One thing that I don't see mentioned but that does bug me: data engineers often use a lot of Python and SQL, even the ones that have heavily adopted software engineering best practices. Yet both languages are not great for this.
Python is dynamically typed, which you can patch a bit with type hints, but it's still easy to go to production with incompatible types, leading to panics in prod. It's uncompiled nature also makes it very slow.
SQL is pretty much impossible to unit test, yet often you will end up with logic that you want to test. E.g. to optimize a query.
For SQL I don't have a solution. It's a 50 year old language that lacks a lot of features you would expect. It's also the defacto standard for database access.
For Python I would say that we should start adopting statically typed compiled languages. Rust has polars as dataframe package, but the language itself isn't that easy to pick up. Go is very easy to learn, but has no serious dataframe package, so you end up doing a lot of that work yourself in goroutines. Maybe there are better options out there.
Data engineering was software engineering from the very beginning. Then a bunch of business analysts who didn't know anything about writing software got jealous and said that if you knew SQL/DBT you were a data engineer. I've had to explain too many times that yes, indeed, I can set up a CI/CD pipeline or set up kafka or deploy Dagster on ECS, to the point where I think I need to change my title just to not be cheapened.
I think even before dbt turned DE into "just write sql & yaml", there was an appreciable difference in DE vs SE. There was defo some DEs writing a lot of java/scala if they were in Spark heavy co's, but my experience is that DEs were doing a lot more platform engineering (similar to what you suggest), SQL and point-and-click (just because that was the nature of the tooling). I wasn't really seeing many DEs spending a lot of time in an IDE.
But I think whats interesting from the post is looking at SEs adopting data infra into their workflow, as opposed to DEs writing more software.
Yep, I specifically asked my company to make sure my job title was not âdata engineerâ when working on data infrastructure, because there was a growing trend of using it to mean âcan write some sqlâ.
Likewise, we had to steer HR away from âdata engineerâ because we got very mixed results with candidates.
yeah, i've seen large fortune 100 data and analytics orgs where the majority of folks with data engineering titles are uncomfortable with even the basics of git.
We have these at my company. They refuse to do any infrastructure work so you have to spoon feed the databases to them ready to go. Itâs pretty annoying.
Agreed. Weird distinction to pay less to people who did certain things and you could a high variance between "data engineers". Some who had only done a course and others that had extensive knowledge of software engineering practices were considered the same.
Hereâs an argument that college freshmen should be introduced to data science and computer science in the same introductory course. Theyâve written a textbook, which seems pretty sensible:
The comments here are... interesting, as they indicate a strong split between analysts and those engineers that can operationalize things. I see another dimension to it all.
My title is senior data engineer at GAMMA/FAANG/whatever we're calling them. I have a CS degree and am firmly in the engineering. My passion, though, is in using software engineering and computer science principles to make very large-scale data processing as stupid fast as we can. To the extent I can ignore it, I don't personally care much about the tooling and frameworks and such (CI/CD, Airflow, Kafka, whatever). I care about how we're affinitizing our data, how we index it, whether and when we can use data sketches to achieve a good tradeoff between accuracy and compute/memory, and so on.
While there are plenty of folks in this thread bashing analysts, one could also bash other "proper" engineers that can do the CI/CD but don't know shit about how to be efficient with petabyte-scale processing.
People who can utilize the tooling to process petabytes of data efficiently arenât the ones that are catching flack. The people Iâm thinking of basically run massive inefficient SQL queries and then throw their hands up when it runs slowly or gets an oom error. They donât even know how to do an explain plan. And if you try to explain to them things like partitioning, indexes, sketches, etc then they are not able to comprehend and argue that itâs not their job to learn, and that itâs the âproper engineersâ job to scale the processing.
>And if you try to explain to them things like partitioning, indexes, sketches, etc then they are not able to comprehend and argue that itâs not their job to learn, and that itâs the âproper engineersâ job to scale the processing.
Make up a person and attack him, literal strawman. You sound pleasant to work with.
Iâm referring to actual people I have worked and interacted with so no not made up.
