## What changes were proposed in this pull request?
Flush batch timely for pandas UDF.
This could improve performance when multiple pandas UDF plans are pipelined.
When batch being flushed in time, downstream pandas UDFs will get pipelined as soon as possible, and pipeline will help hide the donwstream UDFs computation time. For example:
When the first UDF start computing on batch-3, the second pipelined UDF can start computing on batch-2, and the third pipelined UDF can start computing on batch-1.
If we do not flush each batch in time, the donwstream UDF's pipeline will lag behind too much, which may increase the total processing time.
I add flush at two places:
* JVM process feed data into python worker. In jvm side, when write one batch, flush it
* VM process read data from python worker output, In python worker side, when write one batch, flush it
If no flush, the default buffer size for them are both 65536. Especially in the ML case, in order to make realtime prediction, we will make batch size very small. The buffer size is too large for the case, which cause downstream pandas UDF pipeline lag behind too much.
### Note
* This is only applied to pandas scalar UDF.
* Do not flush for each batch. The minimum interval between two flush is 0.1 second. This avoid too frequent flushing when batch size is small. It works like:
```
last_flush_time = time.time()
for batch in iterator:
writer.write_batch(batch)
flush_time = time.time()
if self.flush_timely and (flush_time - last_flush_time > 0.1):
stream.flush()
last_flush_time = flush_time
```
## How was this patch tested?
### Benchmark to make sure the flush do not cause performance regression
#### Test code:
```
numRows = ...
batchSize = ...
spark.conf.set('spark.sql.execution.arrow.maxRecordsPerBatch', str(batchSize))
df = spark.range(1, numRows + 1, numPartitions=1).select(col('id').alias('a'))
pandas_udf("int", PandasUDFType.SCALAR)
def fp1(x):
return x + 10
beg_time = time.time()
result = df.select(sum(fp1('a'))).head()
print("result: " + str(result[0]))
print("consume time: " + str(time.time() - beg_time))
```
#### Test Result:
params | Consume time (Before) | Consume time (After)
------------ | ----------------------- | ----------------------
numRows=100000000, batchSize=10000 | 23.43s | 24.64s
numRows=100000000, batchSize=1000 | 36.73s | 34.50s
numRows=10000000, batchSize=100 | 35.67s | 32.64s
numRows=1000000, batchSize=10 | 33.60s | 32.11s
numRows=100000, batchSize=1 | 33.36s | 31.82s
### Benchmark pipelined pandas UDF
#### Test code:
```
spark.conf.set('spark.sql.execution.arrow.maxRecordsPerBatch', '1')
df = spark.range(1, 31, numPartitions=1).select(col('id').alias('a'))
pandas_udf("int", PandasUDFType.SCALAR)
def fp1(x):
print("run fp1")
time.sleep(1)
return x + 100
pandas_udf("int", PandasUDFType.SCALAR)
def fp2(x, y):
print("run fp2")
time.sleep(1)
return x + y
beg_time = time.time()
result = df.select(sum(fp2(fp1('a'), col('a')))).head()
print("result: " + str(result[0]))
print("consume time: " + str(time.time() - beg_time))
```
#### Test Result:
**Before**: consume time: 63.57s
**After**: consume time: 32.43s
**So the PR improve performance by make downstream UDF get pipelined early.**
Please review https://spark.apache.org/contributing.html before opening a pull request.
Closes#24734 from WeichenXu123/improve_pandas_udf_pipeline.
Lead-authored-by: WeichenXu <weichen.xu@databricks.com>
Co-authored-by: Xiangrui Meng <meng@databricks.com>
Signed-off-by: gatorsmile <gatorsmile@gmail.com>
## 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?
After upgrading cloudpickle to 0.6.1 at https://github.com/apache/spark/pull/20691, one regression was found. Cloudpickle had a critical https://github.com/cloudpipe/cloudpickle/pull/240 for that.
Basically, it currently looks existing globals would override globals shipped in a function's, meaning:
**Before:**
```python
>>> def hey():
... return "Hi"
...
>>> spark.range(1).rdd.map(lambda _: hey()).collect()
['Hi']
>>> def hey():
... return "Yeah"
...
>>> spark.range(1).rdd.map(lambda _: hey()).collect()
['Hi']
```
**After:**
```python
>>> def hey():
... return "Hi"
...
>>> spark.range(1).rdd.map(lambda _: hey()).collect()
['Hi']
>>>
>>> def hey():
... return "Yeah"
...
>>> spark.range(1).rdd.map(lambda _: hey()).collect()
['Yeah']
```
Therefore, this PR upgrades cloudpickle to 0.8.0.
Note that cloudpickle's release cycle is quite short.
Between 0.6.1 and 0.7.0, it contains minor bug fixes. I don't see notable changes to double check and/or avoid.
There is virtually only this fix between 0.7.0 and 0.8.1 - other fixes are about testing.
## How was this patch tested?
Manually tested, tests were added. Verified unit tests were added in cloudpickle.
Closes#23904 from HyukjinKwon/SPARK-27000.
Authored-by: Hyukjin Kwon <gurwls223@apache.org>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
In this PR we've done two things:
1) updated the Spark's copy of cloudpickle to 0.6.1 (current stable)
The main reason Spark stayed with cloudpickle 0.4.x was that the default pickle protocol was changed in later versions.
2) started using pickle.HIGHEST_PROTOCOL for both Python 2 and Python 3 for serializers and broadcast
[Pyrolite](https://github.com/irmen/Pyrolite) has such Pickle protocol version support: reading: 0,1,2,3,4; writing: 2.
## How was this patch tested?
Jenkins tests.
