༄ streamable

concurrent & fluent interface for (async) iterables

• 🇹 Stream[T] is a decorator for Iterable[T] / AsyncIterable[T]
• 🔗 lazy operations chaining
• 🔀 concurrent via threads / processes / coroutines (async)
• 🛡️ battle-tested with Python 3.7 to 3.15 (compatible with PyPy)

1. install

(zero dependencies)

pip install streamable

2. import

from streamable import Stream

3. init

Create a Stream[T] decorating an Iterable[T]/AsyncIterable[T]:

integers: Stream[int] = Stream(range(10))

4. operate

Chain lazy operations (only evaluated during iteration), each returning a new immutable Stream:

inverses: Stream[float] = (
    integers
    .map(lambda n: round(1 / n, 2))
    .catch(ZeroDivisionError, replacement=float("inf"))
)

5. iterate

Iterate over a Stream[T] like any Iterable[T]/AsyncIterable[T]:

>>> list(inverses)
[inf, 1.0, 0.5, 0.33, 0.25, 0.2, 0.17, 0.14, 0.12, 0.11]

Elements are processed on-the-fly as the iteration advances.

↔ showcase: ETL

Let's walk through the Stream's features with an Extract-Transform-Load script:

This toy script gets Pokémons concurrently from PokéAPI, and writes the quadrupeds from the first three generations into a CSV file, in 5-seconds batches:

import csv
from datetime import timedelta
import itertools
import requests
from streamable import Stream

with open("./quadruped_pokemons.csv", mode="w") as file:
    fields = ["id", "name", "is_legendary", "base_happiness", "capture_rate"]
    writer = csv.DictWriter(file, fields, extrasaction='ignore')
    writer.writeheader()

    pipeline: Stream = (
        # Infinite Stream[int] of Pokemon ids starting from Pokémon #1: Bulbasaur
        Stream(itertools.count(1))
        # Limit to 16 requests per second to be friendly to our fellow PokéAPI devs
        .throttle(16, per=timedelta(seconds=1))
        # GET pokemons concurrently using a pool of 8 threads
        .map(lambda poke_id: f"https://pokeapi.co/api/v2/pokemon-species/{poke_id}")
        .map(requests.get, concurrency=8)
        .foreach(requests.Response.raise_for_status)
        .map(requests.Response.json)
        # Stop the iteration when reaching the 1st pokemon of the 4th generation
        .truncate(when=lambda poke: poke["generation"]["name"] == "generation-iv")
        .observe("pokemons")
        # Keep only quadruped Pokemons
        .filter(lambda poke: poke["shape"]["name"] == "quadruped")
        .observe("quadruped pokemons")
        # Write a batch of pokemons every 5 seconds to the CSV file
        .group(interval=timedelta(seconds=5))
        .foreach(writer.writerows)
        .flatten()
        .observe("written pokemons")
        # Catch exceptions and raises the 1st one at the end of the iteration
        .catch(Exception, finally_raise=True)
    )

    # Start a full iteration
    pipeline()

... or the async way!

Let's write an async version of this script:

• using httpx.AsyncCLient together with the .amap operation (the async counterpart of .map).
• instead of calling pipeline() to iterate over it as an Iterable, let's await pipeline to iterate over it as an AsyncIterable.

import asyncio
import csv
from datetime import timedelta
import itertools
import httpx
from streamable import Stream

async def main() -> None:
    with open("./quadruped_pokemons.csv", mode="w") as file:
        fields = ["id", "name", "is_legendary", "base_happiness", "capture_rate"]
        writer = csv.DictWriter(file, fields, extrasaction='ignore')
        writer.writeheader()

        async with httpx.AsyncClient() as http:
            pipeline: Stream = (
                # Infinite Stream[int] of Pokemon ids starting from Pokémon #1: Bulbasaur
                Stream(itertools.count(1))
                # Limit to 16 requests per second to be friendly to our fellow PokéAPI devs
                .throttle(16, per=timedelta(seconds=1))
                # GET pokemons via 8 concurrent coroutines
                .map(lambda poke_id: f"https://pokeapi.co/api/v2/pokemon-species/{poke_id}")
                .amap(http.get, concurrency=8)
                .foreach(httpx.Response.raise_for_status)
                .map(httpx.Response.json)
                # Stop the iteration when reaching the 1st pokemon of the 4th generation
                .truncate(when=lambda poke: poke["generation"]["name"] == "generation-iv")
                .observe("pokemons")
                # Keep only quadruped Pokemons
                .filter(lambda poke: poke["shape"]["name"] == "quadruped")
                .observe("quadruped pokemons")
                # Write a batch of pokemons every 5 seconds to the CSV file
                .group(interval=timedelta(seconds=5))
                .foreach(writer.writerows)
                .flatten()
                .observe("written pokemons")
                # Catch exceptions and raises the 1st one at the end of the iteration
                .catch(Exception, finally_raise=True)
            )

