(I have used all of the above tools in my 15+ yr career. Code as ETL was a huge industry shift)
I think fundamentally, the problem is that any DSL lives on a design tradeoff spectrum between domain-specificity and programming-generality. There is only so much savings that a generic ETL tool will give you, because it doesn't want to be domain specific. On the other hand, the more domain specific a tool is, the more limited it will be in it's capacity to handle other domains.
No amount of graphic or textual interface is gonna save you from this fundamental design tradeoff (it's kind of complexity tradeoff - things inside domain are easy but outside the domain are hard). But then, you can as well handle this tradeoff in a decent library/framework for a general programming language, which is, effectively, a "return to code".
Other way to look at it is that we want limited programming languages, because the implicit assumptions make things easier for the reader of the code (the person who wants to understand). However, limitations make things harder for the writer of the code (the person who wants to solve some problem).
In practice you lose all the benefits of abstraction, unit testing, proper CI/CD, etc. I haven't met an engineer that likes the service. Some projects have resorted to writing code generation tools, so that they can take config files and programmatically generate the JSON serialization of the pipelines that you're supposed to develop by clicking and dragging.
This graph-based visual programming somehow fails to deliver on speed of development. Mouse has 2 buttons, the keyboard approx. 100. Not to mention that LLMs work on the language/code level, and are expected to stay so for a while. We dont' have universal means to express things visually. Except for the graph notation of edgex/virtices. But then there is no universal knowledge, people dont usually disambiguate between sequence diagram, bpmn and state diagram. these are all graphs, right, but not the same semantically.
I'd rather go for a standardized ETL langauge a-la-markdown, and only then get to appreciate the GUI.
this is reminiscent of https://mikehadlow.blogspot.com/2012/05/configuration-comple...
Why? We are pretty deep into the ecosystem.
Abstraction -> the only thing data factory does not allow you is to reference a previous activity as a variable, which makes sense if you don't want to let your customer blow up your product. Parametrize all you want.
Unit testing -> test all you want each activity, pipeline, flow, resume it from where it broke. Clone the entire thing into a test data factory, then deploy that once ready.
CI/CD -> the first step it nags you about is setting up CI/CD. If you want to get fancy, you setup a dev environment and deploy that to production after testing and sign-off.
Abstracting ETL only works when you remember or have the same people on staff that abstracted that ETL process. Data factory 'could' be visual but does not let you pull the same level of non-sense that SSIS would.
For example, we call data factory via API, the pipeline is fully abstracted, it does one thing, but it's inputs and outputs are controlled by the request.
ETL is hot again with machine learning. Companies have massive amounts of data they need to get in shape for models. The promise of a GUI anyone can use with little experience (low costs) means these tools are gaining traction again for the same lessons to be learnt from the last generation who moved on/out the industry.
My personal pain is Tableau trying to move uo the stack into data orchestration and compute. Code is completely inaccessible.
Not really a Tableau user, but I bet they are doing it to increase vendor lock-in. Which has always been an industry wide problem, proliferation of APIs and languages for that reason. I am not even sure how this could be any different, if the remuneration in a capitalist society depends on scarcity of the traded product or service (and the commons are privatized as a means to increase it). Here, the commons are the common understanding, standard APIs, and interoperability.
That said, if this thing is customizable enough, a good data engineer can prepare canned steps that fit the general structure of the customer data process and it may have its place.
When I transitioned to ETL as a code 12 years ago, I felt for the first time what it looks like to be a programmer/developer, since in my case the technology was literally a means to an end. To be honest, I did not even think about tech at all.
Just to give an idea, I was responsible for 5 instances of SQL Server, and with another person we were responsible for the whole end-to-end data operation (i.e. ETL via SSIS + OLAP cube via SSAS + and reporting via SSRS), and today I struggle to have the same level of scale with less than 7 people.
Fast forward to 2024, the inverse is true: I spend way more time thinking about the tools and their specificities than about the problem that I am solving at the end of the day.
I'm not telling that ETL as a code is bad (actually there's a lot of good stuff out there) but for folks in this kind of ecosystem (e.g. azure), definitely the juice to have some homemade code and all the stuff that comes with that, does not worth the squeeze.