Theyâre not engineers and shouldnât have been labeled data engineers. They have some other value to the company, presumably, but trying to repackage them as data engineers does cause issues. Thatâs the topic of this thread.
I mean this very sincerely but I'm a little lost how data engineering is distinct from software engineering. It seems like just a subset of it, my title was software engineer and I've done what sounds like very similar work.
Iâm pretty sure the term came from Google (at least that is where I heard it first described) and just referred to a backend engineer with speciality in this area. Now usually these roles have âdistributed systemsâ in the title, even if you arenât really on the inside of the systems. That or âsystems and infrastructureâ, âdata infrastructureâ, or âAI/ML infrastructureâ or sometimes âMLEâ for those kinds of orgs. Or back to good ole âbig dataâ now that itâs no longer tacked on everything.
>one could also bash other "proper" engineers that can do the CI/CD but don't know shit about how to be efficient with petabyte-scale processing.
But that would be SWEs no?
I was a 'data engineer' (until they changed the terrible title) at a startup and I ended up having to fight with Spark and Apache Beam at times, eventually contributing back to improve throughput for our use cases.
That's not the same thing as a Business Analyst who can run a little pyspark query.
I've never really seen the distinction between data and software engineering. It's more like front-end vs backend. If you're a data engineer and it's all no code tooling, then you're just an analyst or something.
When I worked at bigCo , it was a totally different world. Data engineers used data platform tools to do data work, usually for dataâs sake. Software teams trying to build stuff with data had to finagle their way onto roadmaps.
I think they've been fully converged in most strong practitioners for a long time.
There's a specific type of "data engineer" (quotes to indicate this is what they're called by the business, not to contest their legitimacy) that just writes lots of SQL, but they're usually a bad hire for businesses. They're approximately as expensive as what people call platform engineers, but platform engineers in the data space can usually do modelling as well.
When organizations split teams up by the most SWE-type DEs and the pure SQL ones, the latter all jockey to join the former team which causes a lot of drama too.
"Data engineering and software engineering are converging" says firm selling analytics products/services. I think the perspective here may be a bit skewed.
In your experience, how are folks doing (1)? The post is talking about a framework to add e.g. type safety, schema-as-code, etc. over assets in data infra in a familiar way as to what is common with Postgres; I'm not familiar with much else out there for that?
Python, R, and Julia all have at least one package that defines a tabular data type. That means we can pass tables to functions, use them in classes, write tests for them, etc.
In all of these packages, the base tabular object you get is a local in-memory table. For manipulating remote SQL database tables, the best full-featured object API is provided by R's dbplyr package, IMHO.
I think Apache Spark, Apache Ibis, and some other big data packages can be configured to do this too, but IMHO their APIs are not nearly as helpful. For those who (understandably) don't want to use R and need an alternative to dbplyr, Apache Ibis is probably the best one to look at.
Maybe. On the one side you have something like dbt or Moosestack. On the other hand analytics and data pipelining is still a lot of no code tooling and I doubt it will go away.
However I would love to learn more about how other people use coding agents to do DE tasks.
In Snowflake, I am now writing Python Stored Procedures that make REST API calls to things like Datadog REST API and dumping the JSON into a Snowflake table. I then unpack the JSON and transform it into a normalized table. So far it works reasonably well. This is possible using Snowflake's external access feature. https://docs.snowflake.com/en/developer-guide/external-netwo...
agreed on the presence and stickiness of no-code tooling. but in a future where we want to enable LLMs and agents to do as much of that work as possible, a code-first approach seems far more likely to make that effective. not just because agents are better are writing code than clicking through interfaces (maybe that will change as agents evolve?), but because the SDLC is valuable for agents for the same reasons it's valuable for human developers - collaboration, testing, auditing, versioning, etc.
That's interesting with the Zillow anecdote. I wonder if the nuance in the title is actually correlated with a difference in behavior/culture/best practices/approach?
One things have seen through my more recent exposure to experienced data engineers is the lack of repeatability rigor (CI/CD, IaC, etc.). There's a lot of doing things in notebooks and calling that production-ready. Databricks has git (GitHub only from what I can tell) integration, but that's just checking out and directly committing to trunk, if it's in git then we have SDLC right, right? It's fucking nuts.