Authors: Sloane Simmons, Boris Shminke
This contribution is original work of Sloane Simmons and Boris Shminke and they licensed it to the project under the project's open source license.
Closes#20691 from inpefess/pickle_protocol_4.
Lead-authored-by: Boris Shminke <boris@shminke.me>
Co-authored-by: singularperturbation <sloanes.k@gmail.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
Make .unpersist(), .destroy() non-blocking by default and adjust callers to request blocking only where important.
This also adds an optional blocking argument to Pyspark's RDD.unpersist(), which never had one.
## How was this patch tested?
Existing tests.
Closes#23685 from srowen/SPARK-26771.
Authored-by: Sean Owen <sean.owen@databricks.com>
Signed-off-by: Sean Owen <sean.owen@databricks.com>
## What changes were proposed in this pull request?
https://github.com/apache/spark/pull/21977 added a feature to limit Python worker resource limit.
This PR is kind of a followup of it. It proposes to add a test that checks the actual resource limit set by 'spark.executor.pyspark.memory'.
## How was this patch tested?
Unit tests were added.
Closes#23663 from HyukjinKwon/test_rlimit.
Authored-by: Hyukjin Kwon <gurwls223@apache.org>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
It's the follow-up PR for #22962, contains the following works:
- Remove `__init__` in TaskContext and BarrierTaskContext.
- Add more comments to explain the fix.
- Rewrite UT in a new class.
## How was this patch tested?
New UT in test_taskcontext.py
Closes#23435 from xuanyuanking/SPARK-25921-follow.
Authored-by: Yuanjian Li <xyliyuanjian@gmail.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
## What changes were proposed in this pull request?
During the follow-up work(#23435) for PySpark worker reuse scenario, we found that the worker reuse takes no effect for `sc.parallelize(xrange(...))`. It happened because of the specialize rdd.parallelize logic for xrange(introduced in #3264) generated data by lazy iterable range, which don't need to use the passed-in iterator. But this will break the end of stream checking in python worker and finally cause worker reuse takes no effect. See more details in [SPARK-26549](https://issues.apache.org/jira/browse/SPARK-26549) description.
We fix this by force using the passed-in iterator.
## How was this patch tested?
New UT in test_worker.py.
Closes#23470 from xuanyuanking/SPARK-26549.
Authored-by: Yuanjian Li <xyliyuanjian@gmail.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
Spark always creates secure py4j connections between java and python,
but it also allows users to pass in their own connection. This ensures
that even passed in connections are secure.
Added test cases verifying the failure with a (mocked) insecure gateway.
This is closely related to SPARK-26019, but this entirely forbids the
insecure connection, rather than creating the "escape-hatch".
Closes#23441 from squito/SPARK-26349.
Authored-by: Imran Rashid <irashid@cloudera.com>
Signed-off-by: Bryan Cutler <cutlerb@gmail.com>
## What changes were proposed in this pull request?
Python with rpc and disk encryption enabled along with a python broadcast variable and just read the value back on the driver side the job failed with:
Traceback (most recent call last): File "broadcast.py", line 37, in <module> words_new.value File "/pyspark.zip/pyspark/broadcast.py", line 137, in value File "pyspark.zip/pyspark/broadcast.py", line 122, in load_from_path File "pyspark.zip/pyspark/broadcast.py", line 128, in load EOFError: Ran out of input
To reproduce use configs: --conf spark.network.crypto.enabled=true --conf spark.io.encryption.enabled=true
Code:
words_new = sc.broadcast(["scala", "java", "hadoop", "spark", "akka"])
words_new.value
print(words_new.value)
## How was this patch tested?
words_new = sc.broadcast([“scala”, “java”, “hadoop”, “spark”, “akka”])
textFile = sc.textFile(“README.md”)
wordCounts = textFile.flatMap(lambda line: line.split()).map(lambda word: (word + words_new.value[1], 1)).reduceByKey(lambda a, b: a+b)
count = wordCounts.count()
print(count)
words_new.value
print(words_new.value)
Closes#23166 from redsanket/SPARK-26201.
Authored-by: schintap <schintap@oath.com>
Signed-off-by: Thomas Graves <tgraves@apache.org>
## What changes were proposed in this pull request?
This PR continues to break down a big large file into smaller files. See https://github.com/apache/spark/pull/23021. It targets to follow https://github.com/numpy/numpy/tree/master/numpy.
Basically this PR proposes to break down `pyspark/tests.py` into ...:
```
pyspark
...
├── testing
...
│ └── utils.py
├── tests
│ ├── __init__.py
│ ├── test_appsubmit.py
│ ├── test_broadcast.py
│ ├── test_conf.py
│ ├── test_context.py
│ ├── test_daemon.py
│ ├── test_join.py
│ ├── test_profiler.py
│ ├── test_rdd.py
│ ├── test_readwrite.py
│ ├── test_serializers.py
│ ├── test_shuffle.py
│ ├── test_taskcontext.py
│ ├── test_util.py
│ └── test_worker.py
...
```
## How was this patch tested?
Existing tests should cover.
`cd python` and .`/run-tests-with-coverage`. Manually checked they are actually being ran.
Each test (not officially) can be ran via:
```bash
SPARK_TESTING=1 ./bin/pyspark pyspark.tests.test_context
```
Note that if you're using Mac and Python 3, you might have to `OBJC_DISABLE_INITIALIZE_FORK_SAFETY=YES`.
Closes#23033 from HyukjinKwon/SPARK-26036.
Authored-by: hyukjinkwon <gurwls223@apache.org>
Signed-off-by: hyukjinkwon <gurwls223@apache.org>
