Streams - Infinite Data Structures

“Everything transitory is but an image.”

– Goethe, Faust: Part II


A stream is an infinite async iterable, consuming messages from a channel/topic:

async def process(stream):
    async for value in stream:

The above agent is how you usually define stream processors in Faust, but you can also create stream objects manually at any point with the caveat that this can trigger a Kafka rebalance when doing so at runtime:

stream =  # or:
async for value in stream:

The stream needs to be iterated over to be processed, it will not be active until you do.

When iterated over the stream gives deserialized values, but you can also iterate over key/value pairs (using items()), or raw messages (using events()).

Keys and values can be bytes for manual deserialization, or Model instances, and this is decided by the topic’s key_type and value_type arguments.

See also

The easiest way to process streams is to use agents, but you can also create a stream manually from any topic/channel.

Here we define a model for our stream, create a stream from the “withdrawals” topic and iterate over it:

class Withdrawal(faust.Record):
    account: str
    amount: float

async for w in app.topic('withdrawals', value_type=Withdrawal).stream():

Do note that the worker must be started first (or at least the app), for this to work, and the stream iterator needs to be started as an asyncio.Task, so a more practical example is:

import faust

class Withdrawal(faust.Record):
    account: str
    amount: float

app = faust.App('example-app')

withdrawals_topic = app.topic('withdrawals', value_type=Withdrawal)

async def mytask():
    async for w in

if __name__ == '__main__':

You may also treat the stream as a stream of bytes values:

async for value in app.topic('messages').stream():
    # the topic description has no value_type, so values
    # are now the raw message value in bytes.


A stream can have an arbitrary number of processor callbacks that are executed as values go through the stream.

These are normally used in Faust applications, but are useful for libraries that extend the functionality of streams.

A processor takes a value as argument and returns a value:

def add_default_language(value: MyModel) -> MyModel:
    if not value.language:
        value.language = 'US'
    return value

async def add_client_info(value: MyModel) -> MyModel:
    value.client = await get_http_client_info(value.account_id)
    return value

s =,
               processors=[add_default_language, add_client_info])


Processors can be async callable, or normal callable.

Since the processors are stored in an ordered list, the processors above will execute in order and the final value going out of the stream will be the reduction after all processors are applied:

async for value in s:
    # all processors applied here so `value`
    # will be equivalent to doing:
    #   value = add_default_language(add_client_info(value))

Message Life Cycle

Kafka Topics

Every Faust worker instance will start a single Kafka consumer responsible for fetching messages from all subscribed topics.

Every message in the topic have an offset number (where the first message has an offset of zero), and we use a single offset to track the messages that consumers do not want to see again.

The Kafka consumer commits the topic offsets every three seconds in a background task. The default interval is defined by the broker_commit_interval setting.

As we only have one consumer, and multiple agents can subscribe to the same topic, we need a smart way to track when those events have processed so we can commit and advance the consumer group offset.

We use reference counting for this, so when you define an agent that iterates over the topic as a stream:

async def process(stream):
    async for value in stream:

The act of starting that stream iterator will add the topic to the Conductor service. This internal service is responsible for forwarding messages received by the consumer to the streams:

[Consumer] -> [Conductor] -> [Topic] -> [Stream]

The async for is what triggers this, and the agent code above is roughly equivalent to:

async def custom_agent(app: App, topic: Topic):
     topic_iterator = aiter(topic)
     app.topics.add(topic)  # app.topics is the Conductor
     stream = Stream(topic_iterator, app=app)
     async for value in stream:

If two agents use streams subscribed to the same topic:

topic = app.topic('orders')

 async def processA(stream):
      async for value in stream:
          print(f'A: {value}')

  async def processB(stream):
       async for value in stream:
           print(f'B: {value}')

The Conductor will forward every message received on the “orders” topic to both of the agents, increasing the reference count whenever it enters an agents stream.

The reference count decreases when the event is acknowledged, and when it reaches zero the consumer will consider that offset as “done” and can commit it.


The acknowledgment signifies that the event processing is complete and should not happen again.

An event is automatically acknowledged when:

  • The agent stream advances to a new event (Stream.__anext__)

  • An exception occurs in the agent during event processing.

  • The application shuts down, or a rebalance is required, and the stream finished processing the event.

What this means is that an event is acknowledged when your agent is finished handling it, but you can also manually control when it happens.

To manually control when the event is acknowledged, and its reference count decreased, use await event.ack()

async for event in

print(event.value) await event.ack()

You can also use async for on the event:

async for event in
    async with event:
        # event acked when exiting this block

Note that the conditions in automatic acknowledgment still apply when manually acknowledging a message.

Combining streams

Streams can be combined, so that you receive values from multiple streams in the same iteration:

>>> s1 =
>>> s2 =
>>> async for value in (s1 & s2):
...     ...

