Tutorial 6: going Spark

This section provides a tutorial on how to use Apache Spark in Eskapade. Spark works ‘out of the box’ in the Eskapade docker/vagrant image. For details on how to setup a custom Spark setup, see the Spark section in the Appendix.

In this tutorial we will basically redo Tutorial 1 but use Spark instead of Pandas for data processing. The following paragraphs describe step-by-step how to run a Spark job, use existing links and write your own links for Spark queries.

Note

To get familiar with Spark in Eskapade you can follow the exercises in python/eskapadespark/tutorials/tutorial_6.py.

Running the tutorial macro

The very first step to run the tutorial Spark job is:

$ eskapade_run python/eskapadespark/tutorials/tutorial_6.py

Eskapade will start a Spark session, do nothing, and quit - there are no chains/links defined yet. The Spark session is created via the SparkManager which, like the DataStore, is a singleton that configures and controls Spark sessions centrally. It is activated through the magic line:

process_manager.service(SparkManager).create_session(include_eskapade_modules=True)

Note that when the Spark session is created, the following line appears in logs:

Adding Python modules to egg archive <PATH_TO_ESKAPADE>/lib/es_python_modules.egg

This is the SparkManager that ensures all Eskapade source code is uploaded and available to the Spark cluster when running in a distributed environment. To include the Eskapade code the argument include_eskapade_modules need to be set to True (by default it is False).

If there was an ImportError: No module named pyspark then, most likely, SPARK_HOME and PYTHONPATH are not set up correctly. For details, see the Spark section in the Appendix.

Reading data

Spark can read data from various sources, e.g. local disk, HDFS, HIVE tables. Eskapade provides the SparkDfReader link that uses the pyspark.sql.DataFrameReader to read flat CSV files into Spark DataFrames, RDD’s, and Pandas DataFrames. To read in the Tutorial data, the following link should be added to the Data chain:

data = Chain('Data')
reader = SparkDfReader(name='Read_LA_ozone', store_key='data', read_methods=['csv'])
reader.read_meth_args['csv'] = (DATA_FILE_PATH,)
reader.read_meth_kwargs['csv'] = dict(sep=',', header=True, inferSchema=True)
data.add(reader)

The DataStore holds a pointer to the Spark dataframe in (distributed) memory. This is different from a Pandas dataframe, where the entire dataframe is stored in the DataStore, because a Spark dataframe residing on the cluster may not fit entirely in the memory of the machine running Eskapade. This means that Spark dataframes are never written to disk in DataStore pickles!

Using existing links

Spark has a large set of standard functions for Spark DataFrames and RDD’s. Although the purpose of Eskapade is not to duplicate this functionality, there are some links created for generic functions to facilitate specifying Spark queries directly in the macro, instead of hard-coding them in links. This is handy for bookkeeping queries at a central place and reducing code duplication, especially for smaller analysis steps. For example, the SparkExecuteQuery link takes any string containig SQL statements to perform a custom query with Spark on a dataframe.

Column transformations

To add two columns to the Tutorial data using the conversion functions defined earlier in the macro, two SparkWithColumn links need to be added to the Data chain, one for each additional column:

from pyspark.sql.functions import udf
from pyspark.sql.types import TimestampType, FloatType
...

transform = SparkWithColumn(name='Transform_doy', read_key=reader.store_key,
                            store_key='transformed_data', col_select=['doy'],
                            func=udf(comp_date, TimestampType()), new_column='date')
data.add(transform)
transform = SparkWithColumn(name='Transform_vis', read_key=transform.store_key,
                            store_key='transformed_data', col_select=['vis'],
                            func=udf(mi_to_km, FloatType()), new_column='vis_km')
data.add(transform)

Note that the functions defined in the macro are converted to user-defined functions with pyspark.sql.functions.udf and their output types are explicitly specified in terms of pyspark.sql.types. Omitting these type definitions can lead to obscure errors when executing the job.

Creating custom links

More complex queries deserve their own links since links provide full flexibility w.r.t. specifying custom data operation. For this Tutorial the ‘complex query’ is to just print 42 rows of the Spark dataframe. Of course, more advanced Spark functions can be applied in a similar fashion. A link is created just like was done before, e.g.:

$ eskapade_generate_link --dir python/eskapadespark/links SparkDfPrinter

This creates the link python/eskapadespark/links/sparkdfprinter.py. Do not forget to include the import statements in the __init__.py file as indicated by the eskapade_generate_link command.

