fastkafka 0.1.0rc1

FastKafka is a powerful and easy-to-use Python library for building asynchronous web services that interact with Kafka topics. Built on top of FastAPI, Starlette, Pydantic, AIOKafka and AsyncAPI, FastKafka simplifies the process of writing producers and consumers for Kafka topics.

FastKafka

Effortless Kafka integration for your web services

---

---

FastKafka is a powerful and easy-to-use Python library for building asynchronous services that interact with Kafka topics. Built on top of Pydantic, AIOKafka and AsyncAPI, FastKafka simplifies the process of writing producers and consumers for Kafka topics, handling all the parsing, networking, task scheduling and data generation automatically. With FastKafka, you can quickly prototype and develop high-performance Kafka-based services with minimal code, making it an ideal choice for developers looking to streamline their workflow and accelerate their projects.

Install

FastKafka works on macOS, Linux, and most Unix-style operating systems. You can install it with pip as usual:

pip install fastkafka

Writing server code

Here is an example python script using FastKafka that takes data from a Kafka topic, makes a prediction using a predictive model, and outputs the prediction to another Kafka topic.

Messages

FastKafka uses Pydantic to parse input JSON-encoded data into Python objects, making it easy to work with structured data in your Kafka-based applications. Pydantic’s BaseModel class allows you to define messages using a declarative syntax, making it easy to specify the fields and types of your messages.

This example defines two message classes for use in a FastKafka application:

• The InputData class is used to represent input data for a predictive model. It has three fields: user_id, feature_1, and feature_2. The user_id field is of type NonNegativeInt, which is a subclass of int that only allows non-negative integers. The feature_1 and feature_2 fields are both lists of floating-point numbers and integers, respectively.

• The Prediction class is used to represent the output of the predictive model. It has two fields: user_id and score. The score field is a floating-point number and it represents the prediction made by the model, such as the probability of churn in the next 28 days.

These message classes will be used to parse and validate incoming data in Kafka consumers and producers.

from typing import List
from pydantic import BaseModel, Field, NonNegativeInt

class InputData(BaseModel):
    user_id: NonNegativeInt = Field(..., example=202020, description="ID of a user")
    feature_1: List[float] = Field(
        ...,
        example=[1.2, 2.3, 4.5, 6.7, 0.1],
        description="input feature 1",
    )
    feature_2: List[int] = Field(
        ...,
        example=[2, 4, 3, 1, 0],
        description="input feature 2",
    )

class Prediction(BaseModel):
    user_id: NonNegativeInt = Field(..., example=202020, description="ID of a user")
    score: float = Field(
        ...,
        example=0.4321,
        description="Prediction score (e.g. the probability of churn in the next 28 days)",
        ge=0.0,
        le=1.0,
    )

These message classes will be used to parse and validate incoming data in a Kafka consumer and to produce a JSON-encoded message in a producer. Using Pydantic’s BaseModel in combination with FastKafka makes it easy to work with structured data in your Kafka-based applications.

Application

This example shows how to initialize a FastKafka application.

It starts by defining a dictionary called kafka_brokers, which contains two entries: "localhost" and "production", specifying local development and production Kafka brokers. Each entry specifies the URL, port, and other details of a Kafka broker. This dictionary is used for generating the documentation only and it is not being checked by the actual server.

Next, an object of the FastAPI class is created. It role is to serve the documentation and to start and shutdown FastKafka.

Finally, an object of the FastKafka class is initialized with the minimum set of arguments:

• app: an FastAPI application used for serving the documentation and starting/shutting down the service

• kafka_brokers: a dictionary used for generation of documentation

• bootstrap_servers: a host[:port] string or list of host[:port] strings that a consumer or a producer should contact to bootstrap initial cluster metadata

from os import environ

from fastkafka.application import FastKafka

kafka_brokers = {
    "localhost": {
        "url": "localhost",
        "description": "local development kafka broker",
        "port": 9092,
    },
    "production": {
        "url": "kafka.airt.ai",
        "description": "production kafka broker",
        "port": 9092,
        "protocol": "kafka-secure",
        "security": {"type": "plain"},
    },
}

bootstrap_servers = f"{environ['KAFKA_HOSTNAME']}:{environ['KAFKA_PORT']}"

kafka_app = FastKafka(
    kafka_brokers=kafka_brokers,
    bootstrap_servers=bootstrap_servers,
)

Function decorators

FastKafka provides convenient function decorators @kafka_app.consumes and @kafka_app.produces to allow you to delegate the actual process of

• consuming and producing data to Kafka, and

• decoding and encoding JSON encode messages

from user defined functions to the framework. The FastKafka framework delegates these jobs to AIOKafka and Pydantic libraries.