Anyone have workflows or tooling that are highly compatible with the entrenched notebook approach, and are easy to adopt? I want to prevent theses people from learning well-trodden lessons the hard way.
Last time I worked with Databricks you could just create branches in their interface. PRs etc happened in your git provider, which for us was azure devops back then. We also managed some CI/CD.
You're still dealing with notebooks. Back then there was a tool to connect your IDE to a Databricks cluster. That got killed, not sure if they have something new.
This is insane to read as a data engineer who actually builds software. These sound like amateurs, not experienced data engineers to be perfectly honest.
There are plenty of us out here with many repos, dozens of contributors, and thousands of lines of terraform, python, custom GitHub actions, k8s deployments running airflow and internal full stack web apps that we're building, EMR spark clusters, etc. All living in our own Snowflake/AWS accounts that we manage ourselves.
The data scientists that we service use notebooks extensively, but it's my teams job to clean it up and make it testable and efficient. You can't develop real software in a notebook, it sounds like they need to upskill into a real orchestration platform like airflow and run everything through it.
Unit test the utility functions and helpers, data quality test the data flowing in and out. Build diff reports for understanding big swings in the data to sign off changes.
My email is in my profile I'm happy to discuss further! :-)
I think this may be a databricks thing? From what I've seen there's a gap between data engineers forced to use databricks and everyone else. From what I've seen, at least how it's used in practice, databricks seems to result in a mess of notebooks with poor dependency and version management.
Interesting, databricks has been my first exposure to DE at scale and it does seem to solve many problems (even though it sounds like it's causing some). So what does everyone else do? Run spark etc. themselves?
tbh I see just as much notebook-hell outside of dbx, it's certainly not contained to just them. There's some folks doing good SDLC with Spark jobs in java/scala, but I've never found it to be overly common, I see "dump it on the shared drive" equally as much lol. IME data has always been a bit behind in this area
personally you couldn't pay me to run Spark myself these days (and I used to work for the biggest Hadoop vendor in the mid 2010s doing a lot of Spark!)
We use aws glue for spark (but are increasingly moving towards duckdb because it's faster for our workloads and easier to test and deploy).
For Spark, glue works quite well. We use it as 'spark as a service', keeping our code as close to vanilla pyspark as possible. This leaves us free to write our code in normal python files, write our own (tested) libraries which are used in our jobs, use GitHub for version control and ci and so on
Many data teams often find themselves as 'tool jockeys' instead of becoming true engineers. They primarily learn some company data, and then rely on drag-and-drop or YML configuration functionality within the constraints of the tool's environment.
Their organization often insists they must use standard tools, and their idea of a good job is that the task works fine within their personal version. No automatic testing, no automated deployment, no version control, and handcrafted environments. And then they get yelled at when things break and yelled at for taking too long. And most DEs want to quit the field after a few years.
The real question is not that DE and software engineering are converging. It's why most DEs don't have the self-respect and confidence to engineer systems so that their lives don't suck.
Prefacing this with an acknowledgement that I'm a public sector data analyst by trade so my experience may not be universal.
My view is that it isn't so much a lack of "self-respect and confidence" but an acknowledgment that the path of least resistance is often the best one. Often data teams are something that was tacked on as an afterthought and the organizational environment is oriented towards buying off-the-shelf solutions rather than developing things in house.
Saying that, versional control and replicable environments are becoming standard in the profession and, as data professionals become first class citizens in organizations, we may find that orgs orient themselves towards a more production focused environment.
One thing that I don't see mentioned but that does bug me: data engineers often use a lot of Python and SQL, even the ones that have heavily adopted software engineering best practices. Yet both languages are not great for this.
Python is dynamically typed, which you can patch a bit with type hints, but it's still easy to go to production with incompatible types, leading to panics in prod. It's uncompiled nature also makes it very slow.
SQL is pretty much impossible to unit test, yet often you will end up with logic that you want to test. E.g. to optimize a query.
For SQL I don't have a solution. It's a 50 year old language that lacks a lot of features you would expect. It's also the defacto standard for database access.
