I am working on something adjacent to this problem. We focus much less on data pipelines but on automation, but in the end also have an abstraction for flows that one can use to build data pipeline. The locking-in issue was something I thought a lot about and ended up deciding that our generic steps should just be plain code in typescript/python/go/bash, the only requirement is that those snippets code have a main function and return a result. We built the https://hub.windmill.dev where users can share their scripts directly and we have a team of moderators to approve the one to integrate directly into the main product. The goal with those snippets is that they are generic enough to be reusable outside of Windmill and they might be able to work straight out of the box for orchest for the python ones.
I’ve been leaning towards this direction. I think I/O is the biggest part that in the case of plain code steps still needs fixing. Input being data/stream and parameterization/config and output being some sort of typed data/stream.
At the same time, text feels horribly inefficient as encoding for some of the data objects being passed around in these flows. More specialized and optimized binary formats come to mind (Arrow, HDF5, Protobuf).
Plenty of directions to explore, each with their own advantages and disadvantages. I wonder which direction is favored by users of tools like ours. Will be good to poll (do they even care?).
Yes, inputs/outputs is likely the most interesting problems for our diverse specs of flows.
Because data pipeline is not the primary concerns of Windmill, we took the stance that Inputs and Output of steps were simply JSON in, JSON out. For all the languages, we simply extract the JSON object into the different parameters of the main, and then we wrap the return into the respective language native serializer for the output (e.g JSON.stringify in Typescript). Then each step can use a javascript expression executed by v8 to do some lightweight transformation between the output of any step to the input of that step.
A lot of the simplification we made is actually parsing the main function parameters into the corresponding jsonschema, supporting deeply nested objects when relevant.
That works great for automation that do not have big input/outputs, but not for data. So what we do for data is to use a folder that we symlink to be shared by all steps if a specific flag for that flow is set. It also force us to have the same worker process all the steps inside that flow when otherwise flow steps could have been processed by any workers. It is very fast since it's all local filesystem but not super scalable.
I am not pleased with that solution and believe that if we were to expand on the data problem, we would certainly rely on fast network and HDFS/Amazon EFS/etc to simply share that mounted folder across the network.
Anyway, sorry for the rambling but I do feel like we're all taking different approach to the same underlying problem of building the best abstraction for flows and believe we might learn from each other's choices.
ps: congrats Patterns on the launch, the tool look absolutely amazing.
nb: author of https://github.com/windmill-labs/windmill