## What changes were proposed in this pull request?
This PR proposes to rename `mapPartitionsInPandas` to `mapInPandas` with a separate evaluation type .
Had an offline discussion with rxin, mengxr and cloud-fan
The reason is basically:
1. `SCALAR_ITER` doesn't make sense with `mapPartitionsInPandas`.
2. It cannot share the same Pandas UDF, for instance, at `select` and `mapPartitionsInPandas` unlike `GROUPED_AGG` because iterator's return type is different.
3. `mapPartitionsInPandas` -> `mapInPandas` - see https://github.com/apache/spark/pull/25044#issuecomment-508298552 and https://github.com/apache/spark/pull/25044#issuecomment-508299764
Renaming `SCALAR_ITER` as `MAP_ITER` is abandoned due to 2. reason.
For `XXX_ITER`, it might have to have a different interface in the future if we happen to add other versions of them. But this is an orthogonal topic with `mapPartitionsInPandas`.
## How was this patch tested?
Existing tests should cover.
Closes#25044 from HyukjinKwon/SPARK-28198.
Authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
This PR proposes to add `mapPartitionsInPandas` API to DataFrame by using existing `SCALAR_ITER` as below:
1. Filtering via setting the column
```python
from pyspark.sql.functions import pandas_udf, PandasUDFType
df = spark.createDataFrame([(1, 21), (2, 30)], ("id", "age"))
pandas_udf(df.schema, PandasUDFType.SCALAR_ITER)
def filter_func(iterator):
for pdf in iterator:
yield pdf[pdf.id == 1]
df.mapPartitionsInPandas(filter_func).show()
```
```
+---+---+
| id|age|
+---+---+
| 1| 21|
+---+---+
```
2. `DataFrame.loc`
```python
from pyspark.sql.functions import pandas_udf, PandasUDFType
import pandas as pd
df = spark.createDataFrame([['aa'], ['bb'], ['cc'], ['aa'], ['aa'], ['aa']], ["value"])
pandas_udf(df.schema, PandasUDFType.SCALAR_ITER)
def filter_func(iterator):
for pdf in iterator:
yield pdf.loc[pdf.value.str.contains('^a'), :]
df.mapPartitionsInPandas(filter_func).show()
```
```
+-----+
|value|
+-----+
| aa|
| aa|
| aa|
| aa|
+-----+
```
3. `pandas.melt`
```python
from pyspark.sql.functions import pandas_udf, PandasUDFType
import pandas as pd
df = spark.createDataFrame(
pd.DataFrame({'A': {0: 'a', 1: 'b', 2: 'c'},
'B': {0: 1, 1: 3, 2: 5},
'C': {0: 2, 1: 4, 2: 6}}))
pandas_udf("A string, variable string, value long", PandasUDFType.SCALAR_ITER)
def filter_func(iterator):
for pdf in iterator:
import pandas as pd
yield pd.melt(pdf, id_vars=['A'], value_vars=['B', 'C'])
df.mapPartitionsInPandas(filter_func).show()
```
```
+---+--------+-----+
| A|variable|value|
+---+--------+-----+
| a| B| 1|
| a| C| 2|
| b| B| 3|
| b| C| 4|
| c| B| 5|
| c| C| 6|
+---+--------+-----+
```
The current limitation of `SCALAR_ITER` is that it doesn't allow different length of result, which is pretty critical in practice - for instance, we cannot simply filter by using Pandas APIs but we merely just map N to N. This PR allows map N to M like flatMap.
This API mimics the way of `mapPartitions` but keeps API shape of `SCALAR_ITER` by allowing different results.
### How does this PR implement?
This PR adds mimics both `dapply` with Arrow optimization and Grouped Map Pandas UDF. At Python execution side, it reuses existing `SCALAR_ITER` code path.
Therefore, externally, we don't introduce any new type of Pandas UDF but internally we use another evaluation type code `205` (`SQL_MAP_PANDAS_ITER_UDF`).
This approach is similar with Pandas' Windows function implementation with Grouped Aggregation Pandas UDF functions - internally we have `203` (`SQL_WINDOW_AGG_PANDAS_UDF`) but externally we just share the same `GROUPED_AGG`.
## How was this patch tested?
Manually tested and unittests were added.
Closes#24997 from HyukjinKwon/scalar-udf-iter.
Authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
Currently with `toLocalIterator()` and `toPandas()` with Arrow enabled, if the Spark job being run in the background serving thread errors, it will be caught and sent to Python through the PySpark serializer.
This is not the ideal solution because it is only catch a SparkException, it won't handle an error that occurs in the serializer, and each method has to have it's own special handling to propagate the error.
This PR instead returns the Python Server object along with the serving port and authentication info, so that it allows the Python caller to join with the serving thread. During the call to join, the serving thread Future is completed either successfully or with an exception. In the latter case, the exception will be propagated to Python through the Py4j call.
## How was this patch tested?
Existing tests
Closes#24834 from BryanCutler/pyspark-propagate-server-error-SPARK-27992.
Authored-by: Bryan Cutler <cutlerb@gmail.com>
Signed-off-by: Bryan Cutler <cutlerb@gmail.com>
## What changes were proposed in this pull request?
`spark.sql.execution.arrow.enabled` was added when we add PySpark arrow optimization.
Later, in the current master, SparkR arrow optimization was added and it's controlled by the same configuration `spark.sql.execution.arrow.enabled`.
There look two issues about this:
1. `spark.sql.execution.arrow.enabled` in PySpark was added from 2.3.0 whereas SparkR optimization was added 3.0.0. The stability is different so it's problematic when we change the default value for one of both optimization first.
