Like many people starting out, I was overwhelmed by the sheer number of things I thought I needed to learn. Data warehouses, orchestration tools, distributed processing, streaming systems, cloud platforms, infrastructure. The list seemed endless.
Instead of trying to learn everything at once, I took a different approach.
I created a 12-month self-study roadmap built around one simple idea: learn by building.
Rather than jumping from one tutorial to another, I would build a series of small projects. Each project would introduce a few new concepts while reinforcing what I had learned before. The goal wasn’t to build the most sophisticated applications possible. It was to develop the habit of thinking like an engineer by solving one problem at a time.
The first project in that journey was a GitHub ETL pipeline. It began as a simple Python script that fetched repository data and exported it to a CSV file. As I learned more, I gradually improved it. I replaced CSV files with SQLite, made the pipeline idempotent to prevent duplicate data, and eventually automated it with GitHub Actions.
By the time I finished, I realized something that hadn’t been obvious when I started.
Building the ETL logic was actually the easy part.
The harder questions appeared once I stopped thinking about a script that runs once and started thinking about a system that needs to run over and over again without me watching it.
- How should it be scheduled?
- What happens if it fails halfway through?
- Where should retry logic live?
- How do you package the application so it runs the same way everywhere?
Those questions taught me far more about data engineering than parsing JSON or writing SQL ever did.
That first project left me with another challenge.
GitHub Actions worked well for automating a small ETL pipeline, but I wanted to understand what changes when you use a workflow orchestrator built specifically for data engineering. I wanted to learn how engineers separate orchestration from execution, how containerized workloads fit into that picture, and what a production-minded pipeline actually looks like, even on a small scale.
So for my second project, I decided to build an automated RSS ingestion pipeline.
On the surface, it’s a fairly simple application. It fetches articles from an RSS feed, parses them into structured objects, and stores them in PostgreSQL.
But the goal was never to build an RSS reader.
The goal was to explore the engineering decisions that transform a Python script into a reliable data pipeline.
In this article, I’ll walk through those decisions, the mistakes I made along the way, and the lessons I learned building my first pipeline with Kestra.
Why Build Another ETL Pipeline?
After finishing my first ETL project, I considered moving on to something completely different.
Maybe a data warehouse project. Maybe Apache Spark. Maybe an API with a more complex transformation layer.
Instead, I built… another ETL pipeline.
At first, that probably sounds like a step backwards.
After all, I’d already built an extraction pipeline, made it idempotent, and scheduled it with GitHub Actions. Why repeat the same exercise?
Because I wasn’t trying to learn a new dataset. I was trying to learn a new way of thinking.
One lesson from my first project stuck with me.
Writing the ETL logic wasn’t the difficult part. The difficult part was everything surrounding it.
- How should the pipeline be executed?
- How should it recover from failures?
- How do you package it so it runs consistently on any machine?
- Where does scheduling belong?
And perhaps the biggest question of all:
Where should the responsibilities of the application end, and where should the responsibilities of the orchestration layer begin?
Those questions don’t have much to do with RSS feeds or GitHub repositories. They’re engineering questions, and I realized I could explore them with almost any data source.
That’s why I chose an RSS feed.
Not because RSS is particularly exciting, but because it’s intentionally simple.
The extraction logic only takes a few lines of Python. That meant I could spend less time worrying about business logic and more time thinking about architecture.
For the same reason, I decided to move away from GitHub Actions for this project.
GitHub Actions was a great introduction to scheduling. It showed me how to automate a workflow and gave me my first taste of running an ETL pipeline without manual intervention.
But GitHub Actions isn’t designed specifically for orchestrating data workflows.
I wanted to understand what changes when you use a tool that’s built with data pipelines in mind.
That’s what led me to Kestra.
Rather than asking, “How do I run this Python script every hour?”, I found myself asking a different set of questions.
- How should retries be configured?
- How are workflow executions tracked?
- How should environment variables be passed into a container?
- What does a failed execution look like?
- How do you separate the application from the infrastructure that runs it?
Those questions were exactly what I wanted to explore.
By choosing a simple ETL pipeline, I could focus on the engineering decisions instead of getting distracted by complicated business logic.
Looking back, I think that was the right decision.
This project isn’t interesting because it processes RSS feeds.
It’s interesting because building it forced me to think about reliability, repeatability, and orchestration in a way my first project never did.
The First Architectural Decision: Docker Before Kestra
With the project defined, my first instinct was to jump straight into Kestra.
