Show HN: DuckDB for Kafka Stream Processing

Posted by dm03514 1 day ago

Counter74Comment13OpenOriginal

Hello Everyone! We built SQLFlow as a lightweight stream processing engine.

We leverage DuckDB as the stream processing engine, which gives SQLFlow the ability to process 10's of thousands of messages a second using ~250MiB of memory!

DuckDB also supports a rich ecosystem of sinks and connectors!

https://sql-flow.com/docs/category/tutorials/

https://github.com/turbolytics/sql-flow

We were tired of running JVM's for simple stream processing, and also of bespoke one off stream processors

I would love your feedback, criticisms and/or experiences!

Thank you

Comments

Comment by pulkitsh1234 1 day ago

(not an expert in stream processing).. from the docs here https://sql-flow.com/docs/introduction/basics#output-sink it seems like this works on "batches" of data, how is this different from batch processing ? Where is the "stream" here ?

Comment by dm03514 1 day ago

Ha Yes! A pipeline assumes a "batch" of data, which is backed by an ephemeral duckdb in memory table. The goal is to provide SQL table semantics and implement pipelines in a way where the batch size can be toggled without a change to the pipeline logic.

The stream is achieved by the continuous flow of data from Kafka.

SQLFlow exposes a variable for batch size. Setting the batch size to 1 will make it so SQLFlow reads a kafka message, applies the processor SQL logic and then ensures it successfully commits the SQL results to the sink, one after another.

SQLFlow provides at least once delivery guarantees. It will only commit the source message once it successfully writes to the pipeline output (sink).

https://sql-flow.com/docs/operations/handling-errors

The batch table is just a convention which allows for seamless batch size configuration. If your throughput is low, or if you require message by message processing, SQLFlow can be toggled to a batch of 1. If you need higher throughput and can tolerate the latency, then the batch can be toggled higher.

Comment by srameshc 1 day ago

This looks brilliant, thank you. I love DuckDB and use it for lot of local data processing jobs. We have a data stream, not to the size where we need to push to BigQuery or elsewhere. I was thinking of trying something like sql-flow but I am glad now it makes the job very easy.

Comment by mihevc 1 day ago

How does this compare to https://github.com/Query-farm/tributary ?

Comment by rustyconover 1 day ago

The next major release of Tributary will support Avro, Protobuf and JSON along with the Schema Registry it will also bring the ability to write to Kafka with transactions.

But really you should get excited for DuckDB Labs to build out materialized views. Materialized views where you can ingest more streaming data to update aggregates. This way you could just keep pushing rows through aggregates from Kafka.

It is going to be a POWER HOUSE for streaming analytics.

Contact DuckDB Labs if you want to sponsor the work on materialized views: https://duckdb.org/roadmap

Comment by trueno 1 day ago

Is this to be used in an analytics application backend sort of scenario?

I am familiar with materialized views / dynamic tables from enterprise-grade cloud lake type offerings, but I've never quite understood where duckdb, though impressive, fits into everyones use case. I've toyed with it for personal things, it's very cool having a local instance of something akin to snowflake when it comes to processing and aggregating on Big Data™ but generally I don't see it used in operational settings. For application development people are generally tied to sqlite and postgres.

It all does seem really cool though, I guess I'm just not feeling creative enough to conjure up a stream-to-duckdb use case. Feel free to bombard me with cool ideas.

Comment by buremba 1 day ago

Exactly. I have also been playing with DuckDB for streaming use cases, but it feels hacky to issue micro-batching queries on streaming data in short intervals.

DuckDB has everything that streaming engines such as Flink have; it just needs to support managing intermediate aggregate states and scheduling the materialized views itself.

Comment by dm03514 1 day ago

Oh yes!! I've seen this a couple times. I am far from an expert in tributary so please take with a grain of salt.

Based on the tributary documentation, I understand that tributary embeds kafka consumers into duckdb. This makes duckdb the main process that you run to perform consumption. I think that this makes creating stream processing POCs very accessible. It looks like it is quite easy to start streaming data into duckdb. What I don't see is a full story around Devops, operations, testing, configuration as code etc.

SQLFlow is a service that embeds DuckDB as the storage and processing brains. Because of this, we're able to offer metrics, testing utilities, pipelines as code, and all the other DevOps utilities that are necessary to run a huge number of streaming instances 24x7. SQLFlow was created as a tool that I wish I had to for simple stream processing in production in high availability contexts :)

Comment by mihevc 1 day ago

Nice! Thanks for the context, it's great to know!

Comment by mbay 1 day ago

I see an example with what looks like a lookup-type join against a Postgres DB. Are stream/stream joins supported, though?

The DLQ and Prometheus integration out of the box are nice.

Comment by dm03514 1 day ago

Stream to stream joins are NOT currently supported. This is a regularly requested feature, and I'll look at prioritizing it.

SQLFlow uses duckdb internally for windowing and stream state storage :), and I'll look at extending it to support stream / stream joins.

Could you describe a bit more about your use case? I'd really appreciate it if you could create an issue in the repo describing your use case and desired functionality a bit!

https://github.com/turbolytics/sql-flow/issues

We were looking at solving some of the simplier use cases first before branching out into these more complicated ones :)

Comment by mbay 1 day ago

I worked on stream processing at my previous gig but don't have a need for it currently. Just curious.

Comment by itsfseven 1 day ago

It would be great if this supported Pulsar too!