2. Suppose users want to share some JVM by PySpark and SparkR. They are currently forced to use the optimization for all or none if the configuration is set globally.
This PR proposes two separate configuration groups for PySpark and SparkR about Arrow optimization:
- Deprecate `spark.sql.execution.arrow.enabled`
- Add `spark.sql.execution.arrow.pyspark.enabled` (fallback to `spark.sql.execution.arrow.enabled`)
- Add `spark.sql.execution.arrow.sparkr.enabled`
- Deprecate `spark.sql.execution.arrow.fallback.enabled`
- Add `spark.sql.execution.arrow.pyspark.fallback.enabled ` (fallback to `spark.sql.execution.arrow.fallback.enabled`)
Note that `spark.sql.execution.arrow.maxRecordsPerBatch` is used within JVM side for both.
Note that `spark.sql.execution.arrow.fallback.enabled` was added due to behaviour change. We don't need it in SparkR - SparkR side has the automatic fallback.
## How was this patch tested?
Manually tested and some unittests were added.
Closes#24700 from HyukjinKwon/separate-sparkr-arrow.
Authored-by: HyukjinKwon <gurwls223@apache.org>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
This fixes an error when a PySpark local iterator, for both RDD and DataFrames, goes out of scope and the connection is closed before fully consuming the iterator. The error occurs on the JVM in the serving thread, when Python closes the local socket while the JVM is writing to it. This usually happens when there is enough data to fill the socket read buffer, causing the write call to block.
Additionally, this fixes a problem when an error occurs in the Python worker and the collect job is cancelled with an exception. Previously, the Python driver was never notified of the error so the user could get a partial result (iteration until the error) and the application will continue. With this change, an error in the worker is sent to the Python iterator and is then raised.
The change here introduces a protocol for PySpark local iterators that work as follows:
1) The local socket connection is made when the iterator is created
2) When iterating, Python first sends a request for partition data as a non-zero integer
3) While the JVM local iterator over partitions has next, it triggers a job to collect the next partition
4) The JVM sends a nonzero response to indicate it has the next partition to send
5) The next partition is sent to Python and read by the PySpark deserializer
6) After sending the entire partition, an `END_OF_DATA_SECTION` is sent to Python which stops the deserializer and allows to make another request
7) When the JVM gets a request from Python but has already consumed it's local iterator, it will send a zero response to Python and both will close the socket cleanly
8) If an error occurs in the worker, a negative response is sent to Python followed by the error message. Python will then raise a RuntimeError with the message, stopping iteration.
9) When the PySpark local iterator is garbage-collected, it will read any remaining data from the current partition (this is data that has already been collected) and send a request of zero to tell the JVM to stop collection jobs and close the connection.
Steps 1, 3, 5, 6 are the same as before. Step 8 was completely missing before because errors in the worker were never communicated back to Python. The other steps add synchronization to allow for a clean closing of the socket, with a small trade-off in performance for each partition. This is mainly because the JVM does not start collecting partition data until it receives a request to do so, where before it would eagerly write all data until the socket receive buffer is full.
## How was this patch tested?
Added new unit tests for DataFrame and RDD `toLocalIterator` and tested not fully consuming the iterator. Manual tests with Python 2.7 and 3.6.
Closes#24070 from BryanCutler/pyspark-toLocalIterator-clean-stop-SPARK-23961.
Authored-by: Bryan Cutler <cutlerb@gmail.com>
Signed-off-by: Bryan Cutler <cutlerb@gmail.com>
## What changes were proposed in this pull request?
This increases the minimum support version of pyarrow to 0.12.1 and removes workarounds in pyspark to remain compatible with prior versions. This means that users will need to have at least pyarrow 0.12.1 installed and available in the cluster or an `ImportError` will be raised to indicate an upgrade is needed.
## How was this patch tested?
Existing tests using:
Python 2.7.15, pyarrow 0.12.1, pandas 0.24.2
Python 3.6.7, pyarrow 0.12.1, pandas 0.24.0
Closes#24298 from BryanCutler/arrow-bump-min-pyarrow-SPARK-27276.
Authored-by: Bryan Cutler <cutlerb@gmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
This is a followup of https://github.com/apache/spark/pull/23285. This PR adds the notes into PySpark and SparkR documentation as well.
While I am here, I revised the doc a bit to make it sound a bit more neutral
## How was this patch tested?
Manually built the doc and verified.
Closes#24272 from HyukjinKwon/SPARK-26224.
Authored-by: Hyukjin Kwon <gurwls223@apache.org>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
The hashSeed method allocates 64 bytes instead of 8. Other bytes are always zeros (thanks to default behavior of ByteBuffer). And they could be excluded from hash calculation because they don't differentiate inputs.
## How was this patch tested?
By running the existing tests - XORShiftRandomSuite
Closes#20793 from MaxGekk/hash-buff-size.
Lead-authored-by: Maxim Gekk <maxim.gekk@databricks.com>
Co-authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
This is a minor follow-up PR for SPARK-27096. The original PR reconciled the join types supported between dataset and sql interface. In case of R, we do the join type validation in the R side. In this PR we do the correct validation and adds tests in R to test all the join types along with the error condition. Along with this, i made the necessary doc correction.
## How was this patch tested?
Add R tests.
Closes#24087 from dilipbiswal/joinfix_followup.
Authored-by: Dilip Biswal <dbiswal@us.ibm.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
Added .transform() method to Python DataFrame API to be in sync with Scala API.
## How was this patch tested?
Addition has been tested manually.
