('a)` in its constraints.
This PR is the first step for this issue and remove invalid constraints of `Expand` operator.
A test is added to `ConstraintPropagationSuite`.
Author: Liang-Chi Hsieh <simonh@tw.ibm.com>
Author: Michael Armbrust <michael@databricks.com>
Closes#11995 from viirya/fix-expand-constraints.
## What changes were proposed in this pull request?
JIRA: https://issues.apache.org/jira/browse/SPARK-13995
We infer relative `IsNotNull` constraints from logical plan's expressions in `constructIsNotNullConstraints` now. However, we don't consider the case of (nested) `Cast`.
For example:
val tr = LocalRelation('a.int, 'b.long)
val plan = tr.where('a.attr === 'b.attr).analyze
Then, the plan's constraints will have `IsNotNull(Cast(resolveColumn(tr, "a"), LongType))`, instead of `IsNotNull(resolveColumn(tr, "a"))`. This PR fixes it.
Besides, as `IsNotNull` constraints are most useful for `Attribute`, we should do recursing through any `Expression` that is null intolerant and construct `IsNotNull` constraints for all `Attribute`s under these Expressions.
For example, consider the following constraints:
val df = Seq((1,2,3)).toDF("a", "b", "c")
df.where("a + b = c").queryExecution.analyzed.constraints
The inferred isnotnull constraints should be isnotnull(a), isnotnull(b), isnotnull(c), instead of isnotnull(a + c) and isnotnull(c).
## How was this patch tested?
Test is added into `ConstraintPropagationSuite`.
Author: Liang-Chi Hsieh <simonh@tw.ibm.com>
Closes#11809 from viirya/constraint-cast.
## What changes were proposed in this pull request?
PySpark ml.clustering BisectingKMeans support export/import
## How was this patch tested?
doc test.
cc jkbradley
Author: Yanbo Liang <ybliang8@gmail.com>
Closes#12112 from yanboliang/spark-14305.
## What changes were proposed in this pull request?
Currently in Spark on YARN, configurations can be passed through SparkConf, env and command arguments, some parts are duplicated, like client argument and SparkConf. So here propose to simplify the command arguments.
## How was this patch tested?
This patch is tested manually with unit test.
CC vanzin tgravescs , please help to suggest this proposal. The original purpose of this JIRA is to remove `ClientArguments`, through refactoring some arguments like `--class`, `--arg` are not so easy to replace, so here I remove the most part of command line arguments, only keep the minimal set.
Author: jerryshao <sshao@hortonworks.com>
Closes#11603 from jerryshao/SPARK-12343.
## What changes were proposed in this pull request?
This PR updates the usage comments of `debug` according to the following commits.
- [SPARK-9754](https://issues.apache.org/jira/browse/SPARK-9754) removed `typeCheck`.
- [SPARK-14227](https://issues.apache.org/jira/browse/SPARK-14227) added `debugCodegen`.
## How was this patch tested?
Manual.
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#12094 from dongjoon-hyun/minor_fix_debug_usage.
## What changes were proposed in this pull request?
This PR throws Unsupported Operation exception for create index, drop index, alter index , lock table , lock database, unlock table, and unlock database operations that are not supported in Spark SQL. Currently these operations are executed executed by Hive.
Error:
spark-sql> drop index my_index on my_table;
Error in query:
Unsupported operation: drop index(line 1, pos 0)
## How was this patch tested?
Added test cases to HiveQuerySuite
yhuai hvanhovell andrewor14
Author: sureshthalamati <suresh.thalamati@gmail.com>
Closes#12069 from sureshthalamati/unsupported_ddl_spark-14133.
## What changes were proposed in this pull request?
This PR addresses the following
1. Supports native execution of SHOW DATABASES command
2. Fixes SHOW TABLES to apply the identifier_with_wildcards pattern if supplied.
SHOW TABLE syntax
```
SHOW TABLES [IN database_name] ['identifier_with_wildcards'];
```
SHOW DATABASES syntax
```
SHOW (DATABASES|SCHEMAS) [LIKE 'identifier_with_wildcards'];
```
## How was this patch tested?
Tests added in SQLQuerySuite (both hive and sql contexts) and DDLCommandSuite
Note: Since the table name pattern was not working , tests are added in both SQLQuerySuite to
verify the application of the table pattern.