            # Start a full async iteration
            await pipeline

asyncio.run(main())

🙏 Highlights from the community

• Tryolabs' Top 10 Python libraries of 2024 (LinkedIn, Reddit)
• PyCoder’s Weekly x Real Python
• @PythonHub's tweet
• Upvoters on our showcase Reddit post

💡 Notes

Exceptions are not terminating the iteration

Tip

If an operation raises an exception while processing an element, you can handle it and continue the iteration:

👀 show snippet

from contextlib import suppress

casted_ints: Iterator[int] = iter(
    Stream("0123_56789")
    .map(int)
    .group(3)
    .flatten()
)
collected: List[int] = []

with suppress(ValueError):
    collected.extend(casted_ints)
assert collected == [0, 1, 2, 3]

collected.extend(casted_ints)
assert collected == [0, 1, 2, 3, 5, 6, 7, 8, 9]

Performances

Declaring a Stream is lazy,

odd_int_strings = Stream(range(1_000_000)).filter(lambda n: n % 2).map(str)

and there is zero overhead during iteration compared to builtins, iter(odd_int_strings) visits the operations lineage and returns exactly this iterator:

map(str, filter(lambda n: n % 2, range(1_000_000)))

Operations have been implemented with speed in mind. If you have any ideas for improvement, whether performance-related or not, an issue, PR, or discussion would be very much appreciated! 🙏 (CONTRIBUTING.md)

📒 Operations

.map	transform elements
.foreach	call a side effect function on elements
.group / .groupby	batch a certain number of elements, by a given key, over a time interval
.flatten	explode iterable elements
.filter	remove elements
.distinct	remove duplicates
.truncate	cut the stream
.skip	ignore head elements
.catch	handle exceptions
.throttle	control the rate of iteration
.observe	log elements/errors counters

Important

A Stream exposes a minimalist yet expressive set of operations to manipulate its elements, but creating its source or consuming it is not its responsability, it's meant to be combined with specialized libraries (csv, json, pyarrow, psycopg2, boto3, requests, httpx, ...).

Note

async counterparts: For each operation that takes a function (such as .map), there is an equivalent that accepts an async function (such as .amap). You can freely mix synchronous and asynchronous operations within the same Stream. The result can then be consumed either as an Iterable or as an AsyncIterable. When a stream involving async operations is consumed as an Iterable, a temporary asyncio event loop is attached to it.

🟡 .map / .amap

Applies a transformation on elements:

👀 show snippet

integer_strings: Stream[str] = integers.map(str)

assert list(integer_strings) == ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']

thread-based concurrency

Applies the transformation via concurrency threads.

👀 show snippet

import requests

pokemon_names: Stream[str] = (
    Stream(range(1, 4))
    .map(lambda i: f"https://pokeapi.co/api/v2/pokemon-species/{i}")
    .map(requests.get, concurrency=3)
    .map(requests.Response.json)
    .map(lambda poke: poke["name"])
)
assert list(pokemon_names) == ['bulbasaur', 'ivysaur', 'venusaur']

Note

Memory-efficient: Only concurrency upstream elements are pulled for processing; the next upstream element is pulled only when a result is yielded downstream.

Note

Ordering: it yields results in the upstream order (FIFO), set ordered=False to yield results as they become available (First Done, First Out).

process-based concurrency

Same but set via="process":

👀 show snippet

if __name__ == "__main__":
    state: List[int] = []
    # integers are mapped
    assert integers.map(state.append, concurrency=4, via="process").count() == 10
    # but the `state` of the main process is not mutated
    assert state == []

coroutine-based concurrency: .amap

.amap can apply an async function concurrently.

👀 show snippet

import asyncio
import httpx

async def main() -> None:
    async with httpx.AsyncClient() as http:
        pokemon_names: Stream[str] = (
            Stream(range(1, 4))
            .map(lambda i: f"https://pokeapi.co/api/v2/pokemon-species/{i}")
            .amap(http.get, concurrency=3)
            .map(httpx.Response.json)
            .map(lambda poke: poke["name"])
        )
        # consume as an AsyncIterable[str]
        assert [name async for name in pokemon_names] == ['bulbasaur', 'ivysaur', 'venusaur']

asyncio.run(main())

"starmap"

The star function decorator transforms a function that takes several positional arguments into a function that takes a tuple:

👀 show snippet

from streamable import star

zeros: Stream[int] = (
    Stream(enumerate(integers))
    .map(star(lambda index, integer: index - integer))
)

assert list(zeros) == [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

🟡 .foreach / .aforeach

Applies a side effect on elements:

👀 show snippet

state: List[int] = []
appending_integers: Stream[int] = integers.foreach(state.append)

assert list(appending_integers) == [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
assert state == [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]

concurrency

Similar to .map:

• set the concurrency parameter for thread-based concurrency
• set via="process" for process-based concurrency
• set ordered=False for First Done First Out
• The .aforeach operation can apply an async effect concurrently.