Closes#23877 from Hellsen83/pyspark-dataframe-transform.
Authored-by: Hellsen83 <erik.christiansen83@gmail.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
Currently `DataFrame.toPandas()` with arrow enabled or `ArrowStreamPandasSerializer` for pandas UDF with pyarrow<0.12 creates `datetime64[ns]` type series as intermediate data and then convert to `datetime.date` series, but the intermediate `datetime64[ns]` might cause an overflow even if the date is valid.
```
>>> import datetime
>>>
>>> t = [datetime.date(2262, 4, 12), datetime.date(2263, 4, 12)]
>>>
>>> df = spark.createDataFrame(t, 'date')
>>> df.show()
+----------+
| value|
+----------+
|2262-04-12|
|2263-04-12|
+----------+
>>>
>>> spark.conf.set("spark.sql.execution.arrow.enabled", "true")
>>>
>>> df.toPandas()
value
0 1677-09-21
1 1678-09-21
```
We should avoid creating such intermediate data and create `datetime.date` series directly instead.
## How was this patch tested?
Modified some tests to include the date which overflow caused by the intermediate conversion.
Run tests with pyarrow 0.8, 0.10, 0.11, 0.12 in my local environment.
Closes#23795 from ueshin/issues/SPARK-26887/date_as_object.
Authored-by: Takuya UESHIN <ueshin@databricks.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
When executing `toPandas` with Arrow enabled, partitions that arrive in the JVM out-of-order must be buffered before they can be send to Python. This causes an excess of memory to be used in the driver JVM and increases the time it takes to complete because data must sit in the JVM waiting for preceding partitions to come in.
This change sends un-ordered partitions to Python as soon as they arrive in the JVM, followed by a list of partition indices so that Python can assemble the data in the correct order. This way, data is not buffered at the JVM and there is no waiting on particular partitions so performance will be increased.
Followup to #21546
## How was this patch tested?
Added new test with a large number of batches per partition, and test that forces a small delay in the first partition. These test that partitions are collected out-of-order and then are are put in the correct order in Python.
## Performance Tests - toPandas
Tests run on a 4 node standalone cluster with 32 cores total, 14.04.1-Ubuntu and OpenJDK 8
measured wall clock time to execute `toPandas()` and took the average best time of 5 runs/5 loops each.
Test code
```python
df = spark.range(1 << 25, numPartitions=32).toDF("id").withColumn("x1", rand()).withColumn("x2", rand()).withColumn("x3", rand()).withColumn("x4", rand())
for i in range(5):
start = time.time()
_ = df.toPandas()
elapsed = time.time() - start
```
Spark config
```
spark.driver.memory 5g
spark.executor.memory 5g
spark.driver.maxResultSize 2g
spark.sql.execution.arrow.enabled true
```
Current Master w/ Arrow stream | This PR
---------------------|------------
5.16207 | 4.342533
5.133671 | 4.399408
5.147513 | 4.468471
5.105243 | 4.36524
5.018685 | 4.373791
Avg Master | Avg This PR
------------------|--------------
5.1134364 | 4.3898886
Speedup of **1.164821449**
Closes#22275 from BryanCutler/arrow-toPandas-oo-batches-SPARK-25274.
Authored-by: Bryan Cutler <cutlerb@gmail.com>
Signed-off-by: Bryan Cutler <cutlerb@gmail.com>
Signed-off-by: DylanGuedes <djmgguedesgmail.com>
## What changes were proposed in this pull request?
Addition of float, int and list hints for `pyspark.sql` Hint.
## How was this patch tested?
I did manual tests following the same principles used in the Scala version, and also added unit tests.
Closes#20788 from DylanGuedes/jira-21030.
Authored-by: DylanGuedes <djmgguedes@gmail.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
Implement codegen for `LocalTableScanExec` and `ExistingRDDExec`. Refactor to share code between `LocalTableScanExec`, `ExistingRDDExec`, `InputAdapter` and `RowDataSourceScanExec`.
The difference in `doProduce` between these four was that `ExistingRDDExec` and `RowDataSourceScanExec` triggered adding an `UnsafeProjection`, while `InputAdapter` and `LocalTableScanExec` did not.
In the new trait `InputRDDCodegen` I added a flag `createUnsafeProjection` which the operators set accordingly.
Note: `LocalTableScanExec` explicitly creates its input as `UnsafeRows`, so it was obvious why it doesn't need an `UnsafeProjection`. But if an `InputAdapter` may take input that is `InternalRows` but not `UnsafeRows`, then I think it doesn't need an unsafe projection just because any other operator that is its parent would do that. That assumes that that any parent operator would always result in some `UnsafeProjection` being eventually added, and hence the output of the `WholeStageCodegen` unit would be `UnsafeRows`. If these assumptions hold, I think `createUnsafeProjection` could be set to `(parent == null)`.
Note: Do not codegen `LocalTableScanExec` when it's the only operator. `LocalTableScanExec` has optimized driver-only `executeCollect` and `executeTake` code paths that are used to return `Command` results without starting Spark Jobs. They can no longer be used if the `LocalTableScanExec` gets optimized.
## How was this patch tested?
Covered and used in existing tests.
Closes#23127 from juliuszsompolski/SPARK-26159.
Authored-by: Juliusz Sompolski <julek@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
This PR is to add back `unionAll`, which is widely used. The name is also consistent with our ANSI SQL. We also have the corresponding `intersectAll` and `exceptAll`, which were introduced in Spark 2.4.
## How was this patch tested?
Added a test case in DataFrameSuite
Closes#23131 from gatorsmile/addBackUnionAll.
