/ Stream

A stream processing library porting the ideas from Kafka Streams to Python

A stream processing library porting the ideas from Kafka Streams to Python

Python Stream Processing

Faust is a stream processing library, porting the ideas from Kafka Streams to Python.

It is used at Robinhood to build high performance distributed systems and real-time data pipelines that process billions of events every day.

Faust provides both stream processing and event processing, sharing similarity with tools such as Kafka Streams, Apache Spark/Storm/Samza/Flink,

It does not use a DSL, it's just Python! This means you can use all your favorite Python libraries when stream processing: NumPy, PyTorch, Pandas, NLTK, Django, Flask, SQLAlchemy, ++

Faust requires Python 3.6 or later for the new async/await syntax, and variable type annotations.

Here's an example processing a stream of incoming orders:

app = faust.App('myapp', broker='kafka://localhost')

# Models describe how messages are serialized:
# {"account_id": "3fae-...", amount": 3}
class Order(faust.Record):
    account_id: str
    amount: int

@app.agent(value_type=Order)
async def order(orders):
    async for order in orders:
        # process infinite stream of orders.
        print(f'Order for {order.account_id}: {order.amount}')

The Agent decorator defines a "stream processor" that essentially consumes from a Kafka topic and does something for every event it receives.

The agent is an async def function, so can also perform other operations asynchronously, such as web requests.

This system can persist state, acting like a database. Tables are named distributed key/value stores you can use as regular Python dictionaries.

Tables are stored locally on each machine using a super fast embedded database written in C++, called RocksDB.

Tables can also store aggregate counts that are optionally "windowed" so you can keep track of "number of clicks from the last day," or "number of clicks in the last hour." for example. Like Kafka Streams, we support tumbling, hopping and sliding windows of time, and old windows can be expired to stop data from filling up.

For reliability we use a Kafka topic as "write-ahead-log". Whenever a key is changed we publish to the changelog. Standby nodes consume from this changelog to keep an exact replica of the data and enables instant recovery should any of the nodes fail.

To the user a table is just a dictionary, but data is persisted between restarts and replicated across nodes so on failover other nodes can take over automatically.

You can count page views by URL:

# data sent to 'clicks' topic sharded by URL key.
# e.g. key="http://example.com" value="1"
click_topic = app.topic('clicks', key_type=str, value_type=int)

# default value for missing URL will be 0 with `default=int`
counts = app.Table('click_counts', default=int)

@app.agent(click_topic)
async def count_click(clicks):
    async for url, count in clicks.items():
        counts[url] += count

The data sent to the Kafka topic is partitioned, which means the clicks will be sharded by URL in such a way that every count for the same URL will be delivered to the same Faust worker instance.

Faust supports any type of stream data: bytes, Unicode and serialized structures, but also comes with "Models" that use modern Python syntax to describe how keys and values in streams are serialized:

# Order is a json serialized dictionary,
# having these fields:

class Order(faust.Record):
    account_id: str
    product_id: str
    price: float
    quantity: float = 1.0

orders_topic = app.topic('orders', key_type=str, value_type=Order)

@app.agent(orders_topic)
async def process_order(orders):
    async for order in orders:
        # process each order using regular Python
        total_price = order.price * order.quantity
        await send_order_received_email(order.account_id, order)

Faust is statically typed, using the mypy type checker, so you can take advantage of static types when writing applications.

The Faust source code is small, well organized, and serves as a good resource for learning the implementation of Kafka Streams.

GitHub