Skip to content

Update readme #13

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
pyper-dev merged 2 commits into from
Dec 9, 2024
Merged

Update readme #13

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
cd959bb
update readme
RichardZhu2 Dec 9, 2024
73098b7
Merge branch 'dev' of https://github.com/RichardZhu2/pyper into dev
RichardZhu2 Dec 9, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
221 changes: 55 additions & 166 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,7 +28,7 @@ See the [Documentation](https://pyper-dev.github.io/pyper/)

Key features:

* 💡**Intuitive API**: Easy to learn, easy to think about. Implements clean abstractions to seamlessly unify threaded and asynchronous work.
* 💡**Intuitive API**: Easy to learn, easy to think about. Implements clean abstractions to seamlessly unify threaded, multiprocessed, and asynchronous work.
* 🚀 **Functional Paradigm**: Python functions are the building blocks of data pipelines. Let's you write clean, reusable code naturally.
* 🛡️ **Safety**: Hides the heavy lifting of underlying task execution and resource clean-up. No more worrying about race conditions, memory leaks, or thread-level error handling.
* ⚡ **Efficiency**: Designed from the ground up for lazy execution, using queues, workers, and generators.
Expand All @@ -46,235 +46,124 @@ Note that `python-pyper` is the [pypi](https://pypi.org/project/python-pyper) re

## Usage

In Pyper, the `task` decorator is used to transform functions into composable pipelines.

Let's simulate a pipeline that performs a series of transformations on some data.

```python
import asyncio
import time
from typing import AsyncIterable

from pyper import task


def step1(limit: int):
"""Generate some data."""
def get_data(limit: int):
for i in range(limit):
yield i


async def step2(data: int):
"""Simulate some asynchronous work."""
async def step1(data: int):
await asyncio.sleep(1)
print("Finished async sleep")
return data + 1
print("Finished async wait", data)
return data


def step3(data: int):
"""Simulate some IO-bound (non awaitable) work."""
def step2(data: int):
time.sleep(1)
print("Finished sync sleep")
return 2 * data - 1
print("Finished sync wait", data)
return data


async def print_sum(data: AsyncIterable[int]):
"""Print the sum of values from a data stream."""
total = 0
async for output in data:
total += output
print("Total ", total)
def step3(data: int):
for i in range(10_000_000):
_ = i*i
print("Finished heavy computation", data)
return data


async def main():
# Define a pipeline of tasks using `pyper.task`
run = task(step1) | task(step2, concurrency=20) | task(step3, concurrency=20) > print_sum
await run(limit=20)


if __name__ == "__main__":
asyncio.run(main()) # takes ~2 seconds
```

Pyper provides an elegant abstraction of the concurrent execution of each function via `pyper.task`, allowing you to focus on building out the **logical** functions of your program.

In our pipeline:

* `task(step1)` generates 20 data values

* `task(step2, concurrency=20)` spins up 20 asynchronous workers, taking each value as input and returning an output

* `task(step3, concurrency=20)` spins up 20 threaded workers, taking each value as input and returning an output

The script therefore takes ~2 seconds to complete, as `step2` and `step3` in the pipeline only take the 1 second of sleep time, performed concurrently. If you'd like, experiment with tweaking the `limit` and `concurrency` values for yourself.

---

<details markdown="1">
<summary><u>What does the logic translate to in non-concurrent code?</u></summary>

<br>

Having defined the logical operations we want to perform on our data as functions, all we are doing is piping the output of one function to the input of another. In sequential code, this could look like:

```python
# Analogous to:
# pipeline = task(step1) | task(step2) | task(step3)
async def pipeline(limit: int):
for data in step1(limit):
data = await step2(data)
data = step3(data)
yield data
pipeline = task(get_data, branch=True) \
| task(step1, workers=20) \
| task(step2, workers=20) \
| task(step3, workers=20, multiprocess=True)


# Analogous to:
# run = pipeline > print_sum
async def run(limit: int):
await print_sum(pipeline(limit))
# Call the pipeline
total = 0
async for output in pipeline(limit=20):
total += output
print("Total:", total)


async def main():
await run(20) # takes ~40 seconds
if __name__ == "__main__":
asyncio.run(main())
```

Pyper uses the `|` (motivated by Unix's pipe operator) syntax as a representation of this input-output piping between tasks.
Pyper provides an elegant abstraction of the execution of each function via `pyper.task`, allowing you to focus on building out the **logical** functions of your program. In the `main` function:

</details>
* `pipeline` defines a function; this takes the parameters of its first task (`get_data`) and yields each output from its last task (`step3`)
* Tasks are piped together using the `|` operator (motivated by Unix's pipe operator) as a syntactic representation of passing inputs/outputs between tasks.