Authored-by: gatorsmile <gatorsmile@gmail.com>
Signed-off-by: gatorsmile <gatorsmile@gmail.com>
## What changes were proposed in this pull request?
The DOI foundation recommends [this new resolver](https://www.doi.org/doi_handbook/3_Resolution.html#3.8). Accordingly, this PR re`sed`s all static DOI links ;-)
## How was this patch tested?
It wasn't, since it seems as safe as a "[typo fix](https://spark.apache.org/contributing.html)".
In case any of the files is included from other projects, and should be updated there, please let me know.
Closes#23129 from katrinleinweber/resolve-DOIs-securely.
Authored-by: Katrin Leinweber <9948149+katrinleinweber@users.noreply.github.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
Following [SPARK-26024](https://issues.apache.org/jira/browse/SPARK-26024), I noticed the number of elements in each partition after repartitioning using `df.repartitionByRange` can vary for the same setup:
```scala
// Shuffle numbers from 0 to 1000, and make a DataFrame
val df = Random.shuffle(0.to(1000)).toDF("val")
// Repartition it using 3 partitions
// Sum up number of elements in each partition, and collect it.
// And do it several times
for (i <- 0 to 9) {
var counts = df.repartitionByRange(3, col("val"))
.mapPartitions{part => Iterator(part.size)}
.collect()
println(counts.toList)
}
// -> the number of elements in each partition varies
```
This is expected as for performance reasons this method uses sampling to estimate the ranges (with default size of 100). Hence, the output may not be consistent, since sampling can return different values. But documentation was not mentioning it at all, leading to misunderstanding.
## What changes were proposed in this pull request?
Update the documentation (Spark & PySpark) to mention the impact of `spark.sql.execution.rangeExchange.sampleSizePerPartition` on the resulting partitioned DataFrame.
Closes#23025 from JulienPeloton/SPARK-26024.
Authored-by: Julien <peloton@lal.in2p3.fr>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
## What changes were proposed in this pull request?
- Remove some AccumulableInfo .apply() methods
- Remove non-label-specific multiclass precision/recall/fScore in favor of accuracy
- Remove toDegrees/toRadians in favor of degrees/radians (SparkR: only deprecated)
- Remove approxCountDistinct in favor of approx_count_distinct (SparkR: only deprecated)
- Remove unused Python StorageLevel constants
- Remove Dataset unionAll in favor of union
- Remove unused multiclass option in libsvm parsing
- Remove references to deprecated spark configs like spark.yarn.am.port
- Remove TaskContext.isRunningLocally
- Remove ShuffleMetrics.shuffle* methods
- Remove BaseReadWrite.context in favor of session
- Remove Column.!== in favor of =!=
- Remove Dataset.explode
- Remove Dataset.registerTempTable
- Remove SQLContext.getOrCreate, setActive, clearActive, constructors
Not touched yet
- everything else in MLLib
- HiveContext
- Anything deprecated more recently than 2.0.0, generally
## How was this patch tested?
Existing tests
Closes#22921 from srowen/SPARK-25908.
Lead-authored-by: Sean Owen <sean.owen@databricks.com>
Co-authored-by: hyukjinkwon <gurwls223@apache.org>
Co-authored-by: Sean Owen <srowen@gmail.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
This patch is to bump the master branch version to 3.0.0-SNAPSHOT.
## How was this patch tested?
N/A
Closes#22606 from gatorsmile/bump3.0.
Authored-by: gatorsmile <gatorsmile@gmail.com>
Signed-off-by: gatorsmile <gatorsmile@gmail.com>
## What changes were proposed in this pull request?
In the PR, I propose to add an overloaded method for `sampleBy` which accepts the first argument of the `Column` type. This will allow to sample by any complex columns as well as sampling by multiple columns. For example:
```Scala
spark.createDataFrame(Seq(("Bob", 17), ("Alice", 10), ("Nico", 8), ("Bob", 17),
("Alice", 10))).toDF("name", "age")
.stat
.sampleBy(struct($"name", $"age"), Map(Row("Alice", 10) -> 0.3, Row("Nico", 8) -> 1.0), 36L)
.show()
+-----+---+
| name|age|
+-----+---+
| Nico| 8|
|Alice| 10|
+-----+---+
```
## How was this patch tested?
Added new test for sampling by multiple columns for Scala and test for Java, Python to check that `sampleBy` is able to sample by `Column` type argument.
Closes#22365 from MaxGekk/sample-by-column.
Authored-by: Maxim Gekk <maxim.gekk@databricks.com>
Signed-off-by: hyukjinkwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
This changes the calls of `toPandas()` and `createDataFrame()` to use the Arrow stream format, when Arrow is enabled. Previously, Arrow data was written to byte arrays where each chunk is an output of the Arrow file format. This was mainly due to constraints at the time, and caused some overhead by writing the schema/footer on each chunk of data and then having to read multiple Arrow file inputs and concat them together.
Using the Arrow stream format has improved these by increasing performance, lower memory overhead for the average case, and simplified the code. Here are the details of this change:
**toPandas()**
_Before:_
Spark internal rows are converted to Arrow file format, each group of records is a complete Arrow file which contains the schema and other metadata. Next a collect is done and an Array of Arrow files is the result. After that each Arrow file is sent to Python driver which then loads each file and concats them to a single Arrow DataFrame.
