[FEATURE] Apache Pulsar Connector

[tspannhw]

Is your feature request related to a problem? Please describe.
Add Pulsar as a Source. All the feature of Kafka. Pulsar also supports AMQP, MQTT, WebSockets, and Kafka as protocols.

Describe the solution you'd like
Add Pulsar the same as Kafka

Describe alternatives you've considered
We can use Pulsar's other protocols, but native is the fastest and easiest.

Additional context
Contact me for assistance. https://github.com/tspannhw/SpeakerProfile

[awmatheson]

👋 @tspannhw. Thanks for opening this! An apache pulsar connector would be great eventually. We will have some better examples on how to add connectors to bytewax in the future. For now, people getting started can use the Kafka protocol to take advantage of the KafkaInput, I made an example with this working here. However, it is important to note that some functionality will potentially not work out of the box, like recovery for instance

You can also use the Pulsar Python client to connect directly in the input or output code. For the input, state recovery and managing offsets would have to be managed in the input code.

[spacid]

Hello, we're thinking about using a Stream Processing Framework like Bytewax. However, we would need to connect to Apache Pulsar. Is there, by any chance, an update about the implementation of a native Apache Pulsar connector and/or documentation on how to implement (such) a connector?

[jhanninen]

If someone needs inspiration how to implement an own customised connector for Pulsar, here's our PoC:
https://github.com/HSLdevcom/ajoaikadata/blob/main/src/connectors/pulsar.py

The strategy I used there was to create Bytewax messages as dicts, like this:

class BytewaxMsg(TypedDict):
    data: dict                    # the actual message content
    msgs: NotRequired[list[str]]  # the list of Pulsar msg ids for acking

Disclaimer:

• This connector has not been used for a while, so it may be broken (last run with Bytewax 0.17, but code should be updated for 0.19).
• It has only been quickly tested in local environment, so the implementation is not guaranteed to handle all messages correctly nor optimised performance-wise.
• It contains some use-case-specific tricks, such as ability to filter messages or group multiple messages together and ack them once.