These decorators make it easy to specify the processing logic for your Kafka consumers and producers, allowing you to focus on the core business logic of your application without worrying about the underlying Kafka integration.

This following example shows how to use the @kafka_app.consumes and @kafka_app.produces decorators in a FastKafka application:

• The @kafka_app.consumes decorator is applied to the on_input_data function, which specifies that this function should be called whenever a message is received on the “input_data” Kafka topic. The on_input_data function takes a single argument which is expected to be an instance of the InputData message class. Specifying the type of the single argument is instructing the Pydantic to use InputData.parse_raw() on the consumed message before passing it to the user defined function on_input_data.

• The @produces decorator is applied to the to_predictions function, which specifies that this function should produce a message to the “predictions” Kafka topic whenever it is called. The to_predictions function takes two arguments: user_id and score. It creates a new Prediction message with these values and then returns it. The framework will call the Prediction.json().encode("utf-8") function on the returned value and produce it to the specified topic.

@kafka_app.consumes(topic="input_data", auto_offset_reset="latest", group_id="my_group")
async def on_input_data(msg: InputData):
    global model
    score = await model.predict(feature_1=msg.feature_1, feature_2=msg.feature_2)
    await to_predictions(user_id=msg.user_id, score=score)


@kafka_app.produces(topic="predictions")
async def to_predictions(user_id: int, score: float) -> Prediction:
    prediction = Prediction(user_id=user_id, score=score)
    return prediction


# this is a mock up for testing, should be replaced with the real model
class Model:
    async def predict(self, feature_1: List[int], feature_2: List[float]) -> float:
        return 0.87


model = Model()

Running the service

The service can be started using builtin faskafka run CLI command

We will concatenate the code snippets from above and save them in a file "server.py"

??? Example

This example contains the content of the file "server.py":

```python

from typing import List
from pydantic import BaseModel, Field, NonNegativeInt

class InputData(BaseModel):
    user_id: NonNegativeInt = Field(..., example=202020, description="ID of a user")
    feature_1: List[float] = Field(
        ...,
        example=[1.2, 2.3, 4.5, 6.7, 0.1],
        description="input feature 1",
    )
    feature_2: List[int] = Field(
        ...,
        example=[2, 4, 3, 1, 0],
        description="input feature 2",
    )

class Prediction(BaseModel):
    user_id: NonNegativeInt = Field(..., example=202020, description="ID of a user")
    score: float = Field(
        ...,
        example=0.4321,
        description="Prediction score (e.g. the probability of churn in the next 28 days)",
        ge=0.0,
        le=1.0,
    )



from os import environ

from fastkafka.application import FastKafka

kafka_brokers = {
    "localhost": {
        "url": "localhost",
        "description": "local development kafka broker",
        "port": 9092,
    },
    "production": {
        "url": "kafka.airt.ai",
        "description": "production kafka broker",
        "port": 9092,
        "protocol": "kafka-secure",
        "security": {"type": "plain"},
    },
}

bootstrap_servers = f"{environ['KAFKA_HOSTNAME']}:{environ['KAFKA_PORT']}"

kafka_app = FastKafka(
    kafka_brokers=kafka_brokers,
    bootstrap_servers=bootstrap_servers,
)



@kafka_app.consumes(topic="input_data", auto_offset_reset="latest", group_id="my_group")
async def on_input_data(msg: InputData):
    global model
    score = await model.predict(feature_1=msg.feature_1, feature_2=msg.feature_2)
    await to_predictions(user_id=msg.user_id, score=score)


@kafka_app.produces(topic="predictions")
async def to_predictions(user_id: int, score: float) -> Prediction:
    prediction = Prediction(user_id=user_id, score=score)
    return prediction


# this is a mock up for testing, should be replaced with the real model
class Model:
    async def predict(self, feature_1: List[int], feature_2: List[float]) -> float:
        return 0.87


model = Model()