_After:_
Spark internal rows are converted to ArrowRecordBatches directly, which is the simplest Arrow component for IPC data transfers. The driver JVM then immediately starts serving data to Python as an Arrow stream, sending the schema first. It then starts a Spark job with a custom handler that sends Arrow RecordBatches to Python. Partitions arriving in order are sent immediately, and out-of-order partitions are buffered until the ones that precede it come in. This improves performance, simplifies memory usage on executors, and improves the average memory usage on the JVM driver. Since the order of partitions must be preserved, the worst case is that the first partition will be the last to arrive all data must be buffered in memory until then. This case is no worse that before when doing a full collect.
**createDataFrame()**
_Before:_
A Pandas DataFrame is split into parts and each part is made into an Arrow file. Then each file is prefixed by the buffer size and written to a temp file. The temp file is read and each Arrow file is parallelized as a byte array.
_After:_
A Pandas DataFrame is split into parts, then an Arrow stream is written to a temp file where each part is an ArrowRecordBatch. The temp file is read as a stream and the Arrow messages are examined. If the message is an ArrowRecordBatch, the data is saved as a byte array. After reading the file, each ArrowRecordBatch is parallelized as a byte array. This has slightly more processing than before because we must look each Arrow message to extract the record batches, but performance ends up a litle better. It is cleaner in the sense that IPC from Python to JVM is done over a single Arrow stream.
## How was this patch tested?
Added new unit tests for the additions to ArrowConverters in Scala, existing tests for Python.
## Performance Tests - toPandas
Tests run on a 4 node standalone cluster with 32 cores total, 14.04.1-Ubuntu and OpenJDK 8
measured wall clock time to execute `toPandas()` and took the average best time of 5 runs/5 loops each.
Test code
```python
df = spark.range(1 << 25, numPartitions=32).toDF("id").withColumn("x1", rand()).withColumn("x2", rand()).withColumn("x3", rand()).withColumn("x4", rand())
for i in range(5):
start = time.time()
_ = df.toPandas()
elapsed = time.time() - start
```
Current Master | This PR
---------------------|------------
5.803557 | 5.16207
5.409119 | 5.133671
5.493509 | 5.147513
5.433107 | 5.105243
5.488757 | 5.018685
Avg Master | Avg This PR
------------------|--------------
5.5256098 | 5.1134364
Speedup of **1.08060595**
## Performance Tests - createDataFrame
Tests run on a 4 node standalone cluster with 32 cores total, 14.04.1-Ubuntu and OpenJDK 8
measured wall clock time to execute `createDataFrame()` and get the first record. Took the average best time of 5 runs/5 loops each.
Test code
```python
def run():
pdf = pd.DataFrame(np.random.rand(10000000, 10))
spark.createDataFrame(pdf).first()
for i in range(6):
start = time.time()
run()
elapsed = time.time() - start
gc.collect()
print("Run %d: %f" % (i, elapsed))
```
Current Master | This PR
--------------------|----------
6.234608 | 5.665641
6.32144 | 5.3475
6.527859 | 5.370803
6.95089 | 5.479151
6.235046 | 5.529167
Avg Master | Avg This PR
---------------|----------------
6.4539686 | 5.4784524
Speedup of **1.178064192**
## Memory Improvements
**toPandas()**
The most significant improvement is reduction of the upper bound space complexity in the JVM driver. Before, the entire dataset was collected in the JVM first before sending it to Python. With this change, as soon as a partition is collected, the result handler immediately sends it to Python, so the upper bound is the size of the largest partition. Also, using the Arrow stream format is more efficient because the schema is written once per stream, followed by record batches. The schema is now only send from driver JVM to Python. Before, multiple Arrow file formats were used that each contained the schema. This duplicated schema was created in the executors, sent to the driver JVM, and then Python where all but the first one received are discarded.
I verified the upper bound limit by running a test that would collect data that would exceed the amount of driver JVM memory available. Using these settings on a standalone cluster:
```
spark.driver.memory 1g
spark.executor.memory 5g
spark.sql.execution.arrow.enabled true
spark.sql.execution.arrow.fallback.enabled false
spark.sql.execution.arrow.maxRecordsPerBatch 0
spark.driver.maxResultSize 2g
```
Test code:
```python
from pyspark.sql.functions import rand
df = spark.range(1 << 25, numPartitions=32).toDF("id").withColumn("x1", rand()).withColumn("x2", rand()).withColumn("x3", rand())
df.toPandas()
```
This makes total data size of 33554432×8×4 = 1073741824
With the current master, it fails with OOM but passes using this PR.
**createDataFrame()**
No significant change in memory except that using the stream format instead of separate file formats avoids duplicated the schema, similar to toPandas above. The process of reading the stream and parallelizing the batches does cause the record batch message metadata to be copied, but it's size is insignificant.
Closes#21546 from BryanCutler/arrow-toPandas-stream-SPARK-23030.
Authored-by: Bryan Cutler <cutlerb@gmail.com>
Signed-off-by: hyukjinkwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
Implements INTERSECT ALL clause through query rewrites using existing operators in Spark. Please refer to [Link](https://drive.google.com/open?id=1nyW0T0b_ajUduQoPgZLAsyHK8s3_dko3ulQuxaLpUXE) for the design.
Input Query
``` SQL
SELECT c1 FROM ut1 INTERSECT ALL SELECT c1 FROM ut2
```
Rewritten Query
```SQL
SELECT c1
FROM (
SELECT replicate_row(min_count, c1)
FROM (
SELECT c1,
IF (vcol1_cnt > vcol2_cnt, vcol2_cnt, vcol1_cnt) AS min_count
FROM (
SELECT c1, count(vcol1) as vcol1_cnt, count(vcol2) as vcol2_cnt
FROM (
SELECT c1, true as vcol1, null as vcol2 FROM ut1
UNION ALL
SELECT c1, null as vcol1, true as vcol2 FROM ut2
) AS union_all
GROUP BY c1
HAVING vcol1_cnt >= 1 AND vcol2_cnt >= 1
)
)
)
```
## How was this patch tested?
Added test cases in SQLQueryTestSuite, DataFrameSuite, SetOperationSuite
Author: Dilip Biswal <dbiswal@us.ibm.com>
Closes#21886 from dilipbiswal/dkb_intersect_all_final.
