I've solved 15 of the 50 problems in LeetCode's SQL 50. I started my first end to end ML project predicting telco customer churn this week. And three weeks in, I've realized something that nobody told me before I started: the SQL practice is what's making the ML project possible. Not the other way around.
The mistake almost everyone makes (I made it too, at first)
Most people trying to break into data science do it backwards. They jump straight to model.fit() because that's the part that feels like "real" data science the part every YouTube tutorial thumbnail promises.
But the first time you work with data that isn't already cleaned and packaged for a tutorial, you hit a wall. Real data is messy joins across three tables, inconsistent null handling, duplicate customer IDs from a system migration nobody documented, and a schema that exists only in some engineer's head from two years ago.
I found this out the hard way about two weeks into my churn project, when a join I assumed was oneto one turned out to be oneto many, and my "clean" training set had silently tripled a chunk of my customer records. The model trained fine. The results were nonsense.
What actually closed the gap
What's helped me most hasn't been more machine learning theory I already had a reasonable grip on the algorithms from coursework.
A structured learning path instead of scattered tutorials. I've been working through Microsoft's Fabric learning path alongside the SQL grind it's helped connect the querying skill to the bigger picture of how a real data platform is structured, rather than treating SQL as an isolated interview-prep exercise.
That combination is what actually caught the join bug faster the second time it happened (yes, it happened again, on the sensor dataset I'm using for my next project) I recognized the shape of the problem instead of staring at nonsense output for two hours wondering what went wrong.
If you're stuck at the same wall
If you're earlier in this process and finding that ML theory feels comfortable but real datasets feel like a different sport entirely that's not a sign you're bad at this. It's a sign you're at the normal, unglamorous part of the learning curve that most tutorials skip past because it's less exciting to film.
The fix isn't more ML content. It's slowing down on the data layer until querying and shaping data stops feeling like an obstacle between you and the "real" work because at that point, it becomes the real work.
If you want a structured way into that instead of piecing it together from scattered blog posts like I initially tried to, the Fabric learning path is worth the time.
I'm an Electrical Engineering graduate currently doing an MSc in Data Science, documenting the pivot as I go.
If you've made a similar jump from a non CS background, I'd genuinely like to hear what part of it was hardest for you — drop it in the comments.
Tags: #datascience #sql #beginners #microsoftfabric
