Easily build performant data stream processing pipelines in Python.
Bowline is a Python library that simplifies creating data pipelines that perform sequential computations on data streams.
Key features of Bowline include:
- Performance: Each Bowline
Processorruns in its own process, meaning that Bowline is ideal for high throughput and computationally heavy workloads. - Simplicity: Bowline abstracts away the complexity of
multiprocessingby handling process creation, inter-process communication, and clean process shutdown. - Typing: Data inputs and ouputs are validated with
pydantic. - Create Complex Pipelines: Bowline provides a
ProcessorChainthat chains processes together, and aProcessGraphto enable the creation of more complex, branching pipelines.
python -m pip install bowline-streaming
The following section describe the Processor, ProcessorChain, and ProcessorGraph classes.
Processors are the "building blocks" of the process chains. A Processor defines the function to be executed, data input and output formats, and a processor name. Processors are executed in a background process, and data is transferred via process-safe queues.
The following example shows creating and using a Processor that adds two numbers and prints the results.
from pydantic import BaseModel
from bowline import Processor
class AddInputModel(BaseModel):
x: int
y: int
class AddOutputModel(BaseModel):
result: int
def add_and_print(input: AddInputModel) -> AddOutputModel:
result = input.x + input.y
return AddOutputModel(result=result)
if __name__ == '__main__':
# Create and start processor
addition_processor = Processor(target_function=add_and_print,
name="add",
input_model=AddInputModel,
output_model=AddOutputModel)
addition_processor.start()
# Push data to the
addition_processor.push_input(AddInputModel(x=2, y=2))
addition_processor.push_input(AddInputModel(x=3, y=4))
addition_processor.push_input(AddInputModel(x=123, y=456))
# Get the results
print("Results: ")
for _ in range(3): # We pushed 3 inputs, we expect 3 outputs
# Wait for results to be ready
while not addition_processor.has_output():
pass
result = addition_processor.get_output()
print(result)
# Stop the processor
addition_processor.shutdown()Processor Chains connect multiple Processors in a chain, such that the output of a Processor is passed as the input to the next. This allows for the definition of pipelines that can run in sequence on streaming data.
The following example shows creating and using a ProcessorChain that adds two numbers, then squares the result.
from pydantic import BaseModel
from bowline import Processor
from bowline import ProcessorChain
class AddInputModel(BaseModel):
x: int
y: int
class AddOutputModel(BaseModel):
result: int
class SquareOutputModel(BaseModel):
result: int
def add_two_numbers(input: AddInputModel) -> AddOutputModel:
result = input.x + input.y
return AddOutputModel(result=result)
def square_number(input: AddOutputModel) -> SquareOutputModel:
result = input.result * input.result
return SquareOutputModel(result=result)
if __name__ == '__main__':
# Create processors
add_two_numbers_processor = Processor(target_function=add_two_numbers,
name="add",
input_model=AddInputModel,
output_model=AddOutputModel)
square_number_processor = Processor(target_function=square_number,
name="square",
input_model=AddOutputModel,
output_model=SquareOutputModel)
# Create process chain
processor_chain = ProcessorChain()
processor_chain.add_processor(add_two_numbers_processor)
processor_chain.add_processor(square_number_processor)
# Start the processor chain
processor_chain.start()
# Push some data to the chain. This will add the numbers, then square them.
processor_chain.push_input(AddInputModel(x=2, y=2))
processor_chain.push_input(AddInputModel(x=3, y=4))
processor_chain.push_input(AddInputModel(x=123, y=456))
# Get the results
print(f"Results: ")
for _ in range(3):
# Wait for output
while not processor_chain.has_output():
pass
print(processor_chain.get_output())
# Shut down the processor chain
processor_chain.shutdown()Processor Graphs connect multiple Processors in a graph, such that the output of a Processor can be passed to multiple subsequent Processors. This allows for the definition of pipelines that can run with an arbitrary number of branching processors within them.