## What changes were proposed in this pull request?
Implements EXCEPT ALL clause through query rewrites using existing operators in Spark. In this PR, an internal UDTF (replicate_rows) is added to aid in preserving duplicate rows. Please refer to [Link](https://drive.google.com/open?id=1nyW0T0b_ajUduQoPgZLAsyHK8s3_dko3ulQuxaLpUXE) for the design.
**Note** This proposed UDTF is kept as a internal function that is purely used to aid with this particular rewrite to give us flexibility to change to a more generalized UDTF in future.
Input Query
``` SQL
SELECT c1 FROM ut1 EXCEPT ALL SELECT c1 FROM ut2
```
Rewritten Query
```SQL
SELECT c1
FROM (
SELECT replicate_rows(sum_val, c1)
FROM (
SELECT c1, sum_val
FROM (
SELECT c1, sum(vcol) AS sum_val
FROM (
SELECT 1L as vcol, c1 FROM ut1
UNION ALL
SELECT -1L as vcol, c1 FROM ut2
) AS union_all
GROUP BY union_all.c1
)
WHERE sum_val > 0
)
)
```
## How was this patch tested?
Added test cases in SQLQueryTestSuite, DataFrameSuite and SetOperationSuite
Author: Dilip Biswal <dbiswal@us.ibm.com>
Closes#21857 from dilipbiswal/dkb_except_all_final.
## What changes were proposed in this pull request?
Use SQLConf for PySpark to manage all sql configs, drop all the hard code in config usage.
## How was this patch tested?
Existing UT.
Author: Yuanjian Li <xyliyuanjian@gmail.com>
Closes#21648 from xuanyuanking/SPARK-24665.
## What changes were proposed in this pull request?
Address comments in #21370 and add more test.
## How was this patch tested?
Enhance test in pyspark/sql/test.py and DataFrameSuite
Author: Yuanjian Li <xyliyuanjian@gmail.com>
Closes#21553 from xuanyuanking/SPARK-24215-follow.
## What changes were proposed in this pull request?
The pandas_udf functionality was introduced in 2.3.0, but is not completely stable and still evolving. This adds a label to indicate it is still an experimental API.
## How was this patch tested?
NA
Author: Bryan Cutler <cutlerb@gmail.com>
Closes#21435 from BryanCutler/arrow-pandas_udf-experimental-SPARK-24392.
## What changes were proposed in this pull request?
When using Arrow for createDataFrame or toPandas and an error is encountered with fallback disabled, this will raise the same type of error instead of a RuntimeError. This change also allows for the traceback of the error to be retained and prevents the accidental chaining of exceptions with Python 3.
## How was this patch tested?
Updated existing tests to verify error type.
Author: Bryan Cutler <cutlerb@gmail.com>
Closes#20839 from BryanCutler/arrow-raise-same-error-SPARK-23699.
The exit() builtin is only for interactive use. applications should use sys.exit().
## What changes were proposed in this pull request?
All usage of the builtin `exit()` function is replaced by `sys.exit()`.
## How was this patch tested?
I ran `python/run-tests`.
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: Benjamin Peterson <benjamin@python.org>
Closes#20682 from benjaminp/sys-exit.
## What changes were proposed in this pull request?
This PR adds a configuration to control the fallback of Arrow optimization for `toPandas` and `createDataFrame` with Pandas DataFrame.
## How was this patch tested?
Manually tested and unit tests added.
You can test this by:
**`createDataFrame`**
```python
spark.conf.set("spark.sql.execution.arrow.enabled", False)
pdf = spark.createDataFrame([[{'a': 1}]]).toPandas()
spark.conf.set("spark.sql.execution.arrow.enabled", True)
spark.conf.set("spark.sql.execution.arrow.fallback.enabled", True)
spark.createDataFrame(pdf, "a: map<string, int>")
```
```python
spark.conf.set("spark.sql.execution.arrow.enabled", False)
pdf = spark.createDataFrame([[{'a': 1}]]).toPandas()
spark.conf.set("spark.sql.execution.arrow.enabled", True)
spark.conf.set("spark.sql.execution.arrow.fallback.enabled", False)
spark.createDataFrame(pdf, "a: map<string, int>")
```
**`toPandas`**
```python
spark.conf.set("spark.sql.execution.arrow.enabled", True)
spark.conf.set("spark.sql.execution.arrow.fallback.enabled", True)
spark.createDataFrame([[{'a': 1}]]).toPandas()
```
```python
spark.conf.set("spark.sql.execution.arrow.enabled", True)
spark.conf.set("spark.sql.execution.arrow.fallback.enabled", False)
spark.createDataFrame([[{'a': 1}]]).toPandas()
```
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#20678 from HyukjinKwon/SPARK-23380-conf.
## What changes were proposed in this pull request?
This PR explicitly specifies and checks the types we supported in `toPandas`. This was a hole. For example, we haven't finished the binary type support in Python side yet but now it allows as below:
```python
spark.conf.set("spark.sql.execution.arrow.enabled", "false")
df = spark.createDataFrame([[bytearray("a")]])
df.toPandas()
spark.conf.set("spark.sql.execution.arrow.enabled", "true")
df.toPandas()
```
```
_1
0 [97]
_1
0 a
```
This should be disallowed. I think the same things also apply to nested timestamps too.
