保留正文与补采附件。源页面当前未重新核验；具体限制见交付清单。

Revisit consent button We value your privacy We use cookies to enhance your browsing experience, serve personalised ads or content, and analyse our traffic. By clicking "Accept All", you consent to our use of cookies. Customise Reject All Accept All Customise Consent Preferences We use cookies to help you navigate efficiently and perform certain functions. You will find detailed information about all cookies under each consent category below. The cookies that are categorised as "Necessary" are stored on your browser as they are essential for enabling the basic functionalities of the site. ... Show more NecessaryAlways Active Necessary cookies are required to enable the basic features of this site, such as providing secure log-in or adjusting your consent preferences. These cookies do not store any personally identifiable data.

Functional Functional cookies help perform certain functionalities like sharing the content of the website on social media platforms, collecting feedback, and other third-party features.

Analytics Analytical cookies are used to understand how visitors interact with the website. These cookies help provide information on metrics such as the number of visitors, bounce rate, traffic source, etc.

Performance Performance cookies are used to understand and analyse the key performance indexes of the website which helps in delivering a better user experience for the visitors.

Advertisement Advertisement cookies are used to provide visitors with customised advertisements based on the pages you visited previously and to analyse the effectiveness of the ad campaigns.

Uncategorised Other uncategorised cookies are those that are being analysed and have not been classified into a category as yet. No cookies to display. Reject All Save My Preferences Accept All Toggle SearchToggle Mobile Navigation Publish AI, ML & data-science insights to a global community of data professionals. Sign inSubmit an Article LatestEditor's PicksNewsletterShipAI

Write For TDS LinkedInXToggle Search Data Engineering

I Thought Loading Data Was the Finish Line. It Was the Starting Point.

Building my first dbt models and learning what “analysis-ready” data actually means Ibrahim Salami August 9, 20269 min read Architectural Diagram Generated by Gemini AI When I started this journey, I gave myself a 12-month roadmap to go from data analyst to data engineer. I'm only about two months into it. In that short stretch I've already built two ETL pipelines from scratch, the first one pulling GitHub repo data into SQLite, the second one pulling RSS articles into PostgreSQL with Docker and Kestra handling the orchestration. I wrote about scheduling that second pipeline to run automatically every hour, and at the time, that felt like a real milestone. The data was flowing in on its own, no manual runs, no me remembering to trigger anything. But somewhere between writing that article and starting this one, I ran a query on my own data and realized something. I couldn't sort my articles by date properly. I couldn't tell which blogs were publishing the most. The data had been sitting in Postgres for weeks, technically "loaded," and I hadn't actually looked at it closely until I needed it for something. Turns out I'd built two pipelines and skipped the part that makes the data useful. Extract, load, and then nothing. No transformation, no modeling, no real structure past "it's in a table now." This article is about fixing that. I finally sat down and learned dbt, and in the process learned what "analysis ready" actually means, because it turns out loading data and having usable data are two very different things. Learn this step by step with the interactive Data Engineer roadmap.

The Data Was Loaded. It Just Wasn't Usable.

Here's what my articles table actually looked like once I stopped and paid attention to it. The schema itself was simple, honestly about as simple as a table can get: sql

CREATE TABLE IF NOT EXISTS articles (
    id TEXT PRIMARY KEY,
    title TEXT NOT NULL,
    link TEXT NOT NULL,
    summary TEXT,
    published TEXT
);

Notice that last column. published is a TEXT field. Not a timestamp, not a date, just a plain string that happened to look like a date if you squinted at it. When I queried the ten most recent articles, this is what came back: text

title                                                    | published
----------------------------------------------------------+---------------------------------
Django Weblog: Last Call 2026 Django Developer Survey      | Wed, 08 Jul 2026 19:31:21 +0000
Mike Driscoll: New Book Release: Python Typing              | Wed, 08 Jul 2026 18:46:18 +0000

That looks fine at a glance. It's readable. But try to actually do anything with it. Want the articles from the last 7 days? You can't filter on that without casting it first, every single time, in every single query. Want to sort chronologically and trust the order? Text sorting and date sorting aren't the same thing, and depending on the format, they can quietly disagree with each other. Then there was the second problem, the one I almost missed entirely because it was hiding in plain sight. Look at those titles again: markup

Django Weblog: Last Call 2026 Django Developer SurveyMike Driscoll: New Book Release: Python Typing

Every single title in this feed follows the same pattern. Author or blog name, a colon, then the actual headline. That's real, structured information sitting inside a single text field, completely unusable as a filter or a group-by. I couldn't answer a question as simple as "which blogs post the most on Planet Python" because that information wasn't a column. It was just text, buried. So that was the actual state of things. Two pipelines built, data flowing in on schedule, and I still couldn't answer basic questions about my own data. Loading it was never the finish line. I just hadn't gotten to the starting point yet.