<details markdown="1">
<summary><u>What would the implementation look like without Pyper?</u></summary>
In the pipeline, we are executing three different types of work:

<br>

Concurrent programming in Python is notoriously difficult to get right. In a concurrent data pipeline, some challenges are:

* We want producers to concurrently execute tasks and send results to the next stage as soon as it's done processing
* We want consumers to lazily pick up output as soon as it's available from the previous stage
* We need to somehow unify the execution of threads and coroutines, without letting non-awaitable tasks clog up the event-loop
* `task(step1, workers=20)` spins up 20 `asyncio.Task`s to handle asynchronous IO-bound work

The basic approach to doing this is by using queues-- a simplified and very unabstracted implementation could be:
* `task(step2, workers=20)` spins up 20 `threads` to handle synchronous IO-bound work

```python
async def pipeline(limit: int):
q1 = asyncio.Queue()
q2 = asyncio.Queue()
q3 = asyncio.Queue()

step2_concurrency=20
step3_concurrency=20

async def worker1():
for data in step1(limit):
await q1.put(data)
for _ in range(step2_concurrency):
await q1.put(None)

worker2s_finished = 0
async def worker2():
nonlocal worker2s_finished
while True:
data = await q1.get()
if data is None:
break
output = await step2(data)
await q2.put(output)
worker2s_finished += 1
if worker2s_finished == step2_concurrency:
for _ in range(step3_concurrency):
await q2.put(None)

worker3s_finished = 0
async def worker3():
nonlocal worker3s_finished
loop = asyncio.get_running_loop()
while True:
data = await q2.get()
if data is None:
break
# Pyper uses a custom thread group handler instead of run_in_executor
output = await loop.run_in_executor(None, step3, data)
await q3.put(output)
worker3s_finished += 1
if worker3s_finished == step3_concurrency:
await q3.put(None)

async with asyncio.TaskGroup() as tg:
# Start all workers in the background
tg.create_task(worker1())
for _ in range(step2_concurrency):
tg.create_task(worker2())
for _ in range(step3_concurrency):
tg.create_task(worker3())
# Yield data until all workers have stopped
while True:
data = await q3.get()
if data is None:
break
yield data


async def run(limit: int):
await print_sum(pipeline(limit))
* `task(step3, workers=20, multiprocess=True)` spins up 20 `processes` to handle synchronous CPU-bound work

`task` acts as one intuitive API for unifying the execution of each different type of function.

async def main():
await run(20) # takes ~2 seconds
```
Each task submits their outputs to the next task within the pipeline via queue-based data structures, which is the mechanism underpinning how concurrency and parallelism are achieved. See the [docs](https://pyper-dev.github.io/pyper/docs/UserGuide/BasicConcepts) for a breakdown of what a pipeline looks like under the hood.

This implementation achieves the basic desired concurrent data flow, but still lacks some quality-of-life features that Pyper takes care of, like error handling within threads.

Pyper handles the complexities of managing queues and workers, so that this code can be reduced to the two-line main function in the example above.
---

</details>

<details markdown="1">
<summary><u>Do I have to use <code>async</code>?</u></summary>
<summary><u>See a non-async example</u></summary>

<br>

No-- not every program is asynchronous, so Pyper pipelines are by default synchronous, as long as their tasks are defined as synchronous functions. For example:
Pyper pipelines are by default non-async, as long as their tasks are defined as synchronous functions. For example:

```python
import time
from typing import Iterable

from pyper import task


def step1(limit: int):
def get_data(limit: int):
for i in range(limit):
yield i


def step2(data: int):
def step1(data: int):
time.sleep(1)
return data + 1

print("Finished sync wait", data)
return data

def step3(data: int):
time.sleep(1)
return 2 * data - 1
def step2(data: int):
for i in range(10_000_000):
_ = i*i
print("Finished heavy computation", data)
return data


def print_sum(data: Iterable[int]):
def main():
pipeline = task(get_data, branch=True) \
| task(step1, workers=20) \
| task(step2, workers=20, multiprocess=True)
total = 0
for output in data:
for output in pipeline(limit=20):
total += output
print("Total ", total)


def main():
run = task(step1) \
| task(step2, concurrency=20) \
| task(step3, concurrency=20) \
> print_sum
# Run synchronously
run(limit=20)
print("Total:", total)


if __name__ == "__main__":
main() # takes ~2 seconds
main()
```

A pipeline consisting of _at least one asynchronous function_ becomes an `AsyncPipeline`, which exposes the same logical function, provided `async` and `await` syntax in all of the obvious places. This makes it effortless to unify synchronously defined and asynchronously defined functions where need be.
A pipeline consisting of _at least one asynchronous function_ becomes an `AsyncPipeline`, which exposes the same usage API, provided `async` and `await` syntax in the obvious places. This makes it effortless to combine synchronously defined and asynchronously defined functions where need be.

</details>

Expand Down