. _guide-agents:

Agents - Self-organizing Stream Processors

What is an Agent?

An agent is a distributed system processing the events in a stream.

Every event is a message in the stream and is structured as a key/value pair that can be described using models for type safety and straightforward serialization support.

Streams can be sharded in a round-robin manner, or partitioned by the message key; this decides how the stream divides between available agent instances in the cluster.

Create an agent

To create an agent, you need to use the @app.agent decorator on an async function taking a stream as the argument. Further, it must iterate over the stream using the async for keyword to process the stream:

# faustexample.py

import faust

app = faust.App('example', broker='kafka://localhost:9092')

async def myagent(stream):
    async for event in stream:
        ...  # process event
Start a worker for the agent

The faust worker program can be used to start a worker from the same directory as the faustexample.py file:

$ faust -A faustexample worker -l info

Every new worker that you start will force the cluster to rebalance partitions so that every agent receives a specific portion of the stream.


When an agent reads from a topic, the stream is partitioned based on the key of the message. For example, the stream could have keys that are account ids, and values that are high scores, then partitioning decide that any message with the same account id as key, is delivered to the same agent instance.

Sometimes you’ll have to repartition the stream, to ensure you are receiving the right portion of the data. See Streams - Infinite Data Structures for more information on the Stream.group_by() method.


If you don’t set a key (i.e. key=None), the messages will be delivered to available workers in round-robin order. This is useful to simply distribute work amongst a cluster of workers, and you can always repartition that stream later should you need to access data in a table or similar.

Fault tolerancy

If the worker for a partition fails, or is blocked from the network for some reason, there is no need to worry because Kafka solves this by moving the partition to a worker that works.

Faust also takes advantage of “standby tables” and a custom partition manager that prefes to promote any node with a full copy of the data, saving startup time and ensuring availability.

Here’s a complete example of an app, having an agent that adds numbers:

# examples/agent.py
import faust

# The model describes the data sent to our agent,
# We will use a JSON serialized dictionary
# with two integer fields: a, and b.
class Add(faust.Record):
    a: int
    b: int

# Next, we create the Faust application object that
# configures our environment.
app = faust.App('agent-example')

# The Kafka topic used by our agent is named 'adding',
# and we specify that the values in this topic are of the Add model.
# (you can also specify the key_type if your topic uses keys).
topic = app.topic('adding', value_type=Add)

async def adding(stream):
    async for value in stream:
        # here we receive Add objects, add a + b.
        yield value.a + value.b

Starting a worker will now start a single instance of this agent:

$ faust -A examples.agent worker -l info

To send values to it, open a second console to run this program:

# examples/send_to_agent.py
import asyncio
from .agent import Add, adding

async def send_value() -> None:
    print(await adding.ask(Add(a=4, b=4)))

if __name__ == '__main__':
    loop = asyncio.get_event_loop()
$ python examples/send_to_agent.py

Define commands with the @app.command decorator.

You can also use CLI Commands to add actions for your application on the command line. Use the @app.command decorator to rewrite the example program above (examples/agent.py), like this:

async def send_value() -> None:
    print(await adding.ask(Add(a=4, b=4)))

After adding this to your examples/agent.py module, run your new command using the faust program:

$ faust -A examples.agent send_value

You may also specify command line arguments and options:

from faust.cli import argument, option

    argument('a', type=int, help='First number to add'),
    argument('b', type=int, help='Second number to add'),
    option('--print/--no-print', help='Enable debug output'),
async def send_value(a: int, b: int, print: bool) -> None:
    if print:
        print(f'Sending Add({x}, {y})...')
    print(await adding.ask(Add(a, b)))

Then pass those arguments on the command line:

$ faust -A examples.agent send_value 4 8 --print
Sending Add(4, 8)...

The Agent.ask() method wraps the value sent in a particular structure that includes the return address (reply-to). When the agent sees this type of arrangement, it will reply with the result yielded by the agent as a result of processing the event.

Static types

Faust is typed using the type annotations available in Python 3.6, and can be checked using the mypy type checker.

Add type hints to your agent function like this:

from typing import AsyncIterable
from faust import StreamT

async def adding(stream: StreamT[Add]) -> AsyncIterable[int]:
    async for value in stream:
        yield value.a + value.b

The StreamT type used for the agent’s stream argument is a subclass of AsyncIterable extended with the stream API. You could type this call using AsyncIterable, but then mypy would stop you with a typing error should you use stream-specific methods such as .group_by(), through(), etc.