No it’s not. Try and see what banks, retail, manufacturing, various large enterprises still use. They need scale, observability, modularity, and maintainability.
> They need scale, observability, modularity, and maintainability.
Seems orthogonal to code-vs-gui dimension.
Once you have GUI ETL tools, in my experience, you can't modularise because the ETL tool makes assumptions about where the boundaries are that are different from what suit the domain in question. Observability falls over because you're now limited to the ETL tool instead of the domain. Scale suffers because now the ETL data model needs to be preserved and high-performance tricks might need the entire tool to be worked around, etc, etc.
Code is the highest-performance environment we have for working with huge complex systems made of if statements and loops. Giving that up to go to a tool doesn't actually yield any advantages; there needs to be an abstraction with huge practical benifits and a DAG isn't it. Modeling a DAG in a true programming language isn't hard enough to justify moving away from an IDE.
An ETL pipeline in practice is still uncomfortably close to a big spaghetti of if-thens and loops, tooling and extra models create patterns that often block a lot of the useful properties you list. The real gains come from not writing a custom scheduler, but splitting out the valuable scheduler from the ETL tool means that you have a scheduler, not an ETL tool. Sometimes there is an ecosystem of adaptors that makes a big difference, but if that doesn't meet your engineering requirements then the tool is useless (because you don't have any real levers to pull on the scale/observability/modularity and maintainability front).
From my experience SQL is basically the lowest common denominator everything speaks and even then the Oracle SQL dialect is subtly different to Microsoft SQL for example - things are subtly different enough it introduces frustrations.
There has been movement in last couple of years to hoist everything into a common "datalake" but my understanding has been that ingestion into this lake is not a simple process by any means and requires batch processes that need demanding compute resources and is slow (i.e. takes many hours and runs over night).
Node-RED is a common ETL approach in the scenario you described, but I find it too limiting beyond basic examples.
Sounds like an ideal fit for on-prem/co-located systems. The big problem with on-prem is the egress costs from wherever all your data resides.
With on-prem, doubling your hardware doesn't double your ops expenses, so it makes sense, if you already have a server-room, to fill it to capacity.
I would also add that in my 25 years in the industry I have never actually come across an ETL workflow that was complex enough that it required people working it in code. Those opinions seem to occur before proper analysis happens.
The reality is that we shouldn't be making the transformation in written data, now you have staging tables, procs, views, resulting tables procs, views, a staging database (often called staging or etl), and an unhappy DBA who yells at you every time you cross-apply incorrectly.
ETL should be done before the data lands in SQL.
I think that's orthogonal to the occasional need to use SQL to extract source data. And it's hard to beat the performance of RDBMS for your transformations. It's almost their whole raison d'être.
Edit: save in vcs stringly implies usability of ‘diff’ and ‘grep’
Six years ago I used Pentaho to do it. And it worked really well. It was easy and quick. Though maintenance was hard sometimes and it felt very dated: The javascript version was ancient, I could find a lot of questions answered online, but they were usually 5-10years old. I am wondering whether I should use something like Amphi for my next simple-ETLs.
If you can get a true CDC stream from the database to analytics, that would be ideal, but when that isn't available you spend 100x more time trying to bodge together an equivalent batch/retry system.
I built a solid career replacing no-code-ETL tools with Airflow.
You can open a pretty complicated large DAG in and right away you’ll understand the data flow and processing steps. If you were to do similar in code, it becomes a lot harder unless you comply to good modular design practices.
This is also why common game engine and 3d rendering tools come with a UI for flow driven scripting. It’s intuitive and much easier to organize.
https://successfulsoftware.net/2024/01/16/visual-vs-text-bas...
TLDR: Best depends on the situation and the people involved. Given the wide range of both situations and people, I think there are is plenty of room for both approaches.
Tldr: as with dashboards, self-serve ETLs don't work outside very specific use cases or very simple pipeline.
Reason for that, as with bi tools, is that the complexity is not in manipulating data. There are tons of frameworks to do that efficiently.
The issue is interpreting data and it's semantics and evolving data pools to the business needs.