## 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>
82c18c240a