a803ac3e06
I'm relatively new to Spark and functional programming, so forgive me if this pull request is just a result of my misunderstanding of how Spark should be used.
Currently, if one happens to use a mutable object as `zeroValue` for `RDD.aggregate()`, possibly unexpected behavior can occur.
This is because pyspark's current implementation of `RDD.aggregate()` does not serialize or make a copy of `zeroValue` before handing it off to `RDD.mapPartitions(...).fold(...)`. This results in a single reference to `zeroValue` being used for both `RDD.mapPartitions()` and `RDD.fold()` on each partition. This can result in strange accumulator values being fed into each partition's call to `RDD.fold()`, as the `zeroValue` may have been changed in-place during the `RDD.mapPartitions()` call.
As an illustrative example, submit the following to `spark-submit`:
```
from pyspark import SparkConf, SparkContext
import collections
def updateCounter(acc, val):
print 'update acc:', acc
print 'update val:', val
acc[val] += 1
return acc
def comboCounter(acc1, acc2):
print 'combo acc1:', acc1
print 'combo acc2:', acc2
acc1.update(acc2)
return acc1
def main():
conf = SparkConf().setMaster("local").setAppName("Aggregate with Counter")
sc = SparkContext(conf = conf)
print '======= AGGREGATING with ONE PARTITION ======='
print sc.parallelize(range(1,10), 1).aggregate(collections.Counter(), updateCounter, comboCounter)
print '======= AGGREGATING with TWO PARTITIONS ======='
print sc.parallelize(range(1,10), 2).aggregate(collections.Counter(), updateCounter, comboCounter)
if __name__ == "__main__":
main()
```
One probably expects this to output the following:
```
Counter({1: 1, 2: 1, 3: 1, 4: 1, 5: 1, 6: 1, 7: 1, 8: 1, 9: 1})
```
But it instead outputs this (regardless of the number of partitions):
```
Counter({1: 2, 2: 2, 3: 2, 4: 2, 5: 2, 6: 2, 7: 2, 8: 2, 9: 2})
```
This is because (I believe) `zeroValue` gets passed correctly to each partition, but after `RDD.mapPartitions()` completes, the `zeroValue` object has been updated and is then passed to `RDD.fold()`, which results in all items being double-counted within each partition before being finally reduced at the calling node.
I realize that this type of calculation is typically done by `RDD.mapPartitions(...).reduceByKey(...)`, but hopefully this illustrates some potentially confusing behavior. I also noticed that other `RDD` methods use this `deepcopy` approach to creating unique copies of `zeroValue` (i.e., `RDD.aggregateByKey()` and `RDD.foldByKey()`), and that the Scala implementations do seem to serialize the `zeroValue` object appropriately to prevent this type of behavior.
Author: Nicholas Hwang <moogling@gmail.com>
Closes #7378 from njhwang/master and squashes the following commits:
659bb27 [Nicholas Hwang] Fixed RDD.aggregate() to perform a reduce operation on collected mapPartitions results, similar to how fold currently is implemented. This prevents an initial combOp being performed on each partition with zeroValue (which leads to unexpected behavior if zeroValue is a mutable object) before being combOp'ed with other partition results.
8d8d694 [Nicholas Hwang] Changed dict construction to be compatible with Python 2.6 (cannot use list comprehensions to make dicts)
56eb2ab [Nicholas Hwang] Fixed whitespace after colon to conform with PEP8
391de4a [Nicholas Hwang] Removed used of collections.Counter from RDD tests for Python 2.6 compatibility; used defaultdict(int) instead. Merged treeAggregate test with mutable zero value into aggregate test to reduce code duplication.
2fa4e4b [Nicholas Hwang] Merge branch 'master' of https://github.com/njhwang/spark
ba528bd [Nicholas Hwang] Updated comments regarding protection of zeroValue from mutation in RDD.aggregate(). Added regression tests for aggregate(), fold(), aggregateByKey(), foldByKey(), and treeAggregate(), all with both 1 and 2 partition RDDs. Confirmed that aggregate() is the only problematic implementation as of commit
|
||
|---|---|---|
| assembly | ||
| bagel | ||
| bin | ||
| build | ||
| conf | ||
| core | ||
| data/mllib | ||
| dev | ||
| docker | ||
| docs | ||
| ec2 | ||
| examples | ||
| external | ||
| extras | ||
| graphx | ||
| launcher | ||
| mllib | ||
| network | ||
| project | ||
| python | ||
| R | ||
| repl | ||
| sbin | ||
| sbt | ||
| sql | ||
| streaming | ||
| tools | ||
| unsafe | ||
| yarn | ||
| .gitattributes | ||
| .gitignore | ||
| .rat-excludes | ||
| CONTRIBUTING.md | ||
| LICENSE | ||
| make-distribution.sh | ||
| NOTICE | ||
| pom.xml | ||
| pylintrc | ||
| README.md | ||
| scalastyle-config.xml | ||
| tox.ini | ||
Apache Spark
Spark is a fast and general cluster computing system for Big Data. It provides high-level APIs in Scala, Java, and Python, and an optimized engine that supports general computation graphs for data analysis. It also supports a rich set of higher-level tools including Spark SQL for SQL and DataFrames, MLlib for machine learning, GraphX for graph processing, and Spark Streaming for stream processing.
Online Documentation
You can find the latest Spark documentation, including a programming guide, on the project web page and project wiki. This README file only contains basic setup instructions.
Building Spark
Spark is built using Apache Maven. To build Spark and its example programs, run:
build/mvn -DskipTests clean package
(You do not need to do this if you downloaded a pre-built package.) More detailed documentation is available from the project site, at "Building Spark".
Interactive Scala Shell
The easiest way to start using Spark is through the Scala shell:
./bin/spark-shell
Try the following command, which should return 1000:
scala> sc.parallelize(1 to 1000).count()
Interactive Python Shell
Alternatively, if you prefer Python, you can use the Python shell:
./bin/pyspark
And run the following command, which should also return 1000:
>>> sc.parallelize(range(1000)).count()
Example Programs
Spark also comes with several sample programs in the examples directory.
To run one of them, use ./bin/run-example <class> [params]. For example:
./bin/run-example SparkPi
will run the Pi example locally.
You can set the MASTER environment variable when running examples to submit
examples to a cluster. This can be a mesos:// or spark:// URL,
"yarn-cluster" or "yarn-client" to run on YARN, and "local" to run
locally with one thread, or "local[N]" to run locally with N threads. You
can also use an abbreviated class name if the class is in the examples
package. For instance:
MASTER=spark://host:7077 ./bin/run-example SparkPi
Many of the example programs print usage help if no params are given.
Running Tests
Testing first requires building Spark. Once Spark is built, tests can be run using:
./dev/run-tests
Please see the guidance on how to run tests for a module, or individual tests.
A Note About Hadoop Versions
Spark uses the Hadoop core library to talk to HDFS and other Hadoop-supported storage systems. Because the protocols have changed in different versions of Hadoop, you must build Spark against the same version that your cluster runs.
Please refer to the build documentation at "Specifying the Hadoop Version" for detailed guidance on building for a particular distribution of Hadoop, including building for particular Hive and Hive Thriftserver distributions. See also "Third Party Hadoop Distributions" for guidance on building a Spark application that works with a particular distribution.
Configuration
Please refer to the Configuration guide in the online documentation for an overview on how to configure Spark.