Pysparkling provides a faster, more responsive way to develop programs for PySpark. It enables code intended for Spark applications to execute entirely in Python, without incurring the overhead of initializing and passing data through the JVM and Hadoop. The focus is on having a lightweight and fast implementation for small datasets at the expense of some data resilience features and some parallel processing features.
How does it work? To switch execution of a script from PySpark to pysparkling, have the code initialize a pysparkling Context instead of a SparkContext, and use the pysparkling Context to set up your RDDs. The beauty is you don't have to change a single line of code after the Context initialization, because pysparkling's API is (almost) exactly the same as PySpark's. Since it's so easy to switch between PySpark and pysparkling, you can choose the right tool for your use case.
When would I use it? Say you are writing a Spark application because you need robust computation on huge datasets, but you also want the same application to provide fast answers on a small dataset. You're finding Spark is not responsive enough for your needs, but you don't want to rewrite an entire separate application for the small-answers-fast problem. You'd rather reuse your Spark code but somehow get it to run fast. Pysparkling bypasses the stuff that causes Spark's long startup times and less responsive feel.
Here are a few areas where pysparkling excels:
- Small to medium-scale exploratory data analysis
- Application prototyping
- Low-latency web deployments
- Unit tests
Example: you have a pipeline that processes 100k input documents
and converts them to normalized features. They are used to train a local
scikit-learn classifier. The preprocessing is perfect for a full Spark
task. Now, you want to use this trained classifier in an API
endpoint. Assume you need the same pre-processing pipeline for a single
document per API call. This does not have to be done in parallel, but there
should be only a small overhead in initialization and preferably no
dependency on the JVM. This is what pysparkling
is for.
Links: Documentation, Github, Issue Tracker
pip install pysparkling
- Supports URI schemes
s3://
,hdfs://
,gs://
,http://
andfile://
for Amazon S3, HDFS, Google Storage, web and local file access. Specify multiple files separated by comma. Resolves*
and?
wildcards. - Handles
.gz
,.zip
,.lzma
,.xz
,.bz2
,.tar
,.tar.gz
and.tar.bz2
compressed files. Supports reading of.7z
files. - Parallelization via
multiprocessing.Pool
,concurrent.futures.ThreadPoolExecutor
or any other Pool-like objects that have amap(func, iterable)
method. - Plain pysparkling does not have any dependencies (use
pip install pysparkling
). Some file access methods have optional dependencies:boto
for AWS S3,requests
for http,hdfs
for hdfs
Some demos are in the notebooks docs/demo.ipynb and docs/iris.ipynb .
Word Count
from pysparkling import Context
counts = Context().textFile(
'README.rst'
).map(
lambda line: ''.join(ch if ch.isalnum() else ' ' for ch in line)
).flatMap(
lambda line: line.split(' ')
).map(
lambda word: (word, 1)
).reduceByKey(
lambda a, b: a + b
)
print(counts.collect())
which prints a long list of pairs of words and their counts.
A usual pysparkling
session starts with either parallelizing a list
or
by reading data from a file using the methods Context.parallelize(my_list)
or Context.textFile("path/to/textfile.txt")
. These two methods return an
RDD
which can then be processed with the methods below.
API doc: http://pysparkling.trivial.io/en/latest/api.html#rdd
A Context
describes the setup. Instantiating a Context with the default
arguments using Context()
is the most lightweight setup. All data is just
in the local thread and is never serialized or deserialized.
If you want to process the data in parallel, you can use the multiprocessing
module. Given the limitations of the default pickle
serializer, you can
specify to serialize all methods with cloudpickle
instead. For example,
a common instantiation with multiprocessing
looks like this:
c = Context(
multiprocessing.Pool(4),
serializer=cloudpickle.dumps,
deserializer=pickle.loads,
)
This assumes that your data is serializable with pickle
which is generally
faster. You can also specify a custom serializer/deserializer for data.
API doc: http://pysparkling.trivial.io/en/latest/api.html#context
The functionality provided by this module is used in Context.textFile()
for reading and in RDD.saveAsTextFile()
for writing. You can use this
submodule for writing files directly with File(filename).dump(some_data)
,
File(filename).load()
and File.exists(path)
to read, write and check
for existance of a file. All methods transparently handle http://
, s3://
and file://
locations and compression/decompression of .gz
and
.bz2
files.
Use environment variables AWS_SECRET_ACCESS_KEY
and AWS_ACCESS_KEY_ID
for auth and use file paths of the form s3://bucket_name/filename.txt
.
API doc: http://pysparkling.trivial.io/en/latest/api.html#fileio