Under the Hood: The @agent decorator

You can quickly start a stream processor in Faust without using agents. Do so merely by launching an asyncio task that iterates over a stream:

# examples/noagents.py
import asyncio

app = faust.App('noagents')
topic = app.topic('noagents')

async def mystream():
    async for event in topic.stream():
        print(f'Received: {event!r}')

async def start_streams():
    await app.start()
    await mystream()

if __name__ == '__main__':
    loop = asyncio.get_event_loop()

Essentially what the @agent decorator does, given a function like this:

async def mystream(stream):
    async for event in stream:
        print(f'Received: {event!r}')
        yield event

It wraps your function returning async iterator (since it uses yield) in code similar to this:

def agent(topic):

    def create_agent_from(fun):
        async def _start_agent():
            stream = topic.stream()
            async for result in fun(stream):

Defining Agents

The Topic

The topic argument to the agent decorator defines the main topic that agent reads from (this implies it’s not necessarily the only topic, as is the case when using stream joins, for example).

Topics are defined using the app.topic() helper and return a faust.Topic description:

topic = app.topic('topic_name1', 'topic_name2',

Should the topic description provide multiple topic names, the main topic of the agent will be the first topic in that list ("topic_name1").

The key_type and value_type describe how to serialize and deserialize messages in the topic, and you provide it as a model (such as faust.Record), a faust.Codec, or the name of a serializer.

If not specified it will use the default serializer defined by the app.


If you don’t specify a topic, the agent will use the agent name as the topic: the name will be the fully qualified name of the agent function (e.g., examples.agent.adder).

See also

The Stream

The decorated function is unary, meaning it must accept a single argument.

The object passed in as the argument to the agent is an async iterable Stream instance, created from the topic/channel provided to the decorator:

async def myagent(stream):
    async for item in stream:

Iterating over this stream, using the async for keyword, will in turn iterate over messages in the topic/channel.

You can also use the group_by() method of the Stream API, to partition the stream differently:

# examples/groupby.py
import faust

class BankTransfer(faust.Record):
    account_id: str
    amount: float

app = faust.App('groupby')
topic = app.topic('groupby', value_type=BankTransfer)

async def stream(s):
    async for transfer in s.group_by(BankTransfer.account_id):
        # transfers will now be distributed such that transfers
        # with the same account_id always arrives to the same agent
        # instance

A two-way join works by waiting until it has a message from both topics, so to synchronously wait for a reply from the agent you would have to send messages to both topics. A three-way join means you have to send a message to each of the three topics and only then can a reply be produced.

For this reason, you’re discouraged from using joins in an agent, unless you know what you’re doing:

topic1 = app.topic('foo1')
topic2 = app.topic('foo2')

async def mystream(stream):
    # XXX This is not proper use of an agent, as it performs a join.
    # It works fine as long as you don't expect to be able to use
    # ``agent.ask``, ``agent.map`` and similar
    # methods that wait for a reply.
    async for event in (stream & topic2.stream()).join(...):

For joins, the best practice is to use the @app.task decorator instead, to launch an asyncio.Task when the app starts, that manually iterates over the joined stream:

def mystream():
    async for event in (topic1.stream() & topic2.stream()).join(...):
        # process merged event

See also


Use the concurrency argument to start multiple instances of an agent on every worker instance. Each agent instance (actor) will process items in the stream concurrently (and in no particular order).


Concurrent instances of an agent will process the stream out-of-order, so you aren’t allowed to mutate tables from within the agent function:

An agent having concurrency > 1, can only read from a table, never write.

Here’s an agent example that can safely process the stream out of order.

Our hypothetical backend system publishes a message to the Kafka “news” topic every time a news article is published by an author.

We define an agent that consumes from this topic and for every new article will retrieve the full article over HTTP, then store that in a database somewhere (yeah, pretty contrived):

class Article(faust.Record, isodates=True):
    url: str
    date_published: datetime

news_topic = app.topic('news', value_type=Article)

@app.agent(news_topic, concurrency=10)
async def imports_news(articles):
    async for article in articles:
        response = await aiohttp.ClientSession().get(article.url)
        await store_article_in_db(response)


Sinks can be used to perform additional actions after the agent has processed an event in the stream, such as forwarding alerts to a monitoring system, logging to Slack, etc. A sink can be callable, async callable, a topic/channel or another agent.