🟡 .group / .agroup

Groups into Lists

... up to a given group size:

👀 show snippet

integers_by_5: Stream[List[int]] = integers.group(size=5)

assert list(integers_by_5) == [[0, 1, 2, 3, 4], [5, 6, 7, 8, 9]]

... and/or co-groups by a given key:

👀 show snippet

integers_by_parity: Stream[List[int]] = integers.group(by=lambda n: n % 2)

assert list(integers_by_parity) == [[0, 2, 4, 6, 8], [1, 3, 5, 7, 9]]

... and/or co-groups the elements yielded by the upstream within a given time interval:

👀 show snippet

from datetime import timedelta

integers_within_1_sec: Stream[List[int]] = (
    integers
    .throttle(2, per=timedelta(seconds=1))
    .group(interval=timedelta(seconds=0.99))
)

assert list(integers_within_1_sec) == [[0, 1, 2], [3, 4], [5, 6], [7, 8], [9]]

Tip

Combine the size/by/interval parameters:

👀 show snippet

integers_by_parity_by_2: Stream[List[int]] = (
    integers
    .group(by=lambda n: n % 2, size=2)
)

assert list(integers_by_parity_by_2) == [[0, 2], [1, 3], [4, 6], [5, 7], [8], [9]]

🟡 .groupby / .agroupby

Like .group, but groups into (key, elements) tuples:

👀 show snippet

integers_by_parity: Stream[Tuple[str, List[int]]] = (
    integers
    .groupby(lambda n: "odd" if n % 2 else "even")
)

assert list(integers_by_parity) == [("even", [0, 2, 4, 6, 8]), ("odd", [1, 3, 5, 7, 9])]

Tip

Then "starmap" over the tuples:

👀 show snippet

from streamable import star

counts_by_parity: Stream[Tuple[str, int]] = (
    integers_by_parity
    .map(star(lambda parity, ints: (parity, len(ints))))
)

assert list(counts_by_parity) == [("even", 5), ("odd", 5)]

🟡 .flatten / .aflatten

Ungroups elements assuming that they are Iterables (or AsyncIterables for .aflatten):

👀 show snippet

even_then_odd_integers: Stream[int] = integers_by_parity.flatten()

assert list(even_then_odd_integers) == [0, 2, 4, 6, 8, 1, 3, 5, 7, 9]

thread-based concurrency

Concurrently flattens concurrency iterables via threads (or via coroutines for .aflatten):

👀 show snippet

mixed_ones_and_zeros: Stream[int] = (
    Stream([[0] * 4, [1] * 4])
    .flatten(concurrency=2)
)
assert list(mixed_ones_and_zeros) == [0, 1, 0, 1, 0, 1, 0, 1]

🟡 .filter / .afilter

Keeps only the elements that satisfy a condition:

👀 show snippet

even_integers: Stream[int] = integers.filter(lambda n: n % 2 == 0)

assert list(even_integers) == [0, 2, 4, 6, 8]

🟡 .distinct / .adistinct

Removes duplicates:

👀 show snippet

distinct_chars: Stream[str] = Stream("foobarfooo").distinct()

assert list(distinct_chars) == ["f", "o", "b", "a", "r"]

specifying a deduplication key:

👀 show snippet

strings_of_distinct_lengths: Stream[str] = (
    Stream(["a", "foo", "bar", "z"])
    .distinct(len)
)

assert list(strings_of_distinct_lengths) == ["a", "foo"]

Warning

During iteration, all distinct elements that are yielded are retained in memory to perform deduplication. However, you can remove only consecutive duplicates without a memory footprint by setting consecutive_only=True:

👀 show snippet

consecutively_distinct_chars: Stream[str] = (
    Stream("foobarfooo")
    .distinct(consecutive_only=True)
)

assert list(consecutively_distinct_chars) == ["f", "o", "b", "a", "r", "f", "o"]

🟡 .truncate / .atruncate

Ends iteration once a given number of elements have been yielded:

👀 show snippet

five_first_integers: Stream[int] = integers.truncate(5)

assert list(five_first_integers) == [0, 1, 2, 3, 4]

or when a condition is satisfied:

👀 show snippet

five_first_integers: Stream[int] = integers.truncate(when=lambda n: n == 5)

assert list(five_first_integers) == [0, 1, 2, 3, 4]

If both count and when are set, truncation occurs as soon as either condition is met.

🟡 .skip / .askip

Skips the first specified number of elements:

👀 show snippet

integers_after_five: Stream[int] = integers.skip(5)

assert list(integers_after_five) == [5, 6, 7, 8, 9]

or skips elements until a predicate is satisfied:

👀 show snippet

integers_after_five: Stream[int] = integers.skip(until=lambda n: n >= 5)

assert list(integers_after_five) == [5, 6, 7, 8, 9]

If both count and until are set, skipping stops as soon as either condition is met.