For Python I would say that we should start adopting statically typed compiled languages. Rust has polars as dataframe package, but the language itself isn't that easy to pick up. Go is very easy to learn, but has no serious dataframe package, so you end up doing a lot of that work yourself in goroutines. Maybe there are better options out there.
Data engineering was software engineering from the very beginning. Then a bunch of business analysts who didn't know anything about writing software got jealous and said that if you knew SQL/DBT you were a data engineer. I've had to explain too many times that yes, indeed, I can set up a CI/CD pipeline or set up kafka or deploy Dagster on ECS, to the point where I think I need to change my title just to not be cheapened.
I think even before dbt turned DE into "just write sql & yaml", there was an appreciable difference in DE vs SE. There was defo some DEs writing a lot of java/scala if they were in Spark heavy co's, but my experience is that DEs were doing a lot more platform engineering (similar to what you suggest), SQL and point-and-click (just because that was the nature of the tooling). I wasn't really seeing many DEs spending a lot of time in an IDE.
But I think whats interesting from the post is looking at SEs adopting data infra into their workflow, as opposed to DEs writing more software.
Yep, I specifically asked my company to make sure my job title was not âdata engineerâ when working on data infrastructure, because there was a growing trend of using it to mean âcan write some sqlâ.
Likewise, we had to steer HR away from âdata engineerâ because we got very mixed results with candidates.
Ironic, since "Data Engineers" are probably far more in demand right now than "Software Engineers".
Only in places silly enough to believe software devs/enga cannot write SQL.
yeah, i've seen large fortune 100 data and analytics orgs where the majority of folks with data engineering titles are uncomfortable with even the basics of git.
We have these at my company. They refuse to do any infrastructure work so you have to spoon feed the databases to them ready to go. Itâs pretty annoying.
Agreed. Weird distinction to pay less to people who did certain things and you could a high variance between "data engineers". Some who had only done a course and others that had extensive knowledge of software engineering practices were considered the same.
Ridiculous.
Hereâs an argument that college freshmen should be introduced to data science and computer science in the same introductory course. Theyâve written a textbook, which seems pretty sensible:
https://cs.brown.edu/~sk/Publications/Papers/Published/kf-da...
The comments here are... interesting, as they indicate a strong split between analysts and those engineers that can operationalize things. I see another dimension to it all.
My title is senior data engineer at GAMMA/FAANG/whatever we're calling them. I have a CS degree and am firmly in the engineering. My passion, though, is in using software engineering and computer science principles to make very large-scale data processing as stupid fast as we can. To the extent I can ignore it, I don't personally care much about the tooling and frameworks and such (CI/CD, Airflow, Kafka, whatever). I care about how we're affinitizing our data, how we index it, whether and when we can use data sketches to achieve a good tradeoff between accuracy and compute/memory, and so on.
While there are plenty of folks in this thread bashing analysts, one could also bash other "proper" engineers that can do the CI/CD but don't know shit about how to be efficient with petabyte-scale processing.
People who can utilize the tooling to process petabytes of data efficiently arenât the ones that are catching flack. The people Iâm thinking of basically run massive inefficient SQL queries and then throw their hands up when it runs slowly or gets an oom error. They donât even know how to do an explain plan. And if you try to explain to them things like partitioning, indexes, sketches, etc then they are not able to comprehend and argue that itâs not their job to learn, and that itâs the âproper engineersâ job to scale the processing.
>And if you try to explain to them things like partitioning, indexes, sketches, etc then they are not able to comprehend and argue that itâs not their job to learn, and that itâs the âproper engineersâ job to scale the processing.
Make up a person and attack him, literal strawman. You sound pleasant to work with.
Iâm referring to actual people I have worked and interacted with so no not made up.
Theyâre not engineers and shouldnât have been labeled data engineers. They have some other value to the company, presumably, but trying to repackage them as data engineers does cause issues. Thatâs the topic of this thread.
I mean this very sincerely but I'm a little lost how data engineering is distinct from software engineering. It seems like just a subset of it, my title was software engineer and I've done what sounds like very similar work.