Function Callback

Regular functions take a single argument (the value yielded by the agent):

def mysink(value):
    print(f'AGENT YIELD: {value!r}')

async def myagent(stream):
    async for value in stream:
        yield value * 2
Async Function Callback

If you provide an async function, the agent will await it:

async def mysink(value):
    print(f'AGENT YIELD: {value!r}')
    # OBS This will force the agent instance that yielded this value
    # to sleep for 1.0 second before continuing on the next event
    # in the stream.
    await asyncio.sleep(1)

async def myagent(stream):
    async for value in stream:
        yield value * 2

Specifying a topic as the sink will force the agent to forward yielded values it:

agent_log_topic = app.topic('agent_log')

async def myagent(stream):
    async for value in stream:
        yield value
Another Agent

Specifying another agent as the sink will force the agent to forward yielded values to it:

async def agent_b(stream):
    async for value in stream:
        print(f'AGENT B RECEIVED: {event!r}')

async def agent_a(stream):
    async for value in stream:
        print(f'AGENT A RECEIVED: {event!r}')
        yield value * 2

When agents raise an error

If an agent raises in the middle of processing an event what do we do with acking it?

Currently the source message will be acked and not processed again, simply because it violates “”exactly-once” semantics”.

  • What about retries?

    It’d be safe to retry processing the event if the agent processing is idempotent, but we don’t enforce idempotency in stream processors so it’s not something we can enable by default.

    The retry would also have to stop processing of the topic so that order is maintained: the next offset in the topic can only be processed after the event is retried.

  • How about crashing?

    Crashing the instance to require human intervention is certainly a choice, but far from ideal considering how common mistakes in code or unhandled exceptions are. It may be better to log the error and have ops replay and reprocess the stream on notification.

Using Agents

Cast or Ask?

When communicating with an agent, you can ask for the result of the request to be forwarded to another topic: this is the reply_to topic.

The reply_to topic may be the topic of another agent, a source topic populated by a different system, or it may be a local ephemeral topic collecting replies to the current process.

If you perform a cast, you’re passively sending something to the agent, and it will not reply back.

Systems perform better when no synchronization is required, so you should try to solve your problems in a streaming manner. If B needs to happen after A, try to have A call B instead (which could be accomplished using reply_to=B).

cast(value, *, key=None, partition=None)

A cast is non-blocking as it will not wait for a reply:

await adder.cast(Add(a=2, b=2))

The agent will receive the request, but it will not send a reply.

ask(value, *, key=None, partition=None, reply_to=None, correlation_id=None)

Asking an agent will send a reply back to process that sent the request:

value = await adder.ask(Add(a=2, b=2))
assert value == 4
send(key, value, partition, reply_to=None, correlation_id=None)

The Agent.send method is the underlying mechanism used by cast and ask.

Use it to send the reply to another agent:

await adder.send(value=Add(a=2, b=2), reply_to=another_agent)

Streaming Map/Reduce

These map/reduce operations are shortcuts used to stream lots of values into agents while at the same time gathering the results.

map streams results as they come in (out-of-order), and join waits until all the steps are complete (back-to-order) and return the results in a list with orering preserved:

map(values: Union[AsyncIterable[V], Iterable[V]])

Map takes an async iterable, or a regular iterable, and returns an async iterator yielding results as they come in:

async for reply in agent.map([1, 2, 3, 4, 5, 6, 7, 8]):
    print(f'RECEIVED REPLY: {reply!r}')

The iterator will start before all the messages have been sent, and should be efficient even for infinite lists.

As the map executes concurrently, the replies will not appear in any particular order.

kvmap(items: Union[AsyncIterable[Tuple[K, V], Iterable[Tuple[K, V]]]])
Same as map, but takes an async iterable/iterable of (key, value) tuples, where the key in each pair is used as the Kafka message key.
join(values: Union[AsyncIterable[V], Iterable[V]])

Join works like map but will wait until all of the values have been processed and returns them as a list in the original order.

The await will continue only after the map sequence is over, and all results are accounted for, so do not attempt to use join together with infinite data structures ;-)

results = await pow2.join([1, 2, 3, 4, 5, 6, 7, 8])
assert results == [1, 4, 9, 16, 25, 36, 49, 64]
kvjoin(items: Union[AsyncIterable[Tuple[K, V]], Iterable[Tuple[K, V]]])
Same as join, but takes an async iterable/iterable of (key, value) tuples, where the key in each pair is used as the message key.