My MLOps Learning Journey, Coding Tips & a Small Gaming Distraction 🎮

SARAH GLASMACHER OCT 31 ​​​

Hey there!

it's been a while since I last updated you on my machine learning journey! Actually since my last issue I have changed jobs from data science to machine learning engineering - and while these are very similar, the new job brings with it lots of challenges related to MLOps and putting ML actually into production.

So I'm back to report on my continuous, real-world learning we all need to keep up in tech and to share what’s been working (and what hasn’t ) for me.

🔍 Learning Resource of the Week

This week, I’m deep into an MLOps course https://maven.com/marvelousmlops/mlops-with-databricks by Maria Vechtomova (follow her on LinkedIn if you're interested in all things MLOps) and Başak Eskili. Not gonna lie—the course is a LOT, but I’m learning tons, especially about using best practices on Databricks. If you’re looking to dive deeper into MLOps, this one’s solid, and they’ve got another live cohort coming up in January if you want to join. I was lucky to get it expensed via my job, it might be worth looking into if you could do the same.

Top Takeaways So Far:

• Productionizing Code: Never ever deploy a notebook or the MLOps engineers will haunt you forever. Instead, wrapping your code into proper Python packages makes life way easier for version control. I'm discovering a whole new world about .whl files and am currently working on a blogpost about this.
• Code Organization: Moving functions into .py files and keeping them on GitHub will be a lifesaver for testing and tracking changes. This ties in with avoiding notebooks if possible.
• Databricks Tip: If you need to work in a notebook but want to use your clean packaged code, you can upload the built package (as a .whl file) and install it locally in your notebook session. This way you don't compromise your clean code and can still profit from unique Databricks features.

💡 Thoughts on MLOps in Practice

One thing that keeps popping up as I go deeper into MLOps is how flexible (and sometimes chaotic) it still is. It’s not as set-in-stone as software engineering, so adapting is key. Here’s some aspects to consider:

• Data-Driven Challenges: Since ML is so tied to data, things like data drift are just part of the package. Constantly adapting is pretty much a given.
• Custom Solutions: It's tempting to just copy a workspace template that cloud companies try to sell you - however the advice I agree with is: Don’t try to do exactly what huge companies like Google or Meta do if that’s not needed. Think about your own team, your data size, and the complexity you actually need and work up from there.
• Balance & Adaptation: If you are in charge up increasing data and ML maturity in your company, resist using every shiny new thing—understand why you’re implementing something so it makes sense for your context and find a balance between increasing infrastructure complexity and actually shipping new features.

🎥 Project Updates

I vlogged my first week in this MLOps course and aiming to edit soon for YouTube. Plus, Factorio 2.0 & Space Age, the extension to my favorite game, dropped, so you can imagine the distractions!

Also, you can catch my daily updates on Instagram (coding updates, the occasional SAP data joins, and a slice of cake). I’ve got a work trip to the Data + AI World Tour Munich from Databricks coming up, so I will share live updates from that adventure

Catch you in the next update! Whether it’s next week or next year, I hope this inspires a little extra motivation to keep learning. As always, if you’ve got questions or just want to chat, hit reply!

Happy coding!
Sarah