After all, orchestration was one of the main reasons I chose this project. Why not start there?
Instead, I did something that turned out to save me a lot of frustration later.
I ignored Kestra completely.
That might sound counterintuitive, but I wanted to avoid introducing multiple moving parts before I knew the core application actually worked.
So I built the project in layers.
First, I wrote the Python ETL.
Its job was intentionally small. Fetch the RSS feed, parse each entry into an Article
object, and save the results to PostgreSQL. Nothing more.
feed = feedparser.parse(RSS_URL)
articles = parse_feed(feed)
save_articles(articles)
That’s really all the ETL did. I intentionally kept the application small because the focus of this project wasn’t the transformation logic. It was everything that happened around it.
Once that worked reliably, I turned my attention to the database.
I wanted repeated executions to be safe, so I made the inserts idempotent using PostgreSQL’s ON CONFLICT DO NOTHING
. That way, running the pipeline multiple times wouldn’t create duplicate rows.
INSERT INTO articles (...)
VALUES (...)
ON CONFLICT (id) DO NOTHING;
This one line made the pipeline safe to execute repeatedly. Whether Kestra ran the workflow once or one hundred times, PostgreSQL became responsible for preventing duplicate records.
Only after the ETL and database worked together did I introduce Docker.
That decision changed how I thought about the project.
Initially, Docker felt like another tool I needed to learn.
By the end of the project, I realized it had become something much more important.
It became the unit of execution.
FROM python:3.13-slim
WORKDIR /app
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt
COPY . .
CMD ["python", "fetch_rss.py"]
Packaging the ETL this way meant the application became self-contained. Instead of asking Kestra to understand my Python project, I could simply ask it to run a container that already knew how to execute the pipeline.
Instead of thinking, “Kestra will run my Python script,” I started thinking, “Kestra will run my Docker image.”
That distinction might seem subtle, but it completely changes the relationship between your application and your orchestration layer.
Once the ETL was packaged into a container, it no longer mattered whether it ran on my laptop, inside Kestra, or on another machine entirely.
The runtime environment was always the same.
That consistency gave me something I didn’t have before: confidence.
Before introducing Kestra, I could run the container manually and verify that it fetched the RSS feed, connected to PostgreSQL, and persisted the expected records.
If something failed, I knew the problem wasn’t hidden behind another layer of orchestration.
That validation step turned out to be incredibly valuable later.
During development, I ran into issues with networking, container configuration, and workflow execution. Because the Docker image had already been validated independently, I could immediately rule out the ETL itself and focus on the orchestration layer.
That made debugging dramatically easier.
Looking back, I think this was one of the biggest lessons from the project.
It’s tempting to connect every component together as quickly as possible and hope everything works.
A better approach is to validate each layer before introducing the next one.
In this project, the order looked like this:
- Validate the Python ETL.
- Validate PostgreSQL persistence.
- Validate the Docker image.
- Finally, let Kestra orchestrate a container that I already trusted.
Each layer built on the previous one.
By the time Kestra entered the picture, I wasn’t trying to debug Python, PostgreSQL, Docker, and orchestration at the same time.
I was only solving one problem.
That incremental approach made the entire project feel much more manageable, and it’s a workflow I’ll probably continue using on future data engineering projects.
The Assumption That Turned Out to Be Wrong
With a working Docker image, I was convinced the hard part was over.
- I had a Python ETL that worked.
- I had PostgreSQL running in Docker.
- I had verified that the container could fetch RSS articles and save them to the database.
Now all Kestra had to do was run it.
Or so I thought.
My original assumption was simple.
Kestra would point to my Python files, execute the script, and everything would work exactly as it had from the command line.
It didn’t.
I quickly discovered that there was an important distinction I hadn’t fully appreciated.
Kestra is an orchestrator.
It isn’t responsible for building Python environments or managing application dependencies. Its responsibility is deciding when and how workloads should run.
That realization changed the direction of the project.
Instead of treating Kestra as another place to execute Python code, I started treating it as the layer responsible for orchestrating a workload that already existed.
That workload was my Docker image.
Once I made that mental shift, the architecture became much cleaner.
- The ETL became a self-contained application.
- Docker became the deployment artifact.
- Kestra became the orchestrator.
Each layer had a clear responsibility, and none of them needed to know how the others worked internally.
Interestingly, reaching that point wasn’t completely straightforward.
My first instinct was to find a way for Kestra to execute the Python project directly. That approach sounded simpler, but the more I experimented with it, the more I realized I was asking the orchestrator to take responsibility for something the application should already provide.
