I am currently building my analytics pipelines on top of R's dbplyr package, which allows me to run dplyr operations on database tables just as if they were local data frames. Since it's R, it's not without warts, but at least I get to use a real programming language and the free and fantastic RStudio IDE.
Which require that data take a round trip between the R process and the database. It's not uncommon for these jobs to spend more time in the read/write step than doing meaningful work, and why I prefer dbt when possible to keep transformations as close to the data as possible.
Sadly, the company closed this June, and even the website is already down.
ETA: Seems it isn't. https://www.deepchannel.com/ That what you meant?
In that sense it should only be used for last mile data transformations. Not for wrangling raw data into neat tables.
Edit: I am also curious why you consider SQL to be inferior for the latter?
After that step the data is a lot more uniform, so then it's easy to use SQL.
A lot of these discussions are basically "can we please just use the language/tooling we know", and sometimes the answer's yes! A lot of people breathe a big sigh of relief when they discover you can use like, Spark through Python and what-not. Regardless of whether or not it's a good idea or good use of resources, this space is advancing because the market demands it. But like, I don't think you're ever gonna use dbt with not-SQL; it's all in on SQL. Feels like maybe it was the wrong fit for your team.
I should say I'm not sure if your coworker was an SQL person so dunno if this directly applies. Just saying what my experience has been.
Testing is such an interesting concept in data engineering. One needs consistent test data. We aim to implement that with snapshots eventually but now we have sanity checks at each layer.
> lacking industry guidance and generic tooling to support: CI/CD, versioned deployment artifacts, zero downtime deployments, unit testing, observability, monitoring and alerting.
Ci cd is doable and easy: changes are deployed via GitHub actions, the CI part is a bit missing I guess without tests. Versioned artifacts also, add a tag to the airflow job to know what tagged version of the code is running. Zero downtime deployments we do that all day — I mean using views we can a/b deploy changes to the underlying tables and then do a simple schema change to the view and nobody knows the difference. Unit testing still yet to be done. Observability and alerting we use internal dashboards and sentry.
What am I missing?
The first thing he did was remove real data from training, as he was shocked to see that development wasn’t done against dummy data.
And once the dev environment is running on real data, the changes in how you develop and operationalize just seems to cascade in my experience.
It is best practice to keep live data out of your development environment for normal software development, but it is impossible for machine learning projects.
It is somewhat frustrating to see the engineering/platform team spending time on this kind of abstraction that is essentially useless for doing the work we need to do.
The most significant of these is that version control is not guaranteed to be the primary gate to changes to your system’s state. The size, shape, volume, skew of your data can change independently of any code. You can have the best unit/integration tests and airtight CI/CD but it doesn’t matter if your date source suddenly starts populating a column you were not expecting, or the kurtosis of your data changes and now your once performant jobs are timing out.
Compounding this, there is no one answer for how to handle data changes. Sure you can monitor your schema and grind everything to a halt the second something is different in your source, but then you’re likely running a high false positive rate, and causing backup or data loss issues for downstream consumers.
The truth is change management and scaling is more or less a predictable and solved problem in traditional SWE. But the same principles can’t be applied with the same effectiveness to DE. I have seen a number of SWE-turned-DE struggle precisely because they do not believe there to be any difference between the two. It would be nice if it were so simple.
That has been true in SWE since the days of ENIAC.
How you handle contract violations between dependencies and third parties is not an inherently different problem. What should you do if a data source changes schema? How do you detect it, alert, take action? How are users or customers or consumer jobs impacted? This is all just normal software.
In many ways it feels like you're making my case for me.
What are some good resources for the other things you describe?
You’ll want to read up on what people do on error, especially in the case that contracts change (search “api versioning”.)
It may be that those topics are part of the SWE process but not generally applicable for example UI work and therefore overlooked.
But it may also be something else. That a process can create Grafana dashboards and processes for API updates like a pro.