Iâm pretty sure the term came from Google (at least that is where I heard it first described) and just referred to a backend engineer with speciality in this area. Now usually these roles have âdistributed systemsâ in the title, even if you arenât really on the inside of the systems. That or âsystems and infrastructureâ, âdata infrastructureâ, or âAI/ML infrastructureâ or sometimes âMLEâ for those kinds of orgs. Or back to good ole âbig dataâ now that itâs no longer tacked on everything.
>one could also bash other "proper" engineers that can do the CI/CD but don't know shit about how to be efficient with petabyte-scale processing.
But that would be SWEs no?
I was a 'data engineer' (until they changed the terrible title) at a startup and I ended up having to fight with Spark and Apache Beam at times, eventually contributing back to improve throughput for our use cases.
That's not the same thing as a Business Analyst who can run a little pyspark query.
I've never really seen the distinction between data and software engineering. It's more like front-end vs backend. If you're a data engineer and it's all no code tooling, then you're just an analyst or something.
When I worked at bigCo , it was a totally different world. Data engineers used data platform tools to do data work, usually for dataâs sake. Software teams trying to build stuff with data had to finagle their way onto roadmaps.
this has been my experience too
I think they've been fully converged in most strong practitioners for a long time.
There's a specific type of "data engineer" (quotes to indicate this is what they're called by the business, not to contest their legitimacy) that just writes lots of SQL, but they're usually a bad hire for businesses. They're approximately as expensive as what people call platform engineers, but platform engineers in the data space can usually do modelling as well.
When organizations split teams up by the most SWE-type DEs and the pure SQL ones, the latter all jockey to join the former team which causes a lot of drama too.
"Data engineering and software engineering are converging" says firm selling analytics products/services. I think the perspective here may be a bit skewed.
It's not hard to do data engineering to the standards of software engineering, and many people do it already, provided that
1. You use a real programming language that supports all the abstractions software engineers rely on, not (just) SQL.
2. The data is not too big, so the feedback cycle is not too horrendously slow.
#2 can't ever be fully solved, but testing a data pipeline on randomly subsampled data can help a lot in my experience.
In your experience, how are folks doing (1)? The post is talking about a framework to add e.g. type safety, schema-as-code, etc. over assets in data infra in a familiar way as to what is common with Postgres; I'm not familiar with much else out there for that?
Python, R, and Julia all have at least one package that defines a tabular data type. That means we can pass tables to functions, use them in classes, write tests for them, etc.
In all of these packages, the base tabular object you get is a local in-memory table. For manipulating remote SQL database tables, the best full-featured object API is provided by R's dbplyr package, IMHO.
I think Apache Spark, Apache Ibis, and some other big data packages can be configured to do this too, but IMHO their APIs are not nearly as helpful. For those who (understandably) don't want to use R and need an alternative to dbplyr, Apache Ibis is probably the best one to look at.
Maybe. On the one side you have something like dbt or Moosestack. On the other hand analytics and data pipelining is still a lot of no code tooling and I doubt it will go away. However I would love to learn more about how other people use coding agents to do DE tasks.
In Snowflake, I am now writing Python Stored Procedures that make REST API calls to things like Datadog REST API and dumping the JSON into a Snowflake table. I then unpack the JSON and transform it into a normalized table. So far it works reasonably well. This is possible using Snowflake's external access feature. https://docs.snowflake.com/en/developer-guide/external-netwo...
Right, but isn't snowflake like the most expensive way to run python? why not just host the python outside of snowflake and store it there?
It is to take advantage of snowpark.
agreed on the presence and stickiness of no-code tooling. but in a future where we want to enable LLMs and agents to do as much of that work as possible, a code-first approach seems far more likely to make that effective. not just because agents are better are writing code than clicking through interfaces (maybe that will change as agents evolve?), but because the SDLC is valuable for agents for the same reasons it's valuable for human developers - collaboration, testing, auditing, versioning, etc.
If are orchestrating pipelines in airflow or Prefect you are having to write the client software around those engines, and its a lot of python.
Another anecdatum: the data engineers role at Zillow is called "Software Development Engineer, Big Data"
That's interesting with the Zillow anecdote. I wonder if the nuance in the title is actually correlated with a difference in behavior/culture/best practices/approach?