I also added some nicer message about `spark.sql.execution.arrow.enabled` in the error message.
## How was this patch tested?
Manually tested and tests added in `python/pyspark/sql/tests.py`.
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#20625 from HyukjinKwon/pandas_convertion_supported_type.
## What changes were proposed in this pull request?
Expose range partitioning shuffle introduced by spark-22614
## How was this patch tested?
Unit test in dataframe.py
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: xubo245 <601450868@qq.com>
Closes#20456 from xubo245/SPARK22624_PysparkRangePartition.
## What changes were proposed in this pull request?
This PR proposes to disallow default value None when 'to_replace' is not a dictionary.
It seems weird we set the default value of `value` to `None` and we ended up allowing the case as below:
```python
>>> df.show()
```
```
+----+------+-----+
| age|height| name|
+----+------+-----+
| 10| 80|Alice|
...
```
```python
>>> df.na.replace('Alice').show()
```
```
+----+------+----+
| age|height|name|
+----+------+----+
| 10| 80|null|
...
```
**After**
This PR targets to disallow the case above:
```python
>>> df.na.replace('Alice').show()
```
```
...
TypeError: value is required when to_replace is not a dictionary.
```
while we still allow when `to_replace` is a dictionary:
```python
>>> df.na.replace({'Alice': None}).show()
```
```
+----+------+----+
| age|height|name|
+----+------+----+
| 10| 80|null|
...
```
## How was this patch tested?
Manually tested, tests were added in `python/pyspark/sql/tests.py` and doctests were fixed.
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#20499 from HyukjinKwon/SPARK-19454-followup.
## What changes were proposed in this pull request?
This PR proposes to explicitly specify Pandas and PyArrow versions in PySpark tests to skip or test.
We declared the extra dependencies:
b8bfce51ab/python/setup.py (L204)
In case of PyArrow:
Currently we only check if pyarrow is installed or not without checking the version. It already fails to run tests. For example, if PyArrow 0.7.0 is installed:
```
======================================================================
ERROR: test_vectorized_udf_wrong_return_type (pyspark.sql.tests.ScalarPandasUDF)
----------------------------------------------------------------------
Traceback (most recent call last):
File "/.../spark/python/pyspark/sql/tests.py", line 4019, in test_vectorized_udf_wrong_return_type
f = pandas_udf(lambda x: x * 1.0, MapType(LongType(), LongType()))
File "/.../spark/python/pyspark/sql/functions.py", line 2309, in pandas_udf
return _create_udf(f=f, returnType=return_type, evalType=eval_type)
File "/.../spark/python/pyspark/sql/udf.py", line 47, in _create_udf
require_minimum_pyarrow_version()
File "/.../spark/python/pyspark/sql/utils.py", line 132, in require_minimum_pyarrow_version
"however, your version was %s." % pyarrow.__version__)
ImportError: pyarrow >= 0.8.0 must be installed on calling Python process; however, your version was 0.7.0.
----------------------------------------------------------------------
Ran 33 tests in 8.098s
FAILED (errors=33)
```
In case of Pandas:
There are few tests for old Pandas which were tested only when Pandas version was lower, and I rewrote them to be tested when both Pandas version is lower and missing.
## How was this patch tested?
Manually tested by modifying the condition:
```
test_createDataFrame_column_name_encoding (pyspark.sql.tests.ArrowTests) ... skipped 'Pandas >= 1.19.2 must be installed; however, your version was 0.19.2.'
test_createDataFrame_does_not_modify_input (pyspark.sql.tests.ArrowTests) ... skipped 'Pandas >= 1.19.2 must be installed; however, your version was 0.19.2.'
test_createDataFrame_respect_session_timezone (pyspark.sql.tests.ArrowTests) ... skipped 'Pandas >= 1.19.2 must be installed; however, your version was 0.19.2.'
```
```
test_createDataFrame_column_name_encoding (pyspark.sql.tests.ArrowTests) ... skipped 'Pandas >= 0.19.2 must be installed; however, it was not found.'
test_createDataFrame_does_not_modify_input (pyspark.sql.tests.ArrowTests) ... skipped 'Pandas >= 0.19.2 must be installed; however, it was not found.'
test_createDataFrame_respect_session_timezone (pyspark.sql.tests.ArrowTests) ... skipped 'Pandas >= 0.19.2 must be installed; however, it was not found.'
```
```
test_createDataFrame_column_name_encoding (pyspark.sql.tests.ArrowTests) ... skipped 'PyArrow >= 1.8.0 must be installed; however, your version was 0.8.0.'
test_createDataFrame_does_not_modify_input (pyspark.sql.tests.ArrowTests) ... skipped 'PyArrow >= 1.8.0 must be installed; however, your version was 0.8.0.'
test_createDataFrame_respect_session_timezone (pyspark.sql.tests.ArrowTests) ... skipped 'PyArrow >= 1.8.0 must be installed; however, your version was 0.8.0.'
```
```
test_createDataFrame_column_name_encoding (pyspark.sql.tests.ArrowTests) ... skipped 'PyArrow >= 0.8.0 must be installed; however, it was not found.'
test_createDataFrame_does_not_modify_input (pyspark.sql.tests.ArrowTests) ... skipped 'PyArrow >= 0.8.0 must be installed; however, it was not found.'
test_createDataFrame_respect_session_timezone (pyspark.sql.tests.ArrowTests) ... skipped 'PyArrow >= 0.8.0 must be installed; however, it was not found.'