But only the person that really cares about the data and know it from inside out can catch the "unknown unknowns". And for that you need to give access (and hand over responsibilities) to Analyst persons too and in some ways trust them too, even though they may not have as rigorous processes.
This definitely stood out to me when I was working with a lot of Snowflake’s new products, especially their “native apps”.
I started on some bespoke tooling given the lack of everything that you mentioned and at the time was thinking about how to productize it / turn it into some open source tools but I’ve not gotten back around to it since leaving that job.
I had a really great experience working with sqlglot from Tobiko Data https://tobikodata.com/ — haven’t had a chance to check out their sqlmesh project yet but I have some faith in their work.
It enables source code control, modularity, testing, CI/CD, environments, documentation, git branching, local development experience.
Not a fanboy, but it's who reason for being is basically to enable "software engineering practices for data".
(Dang, and here the tool I've been designing in my head for several years was going to claim that niche... So unique. ;-)
- employees
- tall_employees
- tall_employees_by_salary
These tables (dbt would call them models because they can also be views depending on how they're configured) depend on each other, so you have to build them in a specific order. Without something like this, you're manually running SQL to build your pipelines, and that's very error prone. With something like this you can run tests (they're also SQL in dbt, which is pretty common in data engineering teams), run your pipeline in CI, etc.
There's a bunch of other features, but they all hang off this basic idea.
you can have some sh file which runs SQL in desired order..
Sure, one at a time. dbt builds a DAG of the SQL operations and can run transformations concurrently.
So things finish considerably faster?
> then optimizing SQL is likely much more beneficial
This only gets you so far. Running transformations one-at-a-time when the server can do more is just holding you back. My last dbt project had over 300+ tables and running transformations serially would have taken an order of magnitude longer than running concurrently.
in general, the end goal of such ETL optimization is to have all your CPU or IO saturated. But I imagine for 300 tables transformations it could be too much work.
This particular dbt process was running against the second smallest Redshift instance, which barely noticed when the process kicked off every 10 minutes to run incremental updates. Would run for maybe 60-90 seconds, with only the occasional lock contention with long-running analyst queries that was easily mitigated.
I said goal of "such ETL optimization", not ETL.
Luckily for me it is an hour and 50 lines(likely less) of code to write such parallel executor with dependency tracking if need will arise.
Where we ran into issues was:
- Availability: the general model of dbt is DROP the model (OK by default they TRUNCATE but w/e) and rebuild it. But this means that model's unavailable (read: locked) during rebuilding. Further, you can't do that if any downstream view depends on your model. Even if you decide it's OK to drop the downstream view, rebuild all its upstream dependencies, then rebuild it, it's unavailable the whole time. Further, you probably don't have just one downstream view depending on these upstream models, so you've got to work out all the downstream views to DROP and then rebuild, which is like a secondary dependency tree to walk while walking the primary dependency tree. But, that mind-bending weirdness aside, you're essentially making your whole DB unavailable during a build. Sometimes this is OK, for us it really wasn't.
- Incremental tables added a lot of complexity: rebuilding bigger tables that mostly don't change is a big waste of time and resources. But this is an entirely parallel code path which more or less doubled our code burden.
We tried pretty hard to manage these, which ultimately was a big rabbit hole. We eventually settled on schema versioning where we'd build a whole new DB in a different schema, run tests on it, and promote it if it passed. This let us avoid both the availability and incremental ditches, and as a bonus we could instantly restore past schemas if we discovered issues (restoring DB backups is sloooooooow and lossy).
But, I think the engineering lesson here was we either should have rescinded that weirdo app DB rule, or restricted our custom work to ELTing the data out of our apps and used dbt downstream of that.
The first is that you can update a subset of your pipeline, like say you have an internal dashboard for server stats and another external dashboard that's a product for customers, and you notice a problem w/ the server stats dashboard. You can have dbt just build the models that underlie the server stats dashboard and also exempt any model used by the external dashboard.