Why dbt, Specifically

My first instinct was to just fix this in Python, since that's the tool I already trust. Write a script that reads from articles, parses the dates, splits the titles, writes the results into new columns or a new table. And that would have worked, technically. But the more I thought about it, the more that felt like patching the same hole I'd already dug twice. Both of my pipelines were extract and load, full stop, and if I bolted transformation logic onto a Python script again, I'd just be adding a third untested, undocumented step to a system that already had two. I wouldn't be learning anything new. I'd just be writing more of the same thing I already knew how to write. dbt does this differently, and that difference is kind of the whole point of the tool. Instead of a script that runs once and produces some output you have to trust blindly, dbt models are SQL that gets version controlled, tested, and documented as part of the same workflow. You write a transformation, and in the same project you can assert things about it: this column should never be null, this ID should always be unique. If those assumptions break, you find out immediately, not three weeks later when a chart looks wrong and you have no idea why. It also matches how the industry actually works. Every data engineering job post I've looked at over the past two months mentions dbt, or something dbt-shaped. Learning it wasn't just about fixing my RSS data, it was about learning the tool that's become the default way teams handle the "T" in ETL. So instead of another Python script, I decided to actually sit down and learn dbt properly, on data I already had, with problems I already understood. Here's how that went.

Setting Up (and Immediately Hitting a Wall)

Getting dbt installed should have been the boring part. It wasn't. I tried pip install dbt-postgres and got a wall of dependency resolution errors, dbt-core had no matching distribution for my environment. Turns out I was running Python 3.14, which is new enough that dbt hadn't caught up to it yet. dbt Core officially supports up to 3.13 right now, and there's usually a lag before it supports whatever Python just released. The fix wasn't complicated once I understood the actual problem, install an older, supported Python version alongside my existing one, and build a virtual environment specifically for dbt using that: bash

py -3.12 -m venv dbt-env
dbt-env\Scripts\activate
pip install dbt-postgres

That's a small thing, but it's the kind of small thing that eats an hour if you don't know to look for it, and I think it's worth including here because it's exactly the kind of setup friction that doesn't show up in tutorials. Tutorials assume your environment already works. Mine didn't, and I'd guess I'm not the only one running a Python version that's ahead of what dbt currently supports. Once that was sorted, connecting dbt to my existing Postgres database (already running locally in Docker from my RSS pipeline) was straightforward. dbt init, pick postgres, plug in the host, port, credentials, and database name, and dbt debug confirms the connection: markup

All checks passed!

With that, I had a dbt project sitting on top of the same Postgres instance my RSS pipeline had been writing to for weeks. Time to actually fix the data.

Building the Staging Model

The first real dbt concept I had to understand was the difference between a source and a model. My raw articles table isn't something dbt built, it's external data that already exists, so dbt calls it a source. I defined that in a sources.yml file, which is really just dbt's way of formally acknowledging "this table exists, and I depend on it": yaml

sources:
  - name: rss_pipeline
    schema: public
    tables:
      - name: articles

From there, I built my first actual model, stg_articles, a staging model whose entire job is to clean up the raw data without doing anything fancy yet. This is where both of my original problems got fixed in the same file. For the date: sql

to_timestamp(published, 'Dy, DD Mon YYYY HH24:MI:SS OF') as published_at

For the buried author name: sql

split_part(title, ':', 1) as author,
trim(substring(title from position(':' in title) + 1)) as article_title

I ran dbt run, then went and actually queried the result instead of assuming it worked: markup

published_raw                     published_atSun, 05 Jul 2026 16:29:47 +0000   2026-07-05 16:29:47+00

Real timestamps. And when I checked the author split: markup

author                      | article_titlePython Software Foundation  | Python Packaging Council Inaugural Election Dates

Clean separation, even on titles with more than one colon in them, like a PyCoder's Weekly issue title that had a colon in both the source name and the headline itself. The split logic only breaks on the first colon, so it held up fine.

Adding Tests

This is the part that made the whole project feel less like "I wrote some SQL" and more like actual engineering. I added tests directly in a schema file next to the model: yaml

columns:
  - name: article_id
    tests:
      - unique
      - not_null
  - name: published_at
    tests:
      - not_null

Running dbt test doesn't just check that the SQL runs, it checks that my assumptions about the data actually hold: markup

PASS=4 WARN=0 ERROR=0 SKIP=0 NO-OP=0 REUSED=0 TOTAL=4

That not_null test on published_at specifically is the one that would have caught it if my date format string had been wrong. Instead of silently producing nulls I might not notice for weeks, I'd have seen a failed test the moment I ran it.