Author: Dilip Biswal <dbiswal@us.ibm.com>
Closes#11991 from dilipbiswal/dkb_show_database.
## What changes were proposed in this pull request?
Fixes a compilation failure introduced in PR #12088 under Scala 2.10.
## How was this patch tested?
Compilation.
Author: Cheng Lian <lian@databricks.com>
Closes#12107 from liancheng/spark-14295-hotfix.
## What changes were proposed in this pull request?
Define and use ```KMeansWrapper``` for ```SparkR::kmeans```. It's only the code refactor for the original ```KMeans``` wrapper.
## How was this patch tested?
Existing tests.
cc mengxr
Author: Yanbo Liang <ybliang8@gmail.com>
Closes#12039 from yanboliang/spark-14059.
1.Implement LossFunction trait and implement squared error and cross entropy
loss with it
2.Implement unit test for gradient and loss
3.Implement InPlace trait and in-place layer evaluation
4.Refactor interface for ActivationFunction
5.Update of Layer and LayerModel interfaces
6.Fix random weights assignment
7.Implement memory allocation by MLP model instead of individual layers
These features decreased the memory usage and increased flexibility of
internal API.
Author: Alexander Ulanov <nashb@yandex.ru>
Author: avulanov <avulanov@gmail.com>
Closes#9229 from avulanov/mlp-refactoring.
## What changes were proposed in this pull request?
This PR implements `FileFormat.buildReader()` for the LibSVM data source. Besides that, a new interface method `prepareRead()` is added to `FileFormat`:
```scala
def prepareRead(
sqlContext: SQLContext,
options: Map[String, String],
files: Seq[FileStatus]): Map[String, String] = options
```
After migrating from `buildInternalScan()` to `buildReader()`, we lost the opportunity to collect necessary global information, since `buildReader()` works in a per-partition manner. For example, LibSVM needs to infer the total number of features if the `numFeatures` data source option is not set. Any necessary collected global information should be returned using the data source options map. By default, this method just returns the original options untouched.
An alternative approach is to absorb `inferSchema()` into `prepareRead()`, since schema inference is also some kind of global information gathering. However, this approach wasn't chosen because schema inference is optional, while `prepareRead()` must be called whenever a `HadoopFsRelation` based data source relation is instantiated.
One unaddressed problem is that, when `numFeatures` is absent, now the input data will be scanned twice. The `buildInternalScan()` code path doesn't need to do this because it caches the raw parsed RDD in memory before computing the total number of features. However, with `FileScanRDD`, the raw parsed RDD is created in a different way (e.g. partitioning) from the final RDD.
## How was this patch tested?
Tested using existing test suites.
Author: Cheng Lian <lian@databricks.com>
Closes#12088 from liancheng/spark-14295-libsvm-build-reader.
## What changes were proposed in this pull request?
In this patch, we set the initial `maxNumComponents` to `Integer.MAX_VALUE` instead of the default size ( which is 16) when allocating `compositeBuffer` in `TransportFrameDecoder` because `compositeBuffer` will introduce too many memory copies underlying if `compositeBuffer` is with default `maxNumComponents` when the frame size is large (which result in many transport messages). For details, please refer to [SPARK-14242](https://issues.apache.org/jira/browse/SPARK-14242).
## How was this patch tested?
spark unit tests and manual tests.
For manual tests, we can reproduce the performance issue with following code:
`sc.parallelize(Array(1,2,3),3).mapPartitions(a=>Array(new Array[Double](1024 * 1024 * 50)).iterator).reduce((a,b)=> a).length`
It's easy to see the performance gain, both from the running time and CPU usage.
Author: Zhang, Liye <liye.zhang@intel.com>
Closes#12038 from liyezhang556520/spark-14242.
## What changes were proposed in this pull request?
This PR support multiple Python UDFs within single batch, also improve the performance.