Once I embraced Docker as the execution artifact, the workflow became much simpler.
Some approaches looked promising at first but introduced unnecessary complexity. Others worked, but didn’t align with how Kestra encourages containerized workloads to be executed.
Eventually, I landed on a workflow that felt surprisingly simple.
Instead of trying to teach Kestra how to run Python, I let Kestra do what it does best.
It launches a container.
tasks:
- id: run_etl
type: io.kestra.plugin.scripts.shell.Commands
containerImage: rss-pipeline-etl:latest
taskRunner:
type: io.kestra.plugin.scripts.runner.docker.Docker
commands:
- python /app/fetch_rss.py
Looking at the workflow now, what stands out isn’t how much YAML it contains. It’s how little Kestra actually needs to know about my application. Its only responsibility is to launch a container that already knows how to execute the ETL.
Inside that container, my application already knows exactly what to do.
That small architectural change solved more than just the immediate execution problem.
It also reinforced an idea that’s becoming a recurring theme in my learning journey.
Good engineering often isn’t about adding another layer.
It’s about giving each layer a single responsibility and allowing it to do that job well.
- Python shouldn’t worry about orchestration.
- Kestra shouldn’t worry about Python dependencies.
- Docker shouldn’t know anything about RSS feeds.
Each component solves a different problem.
Once I stopped asking one tool to solve every problem, the entire system became much easier to reason about.
Looking back, this was probably the biggest mindset shift in the whole project.
I didn’t just learn how to use Kestra.
I learned what orchestration actually means.
Once a Pipeline Runs Automatically, Everything Changes
Up until this point, I had been running the pipeline manually.
If something failed, I was sitting in front of my computer. I could read the error, make a change, and try again.
That safety net disappears the moment a pipeline starts running on its own.
One of the first things I configured in Kestra was a simple hourly schedule.
triggers:
- id: hourly_schedule
type: io.kestra.plugin.core.trigger.Schedule
cron: "0 * * * *"
On paper, it was just a cron expression.
In practice, it represented a much bigger shift.
The pipeline no longer depended on me remembering to run it.
Every hour, Kestra would start a new execution, launch the ETL container, and process the latest articles from the RSS feed.
That immediately raised another question.
What happens if one of those executions fails?
During development, I deliberately introduced failures to answer that question. I pointed the pipeline at an invalid database host and watched what happened.
The first execution failed, exactly as expected.
More importantly, it didn’t stop there.
Because the workflow was configured with retries, Kestra automatically attempted the execution again after a short delay.
retry:
type: constant
maxAttempts: 3
interval: PT30S
That was one of my favorite moments in the project.
Once I corrected the configuration, the workflow completed successfully without requiring any changes to the application itself.
That was one of my favorite moments in the project.
Not because retries are particularly complicated, but because they highlighted another separation of responsibilities.
The ETL shouldn’t decide whether it deserves another chance.
That’s an orchestration concern.
By moving retry logic into Kestra, the Python application remained focused on a single responsibility: process the feed and persist the results.
The orchestration layer handled resilience.
Scheduling introduced another challenge that I’d already encountered in my first ETL project.
Repeated executions mean repeated attempts to write data.
If the same RSS article appears in multiple hourly runs, the pipeline shouldn’t insert it twice.
Fortunately, I had already solved a similar problem before.
The database layer was designed to be idempotent using PostgreSQL’s ON CONFLICT DO NOTHING
.
INSERT INTO articles (...)
VALUES (...)
ON CONFLICT (id) DO NOTHING;
That meant every execution could safely attempt to insert the same records without creating duplicates.
The combination of scheduling, retries, and idempotent writes made the pipeline much more forgiving.
If a run failed, Kestra could retry it.
If a successful retry encountered data that had already been written, PostgreSQL would simply ignore the duplicates.
Neither layer needed to know what the other was doing.
They each handled their own responsibility.
The last improvement was visibility.
Early in development, my logs looked exactly like you’d expect from a project that was still being debugged.
I printed entire RSS objects to the console just to make sure the parser was working.
It wasn’t pretty, but it served its purpose.
As the pipeline became more stable, those debug statements became less useful.
I replaced them with logs that described the execution instead of dumping raw data.