Mostly this is useful when you have two topics having the same value type, but can be used in general.

If you have two streams that you want to process independently you should rather start individual tasks:

async def process_stream1(stream):
    async for value in stream:

async def process_stream2(stream):
    async for value in stream:


group_by() – Repartition the stream

The Stream.group_by() method repartitions the stream by taking a “key type” as argument:

import faust

class Order(faust.Record):
    account_id: str
    product_id: str
    amount: float
    price: float

app = faust.App('group-by-example')
orders_topic = app.topic('orders', value_type=Order)

async def process(orders):
    async for order in orders.group_by(Order.account_id):

In the example above the “key type” is a field descriptor, and the stream will be repartitioned by the account_id field found in the deserialized stream value.

The new stream will be using a new intermediate topic where messages have account ids as key, and this is the stream that the agent will finally be iterating over.


Stream.group_by() returns a new stream subscribing to the intermediate topic of the group by operation.

Apart from field descriptors, the key type argument can also be specified as a callable, or an async callable, so if you’re not using models to describe the data in streams you can manually extract the key used for repartitioning:

def get_order_account_id(order):
    return json.loads(order)['account_id']

async def process(orders):
    async for order in orders.group_by(get_order_account_id):

See also

items() – Iterate over keys and values

Use Stream.items() to get access to both message key and value at the same time:

async def process(stream):
    async for key, value in stream.items():

Note that this changes the type of what you iterate over from Stream to AsyncIterator, so if you want to repartition the stream or similar, .items() need to be the last operation:

async for key, value in stream.through('foo').group_by(

events() – Access raw messages

Use to iterate over raw Event values, including access to original message payload and message meta data:

async def process(stream):
    async for event in
        message = event.message
        topic = event.message.topic
        partition = event.message.partition
        offset = event.message.offset

        key_bytes = event.message.key
        value_bytes = event.message.value

        key_deserialized = event.key
        value_deserialized = event.value

        async with event:  # use  `async with event` for manual ack
            # event will be acked when this block returns.

See also

  • The faust.Event class in the API reference – for more information about events.

  • The faust.types.tuples.Message class in the API reference – for more information about the fields available in event.message.

take() – Buffer up values in the stream

Use Stream.take() to gather up multiple events in the stream before processing them, for example to take 100 values at a time:

async def process(stream):
    async for values in stream.take(100):
        assert len(values) == 100
        print(f'RECEIVED 100 VALUES: {values}')

The problem with the above code is that it will block forever if there are 99 messages and the last hundredth message is never received.

To solve this add a within timeout so that up to 100 values will be processed within 10 seconds:

async def process(stream):
    async for values in stream.take(100, within=10):
        print(f'RECEIVED {len(values)}: {values}')

The above code works better: if values are constantly being streamed it will process hundreds and hundreds without delay, but if there are long periods of time with no events received it will still process what it has gathered.

enumerate() – Count values

Use Stream.enumerate() to keep a count of the number of values seen so far in a stream.

This operation works exactly like the Python enumerate() function, but for an asynchronous stream:

async def process(stream):
    async for i, value in stream.enumerate():

The count will start at zero by default, but enumerate also accepts an optional starting point argument.

See also

  • The faust.utils.aiter.aenumerate() function – for a general version of enumerate() that let you enumerate any async iterator, not just streams.

  • The enumerate() function in the Python standard library.

through() – Forward through another topic

Use Stream.through() to forward every value to a new topic, and replace the stream by subscribing to the new topic:

source_topic = app.topic('source-topic')
destination_topic = app.topic('destination-topic')

async def process(stream):
    async for value in stream.through(destination_topic):
        # we are now iterating over stream(destination_topic)

You can also specify the destination topic as a string:

# [...]
async for value in stream.through('foo'):

Through is especially useful if you need to convert the number of partitions in a source topic, by using an intermediate table.

If you simply want to forward a value to another topic, you can send it manually, or use the echo recipe below:

async def process(stream):
    async for value in stream:
        await other_topic.send(value)

filter() – Filter values to omit from stream.

New in version 1.7.

This method is useful for filtering events before repartitioning a stream.

Takes a single argument which must be a callable, either a normal function or an async def function.


async def process(stream):
    async for value in stream.filter(lambda: v > 1000).group_by(...):

echo() – Repeat to one or more topics

Use echo() to repeat values received from a stream to another channel/topic, or many other channels/topics:

async def process(stream):
    async for event in stream.echo('other_topic'):

The operation takes one or more topics, as string topic names or app.topic, so this also works:

source_topic = app.topic('sourcetopic')
echo_topic1 = app.topic('source-copy-1')
echo_topic2 = app.topic('source-copy-2')

async def process(stream):
    async for event in stream.echo(echo_topic1, echo_topic2):

See also



Do not create Stream objects directly, instead use: to instantiate new streams.