```python
>>> from pyspark.sql.types import IntegerType
>>> sqlContext.registerFunction("double", lambda x: x * 2, IntegerType())
>>> sqlContext.registerFunction("add", lambda x, y: x + y, IntegerType())
>>> sqlContext.sql("SELECT double(add(1, 2)), add(double(2), 1)").explain(True)
== Parsed Logical Plan ==
'Project [unresolvedalias('double('add(1, 2)), None),unresolvedalias('add('double(2), 1), None)]
+- OneRowRelation$
== Analyzed Logical Plan ==
double(add(1, 2)): int, add(double(2), 1): int
Project [double(add(1, 2))#14,add(double(2), 1)#15]
+- Project [double(add(1, 2))#14,add(double(2), 1)#15]
+- Project [pythonUDF0#16 AS double(add(1, 2))#14,pythonUDF0#18 AS add(double(2), 1)#15]
+- EvaluatePython [add(pythonUDF1#17, 1)], [pythonUDF0#18]
+- EvaluatePython [double(add(1, 2)),double(2)], [pythonUDF0#16,pythonUDF1#17]
+- OneRowRelation$
== Optimized Logical Plan ==
Project [pythonUDF0#16 AS double(add(1, 2))#14,pythonUDF0#18 AS add(double(2), 1)#15]
+- EvaluatePython [add(pythonUDF1#17, 1)], [pythonUDF0#18]
+- EvaluatePython [double(add(1, 2)),double(2)], [pythonUDF0#16,pythonUDF1#17]
+- OneRowRelation$
== Physical Plan ==
WholeStageCodegen
: +- Project [pythonUDF0#16 AS double(add(1, 2))#14,pythonUDF0#18 AS add(double(2), 1)#15]
: +- INPUT
+- !BatchPythonEvaluation [add(pythonUDF1#17, 1)], [pythonUDF0#16,pythonUDF1#17,pythonUDF0#18]
+- !BatchPythonEvaluation [double(add(1, 2)),double(2)], [pythonUDF0#16,pythonUDF1#17]
+- Scan OneRowRelation[]
```
## How was this patch tested?
Added new tests.
Using the following script to benchmark 1, 2 and 3 udfs,
```
df = sqlContext.range(1, 1 << 23, 1, 4)
double = F.udf(lambda x: x * 2, LongType())
print df.select(double(df.id)).count()
print df.select(double(df.id), double(df.id + 1)).count()
print df.select(double(df.id), double(df.id + 1), double(df.id + 2)).count()
```
Here is the results:
N | Before | After | speed up
---- |------------ | -------------|------
1 | 22 s | 7 s | 3.1X
2 | 38 s | 13 s | 2.9X
3 | 58 s | 16 s | 3.6X
This benchmark ran locally with 4 CPUs. For 3 UDFs, it launched 12 Python before before this patch, 4 process after this patch. After this patch, it will use less memory for multiple UDFs than before (less buffering).
Author: Davies Liu <davies@databricks.com>
Closes#12057 from davies/multi_udfs.
## What changes were proposed in this pull request?
Upgrade snappy to 1.1.2.4 to improve snappy read/write performance.
## How was this patch tested?
Tested by running a job on the cluster and saw 7.5% cpu savings after this change.
Author: Sital Kedia <skedia@fb.com>
Closes#12096 from sitalkedia/snappyRelease.
This patch fixes a compilation / build break in Spark's `java8-tests` and refactors their POM to simplify the build. See individual commit messages for more details.
Author: Josh Rosen <joshrosen@databricks.com>
Closes#12073 from JoshRosen/fix-java8-tests.
## What changes were proposed in this pull request?
Feature importances are exposed in the python API for GBTs.
Other changes:
* Update the random forest feature importance documentation to not repeat decision tree docstring and instead place a reference to it.
## How was this patch tested?
Python doc tests were updated to validate GBT feature importance.
Author: sethah <seth.hendrickson16@gmail.com>
Closes#12056 from sethah/Pyspark_GBT_feature_importance.
## What changes were proposed in this pull request?
If I press `CTRL-C` when running these tests, the temp files will be left in `sql/core` folder and I need to delete them manually. It's annoying. This PR just moves the temp files to the `java.io.tmpdir` folder and add a name prefix for them.
## How was this patch tested?
Existing Jenkins tests
Author: Shixiong Zhu <shixiong@databricks.com>
Closes#12093 from zsxwing/temp-file.
## What changes were proposed in this pull request?
Exclude jline from curator-recipes since it conflicts with scala 2.11 when running spark-shell. Should not affect scala 2.10 since it is builtin.