```

Then, we start the FaskKafka servie by running the following command in the folder where the server.py file is located:

fastkafka run --num-workers=1 server:kafka_app

After running the command, you should see an output like the one below:

[5047]: [INFO] fastkafka.application: _create_producer() : created producer using the config: '{'bootstrap_servers': 'tvrtko-fastkafka-kafka-1:9092'}'
[5047]: [INFO] fastkafka._components.aiokafka_consumer_loop: aiokafka_consumer_loop() starting...
[5047]: [INFO] fastkafka._components.aiokafka_consumer_loop: aiokafka_consumer_loop(): Consumer created using the following parameters: {'bootstrap_servers': 'tvrtko-fastkafka-kafka-1:9092', 'auto_offset_reset': 'latest', 'max_poll_records': 100, 'group_id': 'my_group'}
[5047]: [INFO] fastkafka._components.aiokafka_consumer_loop: aiokafka_consumer_loop(): Consumer started.
[5047]: [INFO] aiokafka.consumer.subscription_state: Updating subscribed topics to: frozenset({'input_data'})
[5047]: [INFO] aiokafka.consumer.consumer: Subscribed to topic(s): {'input_data'}
[5047]: [INFO] fastkafka._components.aiokafka_consumer_loop: aiokafka_consumer_loop(): Consumer subscribed.
[5047]: [INFO] aiokafka.consumer.group_coordinator: Discovered coordinator 1001 for group my_group
[5047]: [INFO] aiokafka.consumer.group_coordinator: Revoking previously assigned partitions set() for group my_group
[5047]: [INFO] aiokafka.consumer.group_coordinator: (Re-)joining group my_group
[5047]: [INFO] aiokafka.consumer.group_coordinator: Joined group 'my_group' (generation 3) with member_id aiokafka-0.8.0-4638ef56-bd7d-4581-95ad-42c90322442d
[5047]: [INFO] aiokafka.consumer.group_coordinator: Elected group leader -- performing partition assignments using roundrobin
[5047]: [INFO] aiokafka.consumer.group_coordinator: Successfully synced group my_group with generation 3
[5047]: [INFO] aiokafka.consumer.group_coordinator: Setting newly assigned partitions {TopicPartition(topic='input_data', partition=0)} for group my_group
Starting process cleanup, this may take a few seconds...
[INFO] fastkafka.server: terminate_asyncio_process(): Terminating the process 5047...
[5047]: [INFO] aiokafka.consumer.group_coordinator: LeaveGroup request succeeded
[5047]: [INFO] fastkafka._components.aiokafka_consumer_loop: aiokafka_consumer_loop(): Consumer stopped.
[5047]: [INFO] fastkafka._components.aiokafka_consumer_loop: aiokafka_consumer_loop() finished.
[INFO] fastkafka.server: terminate_asyncio_process(): Process 5047 terminated.

When the service is started, several log messages are printed to the console, including information about the application startup, AsyncAPI specification generation, and consumer loop status.

During the lifetime of the service, incoming requests will be processed by the FastKafka application and appropriate actions will be taken based on the defined Kafka consumers and producers. For example, if a message is received on the “input_data” Kafka topic, the on_input_data function will be called to process the message, and if the to_predictions function is called, it will produce a message to the “predictions” Kafka topic. The service will continue to run until it is shut down, at which point the application shutdown process will be initiated and the service will stop.

Checking out the documentation

To generate and serve the documentation locally, you can use the built in kafka function that will do all the work for you. In the folder where the server.py file is located, run the following command:

fastkafka docs serve server:kafka_app

After running the command you should see the following output:

exit_code, output = await run_script_and_cancel(
    script=script,
    script_file="server.py",
    cmd=cmd,
    cancel_after=45,
)
print(output.decode("utf-8"))
assert exit_code == 0, exit_code
print(output.decode("utf-8"))

{'localhost': {'description': 'local development kafka broker',
               'port': 9092,
               'url': 'localhost'},
 'production': {'description': 'production kafka broker',
                'port': 9092,
                'protocol': 'kafka-secure',
                'security': {'type': 'plain'},
                'url': 'kafka.airt.ai'}}