The second is that you can effectively defer processing. dbt builds views by default (though it of course will also build tables, including ephemeral tables that are cleaned up after a build), which means your transforms don't take place until you query the data. If you can get away with it, this is what you want, and there are lots of ways to stay on views (caching, use BigQuery, etc). This means your pipeline at least has the potential of being "real time".
But, overall we're talking about pipelines that have 100s or 1000s of models. You need some kind of system for that; a shell script isn't gonna cut it.
run_all_etl.sh
build_internal_dashboard_etl.sh (called from run_all_etl.sh)
build_external_dashboard_etl.sh (called from run_all_etl.sh)
hundred tables. one fail -> whole pipeline fails to ensure e2e and overall consistency.
A B C D
1 2 3
i ii
Ai i-ii
> You obviously can have modularity with shell scripts
Just as kind of a meta point, I think you're assuming that people are introducing a lot of incidental complexity into this problem, but I don't think you're open to the possibility that there's a lot of essential complexity here. I'm talking about DAG tools and the features that hang off of them, and you're saying "hey there's new thing called shell scripting". That's not useful in this conversation. Maybe going forward, let's assume people know about Bash.
If you have kitchen sink of 1000 tables chaotically connected to each other, then yes, you need DAG management tool to manage chaos..
This describes 99% of data teams, so I think we agree!
there is a chance that dbt incentivies data teams to go that route, otherwise they would think about modularity and ownership more.
I think I agree, but I think there are a lot of things you could say about this. What if you're just starting out and you don't know how to efficiently design systems? What if you work in an industry of people under 40 (under 30?) who by definition can't have the kind of experience that would let you design small systems? What if you work for a company that doesn't want to give you the time to figure it out? What if you have a competitor that's just shipping trash and forcing you out of the market with a "we'll clean up the tech debt later" mentality? How do we make tools that are powerful enough to enable us to do hard things, but wise enough to keep us from doing foolish things? Should we rely on our tools to do this? Is this an argument against centralization? Isn't that deeply inefficient?
Honestly I struggle with this all the time. I've got the outlines of some ideas and there are some tech communities grappling with the same issues, but I think we're just at the very beginning of a rethink about how we do software.
IMO: No we shouldn't. And that's why I'm in the "Think about it yourself, bulid your ETL with a clean structure so you know what your 1,000 [SQL|shell] scripts do" in stead of the "Let your tool figure it out for you" one.
> Is this an argument against centralization? Isn't that deeply inefficient?
Or for centralization, if we hark back to the "You can get a loooong way on a single server" argument above. Depends on how you define "centralization", I suppose. (Seems to me everything always depends on your definitions, in the end.)
But wait, does this still have anything to do with IDEs? As opposed to Vim (or, say, Notepad)? If so, I'd say IDEs also belong to the category good centralization. At least older ones, before they became desktoppified Web apps with baked-in AS (Artificial Stupidity)...
> Honestly I struggle with this all the time. I've got the outlines of some ideas and there are some tech communities grappling with the same issues, but I think we're just at the very beginning of a rethink about how we do software.
On that debate I'm not so much conservative as... Reactionary.
Then you get things like testing, documentation, modularity if you want it.
We Platform just help them test or setup some things from time to time, but that workflow they have with dbt works quite well, and all the things they’ve complained in these threads we have yet to suffer (at the magnitude people is complaining).
Not sure if there is something “better” these days, but these teams have not had a single major incident with their dbt workflow.
I’ve been using dbt and MDS for nearly 3.5 years and I believe the entire approach is profoundly broken. There’s really nothing “modern” about it, especially compared to software engineering.
https://on-systems.tech/blog/135-draining-the-data-swamp/
Building on Extracted operational data is hard at best and a business altering security liability at worse.
I believe The MDS trails at least a decade behind modern software engineering practices, lacking industry guidance and generic tooling to support: CI/CD, versioned deployment artifacts, zero downtime deployments, unit testing, observability, monitoring and alerting.
MDS Data engineering is a meme in the industry, the expectation of “100%” combined with the lack of modern tooling makes success really hard to achieve.