Framework for simpler Spark Pipelines
Project description
SparkPipelineFramework
SparkPipelineFramework implements a few design patterns to make it easier to create Spark applications that:
- Separate data transformation logic from the pipeline execution code so you can compose pipelines by just stringing together transformers. (Based on the SparkML Pipeline class but enhanced to work for both ML and non-ML transformations)
- Enables running SQL transformations without writing any code
- Enables versioning of transformations so different pipelines can use older or newer versions of each transformer. This enables you to upgrade each pipeline at your own choice
- Enables Autocomplete of transformations when creating pipelines (in PyCharm).
- Implement many separation-of-concerns e.g., logging, performance monitoring, error reporting
- Supports both non-ML, ML and mixed workloads
- Has an additional(optional) library SparkAutoMapper(https://github.com/icanbwell/SparkAutoMapper) that enables data engineers and data analysts to easily map data without writing code.
- Has an additional(optional) library SparkPipelineFramework.Testing (https://github.com/icanbwell/SparkPipelineFramework.Testing) that allows you to create unit tests without writing code; you just put in the input file and confirm the output file
PyPi Package
This code is available as a package to import into your project. https://pypi.org/project/sparkpipelineframework/
Using it in your project
(For an example project that uses SparkPipelineFramework, see https://github.com/icanbwell/TestSparkPipelineFramework)
- Add sparkpipelineframework package to your project requirements.txt/Pipefile
- make init. (this will setup Spark, Docker (to run Spark) )
- Create a folder called library in your project
using Pycharm
You can run SparkPipelineFrame project from Pycharm
- Add a new Docker Compose interpreter
- Choose
docker-compose.yml
for the configuration file - Choose
dev
for the Service - Click OK and give Pycharm a couple of minutes to index the content of the docker container
- Right click on the Test folder and click "Run 'pytest in tests'"
To create a new pipeline
- Create a class derived from
FrameworkPipeline
- In your init function set self.transformers to the list of transformers to run for this pipeline. For example:
class MyPipeline(FrameworkPipeline):
def __init__(self, parameters: Dict[str, Any], progress_logger: ProgressLogger):
super().__init__(parameters=parameters,
progress_logger=progress_logger)
self.transformers = self.create_steps([
FrameworkCsvLoader(
view="flights",
path_to_csv=parameters["flights_path"]
),
FeaturesCarriers(parameters=parameters).transformers,
])
To Add a SQL transformation
- Create a new folder and a .sql file in that folder. This folder should be in the library folder or any subfolder you choose under the library folder.
- The name of the file is the name of the view that will be created/updated to store the result of your sql code e.g., carriers.sql means we will create/update a view called carriers with the results of your sql code.
- Add your sql to it. This can be any valid SparkSQL and can refer to any view created by the pipeline before this transformer is run. For example:
SELECT carrier, crsarrtime FROM flights
- Run the generate_proxies command as shown in the Generating Proxies section below
- Now go to your Pipeline class init and add to self.transformers. Start the folder name and hit ctrl-space for PyCharm to autocomplete the name
- That's it. Your sql has been automaticaly wrapped in a Transformer which will do logging, monitor performance and do error checking
To Add a Python transformation
- Create a new folder and .py file in that folder. This folder should be in the library folder or any subfolder you choose under the library folder.
- In the .py file, create a new class and derive from Transformer (from spark ML). Implement the _transform() function For example:
from typing import Optional, Dict, Any
from pyspark import keyword_only
from pyspark.sql.dataframe import DataFrame
from spark_pipeline_framework.progress_logger.progress_logger import ProgressLogger
from spark_pipeline_framework.proxy_generator.python_proxy_base import PythonProxyBase
class FeatureTransformer(PythonProxyBase):
# noinspection PyUnusedLocal
@capture_parameters
def __init__(self,
name: str = None,
parameters: Optional[Dict[str, Any]] = None,
progress_logger: Optional[ProgressLogger] = None,
verify_count_remains_same: bool = False
) -> None:
super(FeatureTransformer, self).__init__(name=name,
parameters=parameters,
progress_logger=progress_logger,
verify_count_remains_same=verify_count_remains_same)
def _transform(self, df: DataFrame) -> DataFrame:
pass
- Run the generate_proxies command as shown in the Generating Proxies section below
- Now go to your Pipeline class init and add to self.transformers. Start the folder name and hit ctrl-space for PyCharm to autocomplete the name
To Add a Machine Learning training transformation (called fit
or Estimator
in SparkML lingo)
- Create a new folder and .py file in that folder. This folder should be in the library folder or any subfolder you choose under the library folder.
- In the .py file, create a new class and derive from Estimator (from spark ML). Implement the fit() function
- Run the generate_proxies command as shown in the Generating Proxies section below
- Now go to your Pipeline class init and add to self.estimators. Start the folder name and hit ctrl-space for PyCharm to autocomplete the name
To Add a Machine Learning prediction transformation
- Create a new folder and .py file in that folder. This folder should be in the library folder or any subfolder you choose under the library folder.
- In the .py file, create a new class and derive from Estimator (from spark ML). Implement the _transform() function. Note that that can be the same class you use for training and prediction.