The following example creates a ProcessorGraph in which two numbers are added together, then the result of that
calculation is passed to two subsequent processors: one which squares that result, and the other of which computes its
square root.
import math
from pydantic import BaseModel
from bowline import Processor, ProcessorGraph
class AddInputModel(BaseModel):
x: int
y: int
class AddOutputModel(BaseModel):
result: int
class SquareOutputModel(BaseModel):
result: int
class SquareRootOutputModel(BaseModel):
result: float
def add_two_numbers(input: AddInputModel) -> AddOutputModel:
result = input.x + input.y
print(f"{input.x} + {input.y} = {result}")
return AddOutputModel(result=result)
def square_number(input: AddOutputModel) -> SquareOutputModel:
result = input.result * input.result
print(f"{input.result} squared is {result}")
return SquareOutputModel(result=result)
def square_root(input: AddOutputModel) -> SquareRootOutputModel:
result = math.sqrt(input.result)
print(f"The square root of {input.result} is {result}")
return SquareRootOutputModel(result=result)
if __name__ == '__main__':
# Create processors
add_two_numbers_processor = Processor(target_function=add_two_numbers,
name="addition",
input_model=AddInputModel,
output_model=AddOutputModel)
square_number_processor = Processor(target_function=square_number,
name="square",
input_model=AddOutputModel,
output_model=SquareOutputModel)
square_root_processor = Processor(target_function=square_root,
name="sqrt",
input_model=AddOutputModel,
output_model=SquareRootOutputModel)
# Create ProcessorGraph
# This processor graph will run add_two_numbers(), then run square_number() and square_root() on its result.
processor_graph = ProcessorGraph()
processor_graph.add_processor(add_two_numbers_processor)
processor_graph.add_processor(square_number_processor, add_two_numbers_processor)
processor_graph.add_processor(square_root_processor, add_two_numbers_processor)
# Start the ProcessorGraph
processor_graph.start()
# Push input to graph
processor_graph.push_input(AddInputModel(x=2, y=2))
processor_graph.push_input(AddInputModel(x=3, y=4))
processor_graph.push_input(AddInputModel(x=123, y=456))
# Get results
for _ in range(6): # We provided 3 inputs, and there are 2 terminal Processors, so 6 total results
# Wait for results to be available
while not processor_graph.has_output():
pass
result = processor_graph.get_output()
print(f"Received output {result.output} from processor {result.processor}")
# Shut down processors
processor_graph.shutdown()You can also create pipelines via a yaml configuration file, for both ProcessorChains and ProcessorGraphs.
Example yaml configuration file, in which the add Processor provides input to the square Processor.
chain:
processors:
- add:
target_function: simple_chain.add_two_numbers
input_model: simple_chain.AddInputModel
output_model: simple_chain.AddOutputModel
- square:
target_function: simple_chain.square_number
input_model: simple_chain.AddOutputModel
output_model: simple_chain.SquareOutputModelExample code:
from bowline.utils.config import ProcessorConfig
# Create process chain from config file
config_file_path = "examples/chain-config.yml"
config = ProcessorConfig(config_file_path)
processor_chain = config.generate_processors()
# Start the processor chain
processor_chain.start()Example yaml configuration file, which definesa graph in which the addition Processor provides inputs to the square and sqrt Processors:
graph:
processors:
- addition:
target_function: simple_graph.add_two_numbers
input_model: simple_graph.AddInputModel
output_model: simple_graph.AddOutputModel
processors:
- square:
target_function: simple_graph.square_number
input_model: simple_graph.AddOutputModel
output_model: simple_graph.SquareOutputModel
- sqrt:
target_function: simple_graph.square_root
input_model: simple_graph.AddOutputModel
output_model: simple_graph.SquareRootOutputModelExample code:
from bowline.utils.config import ProcessorConfig
# Create process chain from config file
config_file_path = "examples/graph-config.yml"
config = ProcessorConfig(config_file_path)
processor_graph = config.generate_processors()
# Start the ProcessorGraph
processor_graph.start()Because Bowline uses multiprocessing behind the scenes, all data models must be serializable.
- Clone the repository:
git clone [email protected]:scottbarnesg/bowline.git - Install
bowlineas an interactive package:python -m pip install -e .
- Install the test dependencies:
python -m pip install -e .[dev] - Run the tests:
python -m pytest test/ - Or, to view log output while running tests:
python -m pytest --capture=no --log-cli-level=INFO test/test_processor.py
Why doesn't Bowline use asyncio?
Bowline is built on the multiprocessing library, so each Processor instance runs in its own process.
This means that async functionality is generally not needed (although you are welcome to implement it in the functions you want Bowline to run).
Bowline's primary use case is creating high-throughput, low-latency pipelines for streaming data that you want to perform computationally-heavy operations on. If you are looking for an async-first library that runs tasks in a single process and is designed for lighter workloads, a tool like Faust may be a better fit for you use case.