```
Author: hyukjinkwon <gurwls223@gmail.com>
Closes#20487 from HyukjinKwon/pyarrow-pandas-skip.
## What changes were proposed in this pull request?
It's not obvious from the comments that any added column must be a
function of the dataset that we are adding it to. Add a comment to
that effect to Scala, Python and R Data* methods.
Author: Henry Robinson <henry@cloudera.com>
Closes#20429 from henryr/SPARK-23157.
## What changes were proposed in this pull request?
Add colRegex API to PySpark
## How was this patch tested?
add a test in sql/tests.py
Author: Huaxin Gao <huaxing@us.ibm.com>
Closes#20390 from huaxingao/spark-23081.
## What changes were proposed in this pull request?
Make the default behavior of EXCEPT (i.e. EXCEPT DISTINCT) more
explicit in the documentation, and call out the change in behavior
from 1.x.
Author: Henry Robinson <henry@cloudera.com>
Closes#20254 from henryr/spark-23062.
## What changes were proposed in this pull request?
Currently we check pandas version by capturing if `ImportError` for the specific imports is raised or not but we can compare `LooseVersion` of the version strings as the same as we're checking pyarrow version.
## How was this patch tested?
Existing tests.
Author: Takuya UESHIN <ueshin@databricks.com>
Closes#20054 from ueshin/issues/SPARK-22874.
## What changes were proposed in this pull request?
Upgrade Spark to Arrow 0.8.0 for Java and Python. Also includes an upgrade of Netty to 4.1.17 to resolve dependency requirements.
The highlights that pertain to Spark for the update from Arrow versoin 0.4.1 to 0.8.0 include:
* Java refactoring for more simple API
* Java reduced heap usage and streamlined hot code paths
* Type support for DecimalType, ArrayType
* Improved type casting support in Python
* Simplified type checking in Python
## How was this patch tested?
Existing tests
Author: Bryan Cutler <cutlerb@gmail.com>
Author: Shixiong Zhu <zsxwing@gmail.com>
Closes#19884 from BryanCutler/arrow-upgrade-080-SPARK-22324.
## What changes were proposed in this pull request?
This change adds local checkpoint support to datasets and respective bind from Python Dataframe API.
If reliability requirements can be lowered to favor performance, as in cases of further quick transformations followed by a reliable save, localCheckpoints() fit very well.
Furthermore, at the moment Reliable checkpoints still incur double computation (see #9428)
In general it makes the API more complete as well.
## How was this patch tested?
Python land quick use case:
```python
>>> from time import sleep
>>> from pyspark.sql import types as T
>>> from pyspark.sql import functions as F
>>> def f(x):
sleep(1)
return x*2
...:
>>> df1 = spark.range(30, numPartitions=6)
>>> df2 = df1.select(F.udf(f, T.LongType())("id"))
>>> %time _ = df2.collect()
CPU times: user 7.79 ms, sys: 5.84 ms, total: 13.6 ms
Wall time: 12.2 s
>>> %time df3 = df2.localCheckpoint()
CPU times: user 2.38 ms, sys: 2.3 ms, total: 4.68 ms
Wall time: 10.3 s
>>> %time _ = df3.collect()
CPU times: user 5.09 ms, sys: 410 µs, total: 5.5 ms
Wall time: 148 ms
>>> sc.setCheckpointDir(".")
>>> %time df3 = df2.checkpoint()
CPU times: user 4.04 ms, sys: 1.63 ms, total: 5.67 ms
Wall time: 20.3 s
```
Author: Fernando Pereira <fernando.pereira@epfl.ch>
Closes#19805 from ferdonline/feature_dataset_localCheckpoint.
## What changes were proposed in this pull request?
When converting Pandas DataFrame/Series from/to Spark DataFrame using `toPandas()` or pandas udfs, timestamp values behave to respect Python system timezone instead of session timezone.
For example, let's say we use `"America/Los_Angeles"` as session timezone and have a timestamp value `"1970-01-01 00:00:01"` in the timezone. Btw, I'm in Japan so Python timezone would be `"Asia/Tokyo"`.
The timestamp value from current `toPandas()` will be the following:
```
>>> spark.conf.set("spark.sql.session.timeZone", "America/Los_Angeles")
>>> df = spark.createDataFrame([28801], "long").selectExpr("timestamp(value) as ts")
>>> df.show()
+-------------------+
| ts|
+-------------------+
|1970-01-01 00:00:01|
+-------------------+
>>> df.toPandas()
ts
0 1970-01-01 17:00:01
```
As you can see, the value becomes `"1970-01-01 17:00:01"` because it respects Python timezone.
As we discussed in #18664, we consider this behavior is a bug and the value should be `"1970-01-01 00:00:01"`.
## How was this patch tested?
Added tests and existing tests.
Author: Takuya UESHIN <ueshin@databricks.com>
Closes#19607 from ueshin/issues/SPARK-22395.
## What changes were proposed in this pull request?
Adding date and timestamp support with Arrow for `toPandas()` and `pandas_udf`s. Timestamps are stored in Arrow as UTC and manifested to the user as timezone-naive localized to the Python system timezone.
## How was this patch tested?
Added Scala tests for date and timestamp types under ArrowConverters, ArrowUtils, and ArrowWriter suites. Added Python tests for `toPandas()` and `pandas_udf`s with date and timestamp types.
Author: Bryan Cutler <cutlerb@gmail.com>
Author: Takuya UESHIN <ueshin@databricks.com>
Closes#18664 from BryanCutler/arrow-date-timestamp-SPARK-21375.
