kafkaesk

This project is meant to help facilitate easily publishing and subscribing to events with python and Kafka.

Guiding principal:

• simple http, language agnostic contracts built on top of kafka.

Alternatives:

• pure aiokafka: can be complex to scale correctly
• guillotina_kafka: complex, tied to guillotina
• faust: requires additional data layers, not language agnostic
• confluent kafka + avro: close but ends up being like grpc. compilation for languages. No asyncio.

(consider this python project as syntatic sugar around these ideas)

Publish

using pydantic but can be done with pure json

import kafkaesk
from pydantic import BaseModel

app = kafkaesk.Application()

@app.schema("Content", version=1, retention=24 * 60 * 60)
class ContentMessage(BaseModel):
    foo: str


async def foobar():
    # ...
    # doing something in an async func
    await app.publish("content.edited.Resource", data=ContentMessage(foo="bar"))

Subscribe

import kafkaesk
from pydantic import BaseModel

app = kafkaesk.Application()

@app.schema("Content", version=1, retention=24 * 60 * 60)
class ContentMessage(BaseModel):
    foo: str


@app.subscribe('content.*')
async def get_messages(data: ContentMessage):
    print(f"{data.foo}")

Optional consumer injected parameters:

• schema: str
• record: aiokafka.structs.ConsumerRecord

Depending on the type annotation for the first parameter, you will get different data injected:

• async def get_messages(data: ContentMessage): parses pydantic schema
• async def get_messages(data: bytes): give raw byte data
• async def get_messages(record: aiokafka.structs.ConsumerRecord): give kafka record object
• async def get_messages(data): raw json data in message

kafkaesk contract

This is just a library around using kafka. Kafka itself does not enforce these concepts.

• every message must provide a json schema
• messages produced will be validated against json schema
• each topic will have only one schema
• a single schema can be used for multiple topics
• consumed message schema validation is up to the consumer
• messages will be consumed at least once. Considering this, your handling should be idempotent

message format

{
    "schema": "schema_name:1",
    "data": { ... }
}

Worker

kafkaesk mymodule:app --kafka-servers=localhost:9092

Options:

• --kafka-servers: comma separated list of kafka servers
• --kafka-settings: json encoded options to be passed to https://aiokafka.readthedocs.io/en/stable/api.html#aiokafkaconsumer-class
• --topic-prefix: prefix to use for topics
• --replication-factor: what replication factor topics should be created with. Defaults to min(number of servers, 3).

Application.publish

• stream_id: str: name of stream to send data to
• data: class that inherits from pydantic.BaseModel
• key: Optional[bytes]: key for message if it needs one

Application.subscribe

• stream_id: str: fnmatch pattern of streams to subscribe to
• group: Optional[str]: consumer group id to use. Will use name of function if not provided

Application.schema

• id: str: id of the schema to store
• version: Optional[int]: version of schema to store
• streams: Optional[List[str]]: if streams are known ahead of time, we can pre-create them before we push data
• retention: Optional[int]: retention policy in seconds

Application.configure

• kafka_servers: Optional[List[str]]: kafka servers to connect to
• topic_prefix: Optional[str]: topic name prefix to subscribe to
• kafka_settings: Optional[Dict[str, Any]]: additional aiokafka settings to pass in
• replication_factor: Optional[int]: what replication factor topics should be created with. Defaults to min(number of servers, 3).

Dev

poetry install

Run tests:

docker-compose up
KAFKA=localhost:9092 poetry run pytest tests

Extensions

Logging

This extension includes classes to extend python's logging framework to publish structured log messages to a kafka topic. This extension is made up of three main components: an extended logging.LogRecord, a custom logging.Formatter, and a some custom logging.Handlers.

See logger.py in examples directory.

Log Record

kafkaesk.ext.logging.record.factory is a function that will return kafkaesk.ext.logging.record.PydanticLogRecord objects. The factory() function scans through any args passed to a logger and checks each item to determine if it is a subclass of pydantid.BaseModel. If it is a base model instance and model._is_log_model evaluates to True the model will be removed from args and added to record._pydantic_data. After that factory() will use logging's existing logic to finish creating the log record.

The factory() function is automatically passed to logging.setLogRecordFactory when a kafkaesk.ext.logging.formatter.PydanticFormatter instance is created.

Formatter

kafkaesk.ext.logging.formatter.PydanticFormatter is responsible for parsing all information in a log record and returning a kafkaesk.logging.formatter.PydanticLogModel instance. The formatter will build the return message in two passes. The first pass looks at formatter.format_class and attempts to create an instance of that class utilizing the log record's __dict__. The second pass creates a standard dictionary from all models present in a log record's _pydantic_data property. These results are then merged together and used to create a new PydanticLogModel instance which is returned to the handler.

Handler

This extensions ships with two handlers capable of handling kafkaesk.ext.logging.formatter.PydanticLodModel classes: kafakesk.ext.logging.handler.PydanticStreamHandler and kafkaesk.ext.logging.handler.PydanticKafkaeskHandler.

The stream handler is a very small wrapper around logging.StreamHandler, the signature is the same, the only difference is that the handler will attempt to convert any pydantic models it receives to a json string.

The kafkaesk handler has a few more bits going on in the background. The handler has two required inputs, a kafkaesk.app.Application instance and a stream name. Once initialized any logs emitted by the handler will be saved into an internal queue. There is a worker task that handles pulling logs from the queue and writing those logs to the specified topic.

Features

(or a todo here)

• worker
• service to inspect stats
• be able to handle manual commit use-case
• be able to reject commit/abort message handling
• prometheus

Naming things

It's hard and "kafka" is already a fun name. Hopefully this library isn't literally "kafkaesque" for you.