The next step is to add the desired functionality to the link. In this case, the Spark dataframe needs to be retrieved from the DataStore and a show() method of that dataframe needs to be executed. The execute() method of the link is the right location for this:

def execute(self):
  """Execute the link.

  :returns: status code of execution
  :rtype: StatusCode
  """
  settings = process_manager.service(ConfigObject)
  ds = process_manager.service(DataStore)

  # --- your algorithm code goes here
  self.logger.debug('Now executing link: {link}.', link=self.name)
  df = ds[self.read_key]
  df.show(self.nrows)

  return StatusCode.Success

There is an additional attribute self.nrows which should be set in the link. By default, a generated link process only the read_key and store_key arguments and fails if there are any residual kwargs. To set the nrows attribute, add nrows to the key-value arguments in the __init__() method:

def __init__(self, **kwargs):
    ...

    self._process_kwargs(kwargs, read_key=None, store_key=None, nrows=1)

In order to configure Eskapade to run this link, the link needs to be added to a chain, e.g. Summary, in the tutorial/tutorial_6.py macro. This should look similar to:

printer = SparkDfPrinter(name='Print_spark_df', read_key=transform.store_key, nrows=42)
summary.add(printer)

The name of the dataframe is the output name of the transform link and the number of rows to print is specified by the nrows parameter.

Eskapade should now be ready to finally execute the macro and provide the desired output:

$ eskapade_run python/eskapadespark/tutorials/tutorial_6.py

* * * Welcome to Eskapade * * *
...

+-----+----+----+--------+----+----+---+---+---+---+--------------------+--------+
|ozone|  vh|wind|humidity|temp| ibh|dpg|ibt|vis|doy|                date|  vis_km|
+-----+----+----+--------+----+----+---+---+---+---+--------------------+--------+
|    3|5710|   4|      28|  40|2693|-25| 87|250|  3|1976-01-03 00:00:...| 402.335|
|    5|5700|   3|      37|  45| 590|-24|128|100|  4|1976-01-04 00:00:...| 160.934|
|    5|5760|   3|      51|  54|1450| 25|139| 60|  5|1976-01-05 00:00:...| 96.5604|
...

|    6|5700|   4|      86|  55|2398| 21|121|200| 44|1976-02-13 00:00:...| 321.868|
|    4|5650|   5|      61|  41|5000| 51| 24|100| 45|1976-02-14 00:00:...| 160.934|
|    3|5610|   5|      62|  41|4281| 42| 52|250| 46|1976-02-15 00:00:...| 402.335|
+-----+----+----+--------+----+----+---+---+---+---+--------------------+--------+
only showing top 42 rows
...

* * * Leaving Eskapade. Bye! * * *

That’s it!

Spark Streaming

Eskapade supports the use of Spark Streaming as demonstrated in the word count example tutorials/esk610_spark_streaming_wordcount.py. The data is processed in (near) real-time as micro batches of RDD’s, so-called discretized streaming, where the stream originates from either new incoming files or network connection. As with regular Spark queries, various transformations can be defined and applied in subsequent Eskapade links.

For details on Spark Streaming, see also https://spark.apache.org/docs/2.1.1/streaming-programming-guide.html.

File stream

The word count example using the file stream method can be run by executing in two different terminals:

terminal 1 $ eskapade_run -c stream_type='file' python/eskapadespark/tutorials/esk610_spark_streaming_wordcount.py

terminal 2 $ mkdir /tmp/eskapade_stream_test/
terminal 2 $ for ((i=0; i<=100; i++)); do echo "Hello world" > /tmp/eskapade_stream_test/dummy_$(printf %05d ${i}); sleep 0.2; done

Where bash for-loop will create a new file containing Hello world in the /tmp/eskapade_stream_test directory every 0.2 second. Spark Streaming will pick up and process these files and in terminal 1 a word count of the processed data will be displayed. Output is stored in results/esk610_spark_streaming/data/v0/dstream/wordcount. Only new files in /tmp/eskapade_stream_test are processed, do not forget to delete this directory.

TCP stream

The word count example using the TCP stream method can be run by executing in two different terminals:

terminal 1 $ eskapade_run -c stream_type='tcp' python/eskapadespark/tutorials/esk610_spark_streaming_wordcount.py

terminal 2 $ nc -lk 9999

Where nc (netcat) will stream data to port 9999 and Spark Streaming will listen to this port and process incoming data. In terminal 2 random words can be type (followed by enter) and in terminal 1 a word count of the processed data will by displayed. Output is stored in results/esk610_spark_streaming/data/v0/dstream/wordcount.