## How was this patch tested?
Ran spark-shell manually.
Author: Michel Lemay <mlemay@gmail.com>
Closes#12043 from michellemay/spark-13710-fix-jline-on-windows.
## What changes were proposed in this pull request?
Track executor information like host and port, cache size, running tasks.
TODO: tests
## How was this patch tested?
N/A
Author: Wenchen Fan <wenchen@databricks.com>
Closes#11888 from cloud-fan/status-tracker.
It looks like the docs were recently updated to reflect the History Server's support for incomplete applications, but they still had wording that suggested only completed applications were viewable. This fixes that.
My editor also introduced several whitespace removal changes, that I hope are OK, as text files shouldn't have trailing whitespace. To verify they're purely whitespace changes, add `&w=1` to your browser address. If this isn't acceptable, let me know and I'll update the PR.
I also didn't think this required a JIRA. Let me know if I should create one.
Not tested
Author: Michael Gummelt <mgummelt@mesosphere.io>
Closes#12045 from mgummelt/update-history-docs.
## What changes were proposed in this pull request?
This PR try to use `incUpdatedBlockStatuses ` to update the `updatedBlockStatuses ` when removing blocks, making sure `BlockManager` correctly updates `updatedBlockStatuses`
## How was this patch tested?
test("updated block statuses") in BlockManagerSuite.scala
Author: jeanlyn <jeanlyn92@gmail.com>
Closes#12091 from jeanlyn/updateBlock.
This PR is to provide native parsing support for DDL commands: `Alter View`. Since its AST trees are highly similar to `Alter Table`. Thus, both implementation are integrated into the same one.
Based on the Hive DDL document:
https://cwiki.apache.org/confluence/display/Hive/LanguageManual+DDL and https://cwiki.apache.org/confluence/display/Hive/PartitionedViews
**Syntax:**
```SQL
ALTER VIEW view_name RENAME TO new_view_name
```
- to change the name of a view to a different name
**Syntax:**
```SQL
ALTER VIEW view_name SET TBLPROPERTIES ('comment' = new_comment);
```
- to add metadata to a view
**Syntax:**
```SQL
ALTER VIEW view_name UNSET TBLPROPERTIES [IF EXISTS] ('comment', 'key')
```
- to remove metadata from a view
**Syntax:**
```SQL
ALTER VIEW view_name ADD [IF NOT EXISTS] PARTITION spec1[, PARTITION spec2, ...]
```
- to add the partitioning metadata for a view.
- the syntax of partition spec in `ALTER VIEW` is identical to `ALTER TABLE`, **EXCEPT** that it is **ILLEGAL** to specify a `LOCATION` clause.
**Syntax:**
```SQL
ALTER VIEW view_name DROP [IF EXISTS] PARTITION spec1[, PARTITION spec2, ...]
```
- to drop the related partition metadata for a view.
Added the related test cases to `DDLCommandSuite`
Author: gatorsmile <gatorsmile@gmail.com>
Author: xiaoli <lixiao1983@gmail.com>
Author: Xiao Li <xiaoli@Xiaos-MacBook-Pro.local>
Closes#11987 from gatorsmile/parseAlterView.
## What changes were proposed in this pull request?
Redirect error message to logWarning
## How was this patch tested?
Unit tests, manual tests
JoshRosen
Author: Nishkam Ravi <nishkamravi@gmail.com>
Closes#12052 from nishkamravi2/master_warning.
## What changes were proposed in this pull request?
Fixes a minor bug in the record reader constructor that was possibly introduced during refactoring.
## How was this patch tested?
N/A
Author: Sameer Agarwal <sameer@databricks.com>
Closes#12070 from sameeragarwal/vectorized-rr.
## What changes were proposed in this pull request?
This PR proposes a new data-structure based on a vectorized hashmap that can be potentially _codegened_ in `TungstenAggregate` to speed up aggregates with group by. Micro-benchmarks show a 10x improvement over the current `BytesToBytes` aggregation map.