🟡 .catch / .acatch

Catches a given type of exception, and optionally yields a replacement value:

👀 show snippet

inverses: Stream[float] = (
    integers
    .map(lambda n: round(1 / n, 2))
    .catch(ZeroDivisionError, replacement=float("inf"))
)

assert list(inverses) == [float("inf"), 1.0, 0.5, 0.33, 0.25, 0.2, 0.17, 0.14, 0.12, 0.11]

You can specify an additional when condition for the catch:

👀 show snippet

import requests
from requests.exceptions import ConnectionError

status_codes_ignoring_resolution_errors: Stream[int] = (
    Stream(["https://github.com", "https://foo.bar", "https://github.com/foo/bar"])
    .map(requests.get, concurrency=2)
    .catch(ConnectionError, when=lambda error: "Max retries exceeded with url" in str(error))
    .map(lambda response: response.status_code)
)

assert list(status_codes_ignoring_resolution_errors) == [200, 404]

It has an optional finally_raise: bool parameter to raise the first exception caught (if any) when the iteration terminates.

Tip

Leverage when to apply side effects on catch:

👀 show snippet

errors: List[Exception] = []

def store_error(error: Exception) -> bool:
    errors.append(error)  # applies effect
    return True  # signals to catch the error

integers_in_string: Stream[int] = (
    Stream("012345foo6789")
    .map(int)
    .catch(ValueError, when=store_error)
)

assert list(integers_in_string) == [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
assert len(errors) == len("foo")

🟡 .throttle

Limits the number of yields per time interval:

👀 show snippet

from datetime import timedelta

three_integers_per_second: Stream[int] = integers.throttle(3, per=timedelta(seconds=1))

# takes 3s: ceil(10 integers / 3 per_second) - 1
assert list(three_integers_per_second) == [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]

🟡 .observe

Logs the progress of iterations:

👀 show snippet

>>> assert list(integers.throttle(2, per=timedelta(seconds=1)).observe("integers")) == [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]

INFO: [duration=0:00:00.001793 errors=0] 1 integers yielded
INFO: [duration=0:00:00.004388 errors=0] 2 integers yielded
INFO: [duration=0:00:01.003655 errors=0] 4 integers yielded
INFO: [duration=0:00:03.003196 errors=0] 8 integers yielded
INFO: [duration=0:00:04.003852 errors=0] 10 integers yielded

Note

To avoid flooding, logs are emitted only when the number of yielded elements (or errors) reaches powers of 2.

Tip

To mute these logs, set the logging level above INFO:

👀 show snippet

import logging
logging.getLogger("streamable").setLevel(logging.WARNING)

🟡 +

Concatenates streams:

👀 show snippet

assert list(integers + integers) == [0, 1, 2, 3 ,4, 5, 6, 7, 8, 9, 0, 1, 2, 3 ,4, 5, 6, 7, 8, 9]

🟡 zip

Use the builtins' zip function:

👀 show snippet

from streamable import star

cubes: Stream[int] = (
    Stream(zip(integers, integers, integers))  # Stream[Tuple[int, int, int]]
    .map(star(lambda a, b, c: a * b * c))  # Stream[int]
)

assert list(cubes) == [0, 1, 8, 27, 64, 125, 216, 343, 512, 729]

Shorthands for consuming the stream

Although consuming the stream is beyond the scope of this library, it provides two basic shorthands to trigger an iteration:

🟡 .count / .acount

.count iterates over the stream until exhaustion and returns the number of elements yielded:

👀 show snippet

assert integers.count() == 10

The .acount (async method) iterates over the stream as an AsyncIterable until exhaustion and returns the number of elements yielded:

👀 show snippet

assert asyncio.run(integers.acount()) == 10

🟡 () / await

Calling the stream iterates over it until exhaustion, and returns it:

👀 show snippet

state: List[int] = []
appending_integers: Stream[int] = integers.foreach(state.append)
assert appending_integers() is appending_integers
assert state == [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]

Awaiting the stream iterates over it as an AsyncIterable until exhaustion, and returns it:

👀 show snippet

async def test_await() -> None:
    state: List[int] = []
    appending_integers: Stream[int] = integers.foreach(state.append)
    appending_integers is await appending_integers
    assert state == [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
asyncio.run(test_await())

🟡 .pipe

Calls a function, passing the stream as first argument, followed by *args/**kwargs if any (inspired by the .pipe from pandas or polars):

👀 show snippet

import pandas as pd

(
    integers
    .observe("ints")
    .pipe(pd.DataFrame, columns=["integer"])
    .to_csv("integers.csv", index=False)
)