The generated documentation is as follows:

Next, you can see the documentation generated from the @consumes decorator when used on the function on_input_data with a single parameter of type InputData:

class InputData(BaseModel):
    user_id: NonNegativeInt = Field(..., example=202020, description="ID of a user")
    feature_1: List[float] = Field(
        ...,
        example=[1.2, 2.3, 4.5, 6.7, 0.1],
        description="input feature 1",
    )
    feature_2: List[int] = Field(
        ...,
        example=[2, 4, 3, 1, 0],
        description="input feature 2",
    )

@kafka_app.consumes(topic="input_data", auto_offset_reset="latest", group_id="my_group")
async def on_input_data(msg: InputData):
    global model
    score = await model.predict(feature_1=msg.feature_1, feature_2=msg.feature_2)
    await to_predictions(user_id=msg.user_id, score=score)

The resulting documentation is generated as follows:

Testing the service

from os import environ

import anyio
import asyncer
from tqdm.notebook import tqdm, trange

from fastkafka.helpers import (
    consumes_messages,
    create_missing_topics,
    create_admin_client,
    produce_messages,
    wait_for_get_url,
)

from fastkafka.testing import LocalKafkaBroker

num_workers = 4
listener_port = 9913
msgs = [
    dict(user_id=i, feature_1=[(i / 1_000) ** 2], feature_2=[i % 177])
    for i in trange(5_000, desc="generating messages")
]

bootstrap_servers = f"{environ['KAFKA_HOSTNAME']}:{environ['KAFKA_PORT']}"

async with LocalKafkaBroker(
    zookeeper_port=9911, listener_port=listener_port
) as bootstrap_servers:
    create_missing_topics(
        ["input_data", "predictions"],
        bootstrap_servers=bootstrap_servers,
        num_partitions=4,
    )

    async with asyncer.create_task_group() as tg:
        output = tg.soonify(run_script_and_cancel)(
            script=script.replace(
                "bootstrap_servers = f\"{environ['KAFKA_HOSTNAME']}:{environ['KAFKA_PORT']}\"",
                f'bootstrap_servers = "localhost:{listener_port}"',
            ),
            script_file=f"server.py",
            cmd=f"fastkafka run --num-workers={num_workers} server:kafka_app",
            cancel_after=120,
        )

        tg.soonify(consumes_messages)(
            msgs_count=len(msgs),
            topic="predictions",
            bootstrap_servers=bootstrap_servers,
        )

        await anyio.sleep(10)

        tg.soonify(produce_messages)(
            msgs=msgs, topic="input_data", bootstrap_servers=bootstrap_servers
        )

print(output.value[1].decode("UTF-8"))

generating messages:   0%|          | 0/20000 [00:00<?, ?it/s]

[INFO] fastkafka.testing: Java is already installed.
[INFO] fastkafka.testing: But not exported to PATH, exporting...
[INFO] fastkafka.testing: Kafka is already installed.
[INFO] fastkafka.testing: But not exported to PATH, exporting...
[INFO] fastkafka.helpers: create_missing_topics(['input_data', 'predictions']): new_topics = [NewTopic(topic=predictions,num_partitions=4)]
[INFO] aiokafka.consumer.subscription_state: Updating subscribed topics to: frozenset({'predictions'})
[INFO] aiokafka.consumer.group_coordinator: Metadata for topic has changed from {} to {'predictions': 4}. 

consuming from 'predictions':   0%|          | 0/20000 [00:00<?, ?it/s]

producing to 'input_data':   0%|          | 0/20000 [00:00<?, ?it/s]

[INFO] fastkafka.components._subprocess: terminate_asyncio_process(): Terminating the process 5435...
[INFO] fastkafka.components._subprocess: terminate_asyncio_process(): Process 5435 was already terminated.
[INFO] fastkafka.components._subprocess: terminate_asyncio_process(): Terminating the process 5052...
[INFO] fastkafka.components._subprocess: terminate_asyncio_process(): Process 5052 was already terminated.

PendingValueException: The return value of this task is still pending. Maybe you forgot to access it after the async with asyncer.create_task_group() block. If you need to access values of async tasks inside the same task group, you probably need a different approach, for example with AnyIO Streams.