One things have seen through my more recent exposure to experienced data engineers is the lack of repeatability rigor (CI/CD, IaC, etc.). There's a lot of doing things in notebooks and calling that production-ready. Databricks has git (GitHub only from what I can tell) integration, but that's just checking out and directly committing to trunk, if it's in git then we have SDLC right, right? It's fucking nuts.
Anyone have workflows or tooling that are highly compatible with the entrenched notebook approach, and are easy to adopt? I want to prevent theses people from learning well-trodden lessons the hard way.
Last time I worked with Databricks you could just create branches in their interface. PRs etc happened in your git provider, which for us was azure devops back then. We also managed some CI/CD.
You're still dealing with notebooks. Back then there was a tool to connect your IDE to a Databricks cluster. That got killed, not sure if they have something new.
This is insane to read as a data engineer who actually builds software. These sound like amateurs, not experienced data engineers to be perfectly honest.
There are plenty of us out here with many repos, dozens of contributors, and thousands of lines of terraform, python, custom GitHub actions, k8s deployments running airflow and internal full stack web apps that we're building, EMR spark clusters, etc. All living in our own Snowflake/AWS accounts that we manage ourselves.
The data scientists that we service use notebooks extensively, but it's my teams job to clean it up and make it testable and efficient. You can't develop real software in a notebook, it sounds like they need to upskill into a real orchestration platform like airflow and run everything through it.
Unit test the utility functions and helpers, data quality test the data flowing in and out. Build diff reports for understanding big swings in the data to sign off changes.
My email is in my profile I'm happy to discuss further! :-)
I think this may be a databricks thing? From what I've seen there's a gap between data engineers forced to use databricks and everyone else. From what I've seen, at least how it's used in practice, databricks seems to result in a mess of notebooks with poor dependency and version management.
Interesting, databricks has been my first exposure to DE at scale and it does seem to solve many problems (even though it sounds like it's causing some). So what does everyone else do? Run spark etc. themselves?
tbh I see just as much notebook-hell outside of dbx, it's certainly not contained to just them. There's some folks doing good SDLC with Spark jobs in java/scala, but I've never found it to be overly common, I see "dump it on the shared drive" equally as much lol. IME data has always been a bit behind in this area
personally you couldn't pay me to run Spark myself these days (and I used to work for the biggest Hadoop vendor in the mid 2010s doing a lot of Spark!)
We use aws glue for spark (but are increasingly moving towards duckdb because it's faster for our workloads and easier to test and deploy).
For Spark, glue works quite well. We use it as 'spark as a service', keeping our code as close to vanilla pyspark as possible. This leaves us free to write our code in normal python files, write our own (tested) libraries which are used in our jobs, use GitHub for version control and ci and so on
For CI, try dagger. It's code based and runs locally too, so you can write tests. But it is a moving target and more complex than Docker.
That is what dbt solves. Version your SQL and continuously rehydrate the data to match the most recent models.
Many data teams often find themselves as 'tool jockeys' instead of becoming true engineers. They primarily learn some company data, and then rely on drag-and-drop or YML configuration functionality within the constraints of the tool's environment.
Their organization often insists they must use standard tools, and their idea of a good job is that the task works fine within their personal version. No automatic testing, no automated deployment, no version control, and handcrafted environments. And then they get yelled at when things break and yelled at for taking too long. And most DEs want to quit the field after a few years.
The real question is not that DE and software engineering are converging. It's why most DEs don't have the self-respect and confidence to engineer systems so that their lives don't suck.
Prefacing this with an acknowledgement that I'm a public sector data analyst by trade so my experience may not be universal.
My view is that it isn't so much a lack of "self-respect and confidence" but an acknowledgment that the path of least resistance is often the best one. Often data teams are something that was tacked on as an afterthought and the organizational environment is oriented towards buying off-the-shelf solutions rather than developing things in house.
Saying that, versional control and replicable environments are becoming standard in the profession and, as data professionals become first class citizens in organizations, we may find that orgs orient themselves towards a more production focused environment.