Before
print(feed.entries[0])
After
print("=== RSS PIPELINE START ===")
first = feed.entries[0]
print("First entry preview:")
print(f"Title: {first.get('title')}")
print(f"Link: {first.get('link')}")
print(f"Feed title: {feed.feed.title}")
print(f"Fetched articles: {len(articles)}")
print(f"Saved {len(articles)} articles to the database.")
print("=== RSS PIPELINE END ===")
Each run now tells a simple story.
=== RSS PIPELINE START ===
First entry preview:
Title: Christian Ledermann: Migrate From mypy To ty And pyrefly
Link: https://dev.to/
Feed title: Planet Python
Fetched articles: 25
Saved 25 articles to the database.
=== RSS PIPELINE END ===
Those changes didn’t make the application smarter. They made it easier to understand. And that’s an important distinction.
Good observability isn’t about producing more logs. It’s about producing the right logs.
By the end of the project, I realized something interesting. The Python code hadn’t grown dramatically. Most of the work had happened around it.
Looking back, most of the engineering effort wasn’t spent making the ETL smarter. It was spent making it more reliable.
- Scheduling ensured it ran without me.
- Retries helped it recover from transient failures.
- Idempotency protected the database from duplicate writes.
- Logging made every execution easier to understand.
Individually, none of these changes were particularly complex. Together, they transformed a simple Python script into something that behaved much more like a production system.
The Final Architecture
By the end of the project, the architecture had settled into something that felt surprisingly simple.
Looking at the final architecture, it’s tempting to think the project was always heading in this direction.
It wasn’t.
Each layer was added only after the previous one had been validated.
- The ETL came first.
- Then PostgreSQL.
- Then Docker.
- Finally, Kestra.
That order mattered.
Because every component had already been tested independently, I never found myself debugging Python, Docker, PostgreSQL, and Kestra at the same time. Each decision reduced the number of unknowns instead of increasing them.
More importantly, every component ended up with a clear responsibility.
- Python knows how to fetch, parse, and persist RSS articles.
- PostgreSQL knows how to store data safely and prevent duplicates.
- Docker provides a consistent execution environment.
- Kestra decides when the workload should run and what should happen if it fails.
None of those components are trying to do each other’s jobs.
Ironically, that’s what made the finished system feel much simpler than I expected.
What This Project Changed About the Way I Think
When I started learning data engineering, I assumed the difficult part would be writing ETL code.
That’s what most beginner tutorials focus on.
- You learn how to call an API.
- You transform the data.
- You save it somewhere.
Repeat.
Those are valuable skills, but they’re only one part of the picture.
This project taught me that the engineering begins after the script works.
Once a pipeline is expected to run every hour, survive transient failures, avoid duplicate data, and produce logs that explain what happened, the questions become much more interesting.
You’re no longer thinking about individual functions. You’re thinking about systems.
One of the biggest mindset shifts for me was understanding the difference between execution and orchestration.
At first, those ideas felt almost interchangeable. Now they feel completely separate.
The Python application should focus on business logic. The orchestrator should focus on when, where, and how that application runs.
Keeping those responsibilities separate made the entire project easier to reason about.
It also changed how I think about Docker.
Before this project, I saw Docker primarily as a way to package applications.
Now I see it as a deployment artifact.
Once the ETL had been packaged into a container and validated independently, I could stop worrying about whether it would behave differently inside Kestra.
That confidence turned out to be one of the biggest advantages of containerizing the application.
Perhaps the most important lesson, though, had nothing to do with Kestra or Docker. It was the value of building incrementally.
Every major decision followed the same pattern.
- Build the smallest thing that works.
- Validate it.
- Only then introduce the next layer.
That approach made the project feel much less overwhelming than trying to connect everything together from the beginning.
Looking back, I think that’s a lesson I’ll carry into every future project, regardless of the technology.
Looking Ahead
This RSS pipeline is only the second project in my data engineering learning journey. Compared to my first ETL pipeline, the code itself isn’t dramatically more complex. What changed was the way I approached the problem.
Instead of asking, “How do I write this script?”, I found myself asking questions like:
- Where should this responsibility live?
- What happens if it fails?
- Can I run it repeatedly without worrying about duplicate data
- Can I trust it to run when I’m not watching?
Those questions pushed me to think less like someone writing Python code and more like someone designing a system. And I suspect that’s the real value of building projects.
Every project teaches a new tool. But the best projects slowly change the way you think. This one certainly did.
I’m sure the next project will challenge a completely different set of assumptions. Honestly, I’m looking forward to finding out what they are.
This is part of my ongoing series documenting my transition from systems analyst to data engineer. If you’ve been following along, thank you.