## How was this patch tested?
Intel(R) Core(TM) i7-4960HQ CPU 2.60GHz
BytesToBytesMap: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
-------------------------------------------------------------------------------------------
hash 108 / 119 96.9 10.3 1.0X
fast hash 63 / 70 166.2 6.0 1.7X
arrayEqual 70 / 73 150.8 6.6 1.6X
Java HashMap (Long) 141 / 200 74.3 13.5 0.8X
Java HashMap (two ints) 145 / 185 72.3 13.8 0.7X
Java HashMap (UnsafeRow) 499 / 524 21.0 47.6 0.2X
BytesToBytesMap (off Heap) 483 / 548 21.7 46.0 0.2X
BytesToBytesMap (on Heap) 485 / 562 21.6 46.2 0.2X
Vectorized Hashmap 54 / 60 193.7 5.2 2.0X
Author: Sameer Agarwal <sameer@databricks.com>
Closes#12055 from sameeragarwal/vectorized-hashmap.
# What changes were proposed in this pull request?
https://issues.apache.org/jira/browse/SPARK-11892
Add save/load for spark ml.OneVsRest and its model. Also add OneVsRest and OneVsRestModel in MetaAlgorithmReadWrite.
# How was this patch tested?
Test with Scala unit test.
Author: Xusen Yin <yinxusen@gmail.com>
Closes#9934 from yinxusen/SPARK-11892.
## What changes were proposed in this pull request?
jira: https://issues.apache.org/jira/browse/SPARK-13782
Model export/import for BisectingKMeans in spark.ml and mllib
## How was this patch tested?
unit tests
Author: Yuhao Yang <hhbyyh@gmail.com>
Closes#11933 from hhbyyh/bisectingsave.
## What changes were proposed in this pull request?
1. Currently log4j which uses distributed cache only adds to AM's classpath, not executor's, this is introduced in #9118, which breaks the original meaning of that PR, so here add log4j file to the classpath of both AM and executors.
2. Automatically upload metrics.properties to distributed cache, so that it could be used by remote driver and executors implicitly.
## How was this patch tested?
Unit test and integration test is done.
Author: jerryshao <sshao@hortonworks.com>
Closes#11885 from jerryshao/SPARK-14062.
## What changes were proposed in this pull request?
This issue improves an input layer validation and adds related testcases to MultilayerPerceptronClassifier.
```scala
- // TODO: how to check ALSO that all elements are greater than 0?
- ParamValidators.arrayLengthGt(1)
+ (t: Array[Int]) => t.forall(ParamValidators.gt(0)) && t.length > 1
```
## How was this patch tested?
Pass the Jenkins tests including the new testcases.
Author: Dongjoon Hyun <dongjoon@apache.org>
Closes#11964 from dongjoon-hyun/SPARK-14164.
### What changes were proposed in this pull request?
This PR removes the ANTLR3 based parser, and moves the new ANTLR4 based parser into the `org.apache.spark.sql.catalyst.parser package`.
### How was this patch tested?
Existing unit tests.
cc rxin andrewor14 yhuai
Author: Herman van Hovell <hvanhovell@questtec.nl>
Closes#12071 from hvanhovell/SPARK-14211.
## What changes were proposed in this pull request?
Major changes:
1. Implement `FileFormat.buildReader()` for the CSV data source.
1. Add an extra argument to `FileFormat.buildReader()`, `physicalSchema`, which is basically the result of `FileFormat.inferSchema` or user specified schema.
This argument is necessary because the CSV data source needs to know all the columns of the underlying files to read the file.
## How was this patch tested?
Existing tests should do the work.
Author: Cheng Lian <lian@databricks.com>
Closes#12002 from liancheng/spark-14206-csv-build-reader.
## What changes were proposed in this pull request?
This change resolves an issue where `DataFrameNaFunctions.fill` changes a `FloatType` column to a `DoubleType`. We also clarify the contract that replacement values will be cast to the column data type, which may change the replacement value when casting to a lower precision type.
## How was this patch tested?
This patch has associated unit tests.
Author: Travis Crawford <travis@medium.com>
Closes#11967 from traviscrawford/SPARK-14081-dataframena.
## What changes were proposed in this pull request?
This pr is to add a config to control the maximum number of files as even small files have a non-trivial fixed cost. The current packing can put a lot of small files together which cases straggler tasks.
## How was this patch tested?
I added tests to check if many files get split into partitions in FileSourceStrategySuite.
Author: Takeshi YAMAMURO <linguin.m.s@gmail.com>
Closes#12068 from maropu/SPARK-14259.