Building a Mart, and Finally Asking a Real Question

With clean staging data in place, I built one more model on top of it, articles_by_author, which aggregates the data into something I could actually ask a question of: which blogs post the most, and how recently. sql

select
    author,
    count(*) as total_articles,
    max(published_at) as most_recent_article,
    min(published_at) as earliest_article
from {{ ref('stg_articles') }}
group by author
order by total_articles desc

That ref() function instead of source() matters here, it's how dbt knows this model depends on stg_articles, not on the raw table directly. That dependency tracking is what builds the lineage graph later. The result was the first genuinely new thing I could see in this data since I started collecting it two months ago: markup

author                       total_articles   most_recent_articlePython Software Foundation   5                2026-07-09 14:11:06+00Django Weblog                4                2026-07-08 19:31:21+00

A question I couldn't answer a week earlier, answered in one query, on data I'd already had sitting around the whole time.

Seeing the Whole Thing

The last step was running dbt docs generate and dbt docs serve, which builds an interactive documentation site with a lineage graph, basically a visual map of how data flows through the project. Mine showed exactly three connected nodes:

Where This Leaves Me

So that's the project. Two clean models, seven passing tests, and a lineage graph that actually shows a real chain from raw data to something I can ask questions of. Compared to where I started this piece, unable to sort by date or tell which blogs posted the most, that's a real shift, even if the underlying dataset didn't change at all. Same data. Very different usefulness. I want to be honest about what this isn't, though. This is still running entirely on my own machine. The Postgres database, the dbt project, all of it lives locally in Docker, which means none of this exists anywhere the moment my laptop is off. There's also only one RSS feed feeding into this right now, so "which blogs post the most" is a fairly small question with a fairly small dataset behind it. And I haven't touched anything around alerting or monitoring if a test starts failing quietly in the background. None of that takes away from what I actually learned here, though. I think there's a difference between a project being finished and a project having taught you what it was supposed to teach you. This one did the second thing. I understand the difference between a source and a model now. I understand why tests aren't optional if you actually want to trust your own data. And I understand, in a very concrete way this time, why "the data is loaded" and "the data is usable" are two completely different claims. The next problem is obvious, honestly. Everything I built here still depends on my laptop being on and Docker running. That's the next wall I'm going to hit, and probably the next thing I write about, taking this whole stack off my machine and putting it somewhere it can actually run without me. Two months into a twelve month roadmap, and I think that's about right. Slower than I'd like some days, but every wall I've hit so far has taught me something I couldn't have learned by reading about it first. Thanks for reading! This is part of my ongoing series documenting my transition from systems analyst to data engineer. If you’ve been following along, thank you. Connect with me on LinkedIn ,YouTube , and Twitter .

Register Now to Continue Reading

Unlock Towards Data Science with a Free Account

All articles on Towards Data Science are available to registered users. Create a free account today to access this article and thousands more. Register for freeSign in to your account

Written By

Ibrahim Salami

See all from Ibrahim Salami Career ChangeData EngineeringData PipelineData ScienceDbt Share This Article

Towards Data Science is a community publication. Submit your insights to reach our global audience and earn through the TDS Author Payment Program. Write for TDS

Related Articles

A definitive explanation to Hinge Loss for Support Vector Machines.

Artificial Intelligence A fully comprehensive, clear & concise explanation to help you uncover the real truth about the enigmatic cost function. Vagif Aliyev November 23, 20206 min read

Analyze US Debt to the Penny with the US Treasury Debt API, Python, and Tableau

Data Science The United States Treasury makes it easy to retrieve US federal government debt data with its Debt Information API Randy Runtsch April 23, 20226 min read

Playing Moneyball: Creating an efficient NBA team with linear programming

Data Science Inspired in the movie Moneyball, we build an efficient NBA team, learning about the basics of linear programming Nicolas Garcia Aramouni January 23, 20228 min read

4 Vital Tips for Choosing a Business Use Case for Data Science

Data Science According to Gartner, only 20% of data science projects bring business outcomes by 2022. So, applying data science is a promising but... Andrey Koptelov October 4, 20216 min read

Forget About Cloud Computing. On-Premises Is All the Rage Again

Data Engineering From startups to enterprise, companies are lowering costs and regaining control over their operations Ari Joury, PhD March 14, 20257 min read

Map, Filter, Reduce - Higher-Order Functions for Functional Data Pipeline Design

Data Engineering A deep-dive on higher-order functions and how they can be used in designing data transformations for functional data pipelines Ong Chin Hwee February 7, 202210 min read

What I Got Wrong: Looking Back at My 2022 Predictions for the Modern Data Stack

Data Science Where we started vs. where we are now in the data world Prukalpa Sankar December 22, 202214 min read

3 Challenges to Being a Data Scientist in 2024

Data Science Given the current climate, is data science for you? Haden Pelletier June 27, 20247 min read

Data as a product vs data products. What are the differences?

Data Engineering Understand what data as a product is by looking at an example. Xavier Gumara Rigol July 8, 20216 min read

Your home for data science and AI. The world's leading publication for data science, data analytics, data engineering, machine learning, and artificial intelligence professionals. Subscribe to Our NewsletterWrite For TDS•About•Advertise•Privacy Policy•Terms of Use © Insight Media Group, LLC 2026•The New Stack•roadmap.sh•Creator Network

Recommended Articles

Close