- Run the generate_proxies command as shown in the Generating Proxies section below
- Now go to your Pipeline class init and add to self.transformers. Start the folder name and hit ctrl-space for PyCharm to autocomplete the name
Including pipelines in other pipelines
Pipelines are fully composable so you can include one pipeline as a transformer in another pipeline. For example:
class MyPipeline(FrameworkPipeline):
def __init__(self, parameters: Dict[str, Any], progress_logger: ProgressLogger):
super(MyPipeline, self).__init__(parameters=parameters,
progress_logger=progress_logger)
self.transformers = self.create_steps([
FrameworkCsvLoader(
view="flights",
path_to_csv=parameters["flights_path"]
),
PipelineFoo(parameters=parameters).transformers,
FeaturesCarriers(parameters=parameters).transformers,
])
Generating Proxies
- Run the following command to generate proxy classes. These automatically wrap your sql with a Spark Transformer that can be included in a Pipeline with no additional code.
python3 spark_pipeline_framework/proxy_generator/generate_proxies.py
.
You can also add this to your project Makefile to make it easier to run:
.PHONY:proxies
proxies:
python3 spark_pipeline_framework/proxy_generator/generate_proxies.py
Testing
Test a pipeline
A pipeline can be tested by providing test data in csv (or parquet), running the pipeline and then asserting for data in any view or dataframe.
from pathlib import Path
from typing import Dict, Any
from pyspark.sql.dataframe import DataFrame
from pyspark.sql.session import SparkSession
from pyspark.sql.types import StructType
from library.features.carriers.v1.features_carriers_v1 import FeaturesCarriersV1
from library.features.carriers_python.v1.features_carriers_python_v1 import FeaturesCarriersPythonV1
from spark_pipeline_framework.pipelines.framework_pipeline import FrameworkPipeline
from spark_pipeline_framework.progress_logger.progress_logger import ProgressLogger
from spark_pipeline_framework.transformers.framework_csv_loader import FrameworkCsvLoader
from spark_pipeline_framework.utilities.flattener import flatten
class MyPipeline(FrameworkPipeline):
def __init__(self, parameters: Dict[str, Any], progress_logger: ProgressLogger):
super(MyPipeline, self).__init__(parameters=parameters,
progress_logger=progress_logger)
self.transformers = self.create_steps([
FrameworkCsvLoader(
view="flights",
path_to_csv=parameters["flights_path"],
progress_logger=progress_logger
),
FeaturesCarriersV1(parameters=parameters, progress_logger=progress_logger).transformers,
FeaturesCarriersPythonV1(parameters=parameters, progress_logger=progress_logger).transformers
])
def test_can_run_framework_pipeline(spark_session: SparkSession) -> None:
# Arrange
data_dir: Path = Path(__file__).parent.joinpath('./')
flights_path: str = f"file://{data_dir.joinpath('flights.csv')}"
schema = StructType([])
df: DataFrame = spark_session.createDataFrame(
spark_session.sparkContext.emptyRDD(), schema)
spark_session.sql("DROP TABLE IF EXISTS default.flights")
# Act
parameters = {
"flights_path": flights_path
}
with ProgressLogger() as progress_logger:
pipeline: MyPipeline = MyPipeline(parameters=parameters, progress_logger=progress_logger)
transformer = pipeline.fit(df)
transformer.transform(df)
# Assert
result_df: DataFrame = spark_session.sql("SELECT * FROM flights2")
result_df.show()
assert result_df.count() > 0
Testing a single Transformer directly
Each Transformer can be tested individually by setting up the data to pass into it (e.g., loading from csv) and then testing the result of running the transformer.
from pathlib import Path
from pyspark.sql.dataframe import DataFrame
from pyspark.sql.session import SparkSession
from pyspark.sql.types import StructType
from spark_pipeline_framework.transformers.framework_csv_loader import FrameworkCsvLoader
from spark_pipeline_framework.utilities.attr_dict import AttrDict
from library.features.carriers.v1.features_carriers_v1 import FeaturesCarriersV1
def test_carriers_v1(spark_session: SparkSession):
# Arrange
data_dir: Path = Path(__file__).parent.joinpath('./')
flights_path: str = f"file://{data_dir.joinpath('flights.csv')}"
schema = StructType([])
df: DataFrame = spark_session.createDataFrame(
spark_session.sparkContext.emptyRDD(), schema)
spark_session.sql("DROP TABLE IF EXISTS default.flights")
FrameworkCsvLoader(
view="flights",
path_to_csv=flights_path
).transform(dataset=df)
parameters = {}
FeaturesCarriersV1(parameters=parameters).transformers[0].transform(dataset=df)
result_df: DataFrame = spark_session.sql("SELECT * FROM flights2")
result_df.show()
assert result_df.count() > 0
Contributing
Run make init
This will install Java, Scala, Spark and other packages
Publishing a new package
- Create a new release
- The GitHub Action should automatically kick in and publish the package
- You can see the status in the Actions tab
Asynchronous Processing
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for sparkpipelineframework-3.0.23.tar.gz
Algorithm | Hash digest | |
---|---|---|
SHA256 | 31c38cd311bb071121f14488fa85e74675752e38f53c75605ad3f28027df99a9 |
|
MD5 | 95dca2e3036ebc37965f3776b4efc0e0 |
|
BLAKE2b-256 | cea13ce5e6a363cdb93aafaf4d4f83fb39764b910e3c5e18a138af2fb431588b |
Hashes for sparkpipelineframework-3.0.23-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 2cba06dbd24f2c0714cbe8f13a592b6b149886de45a8f551cd7e13660c89c1aa |
|
MD5 | 9060fe3d677bb7ada74d2dca28f62082 |
|
BLAKE2b-256 | cfd9e6204ebae68c4cf9e206d0647f342ac50707255a7f69bef60f764d084baf |