A platform engineer I mentor asked me the only question that really matters here. Not can I learn data science, but where do I start so I am not throwing away the ten years I already have. That is the right question, and most roadmaps answer it badly. They march you through statistics theory and Kaggle notebooks as if you were a fresh graduate, when in fact you already own a large share of the job. This guide lays out the order I actually recommend, and links a specific resource on this site for every step, so you can go from reading to building in the same week.
Where this path actually leads
Data scientist is a loose title covering at least three jobs, and picking the target before you study saves you months. A data analyst answers what happened, mostly with SQL, dashboards and clear writing. A data scientist answers what will happen or what we should do, adding statistics and modelling. A machine learning engineer makes models run reliably in production, which is the role closest to the one you hold today. Most people coming from operations land fastest, and get paid best, on the machine learning or MLOps track, because it rewards the engineering half you already bring and asks for less depth in research statistics. Start by reading how the day to day compares in what a data scientist actually does versus infra, SRE and VMware work, then sit with the harder shift in operations mindset versus data science mindset.
Naming the target early also protects your time. Research heavy roles lean on statistics and experimentation you would build slowly. Machine learning and platform roles lean on shipping, reliability and cost, which you can demonstrate almost immediately. Point your study and your portfolio at the second group and every project doubles as proof you can do the job, rather than a certificate that says you studied it. If you want the deep version of the engineering track, the AI Engineering complete guide is where it lives, and the broader Data Science complete guide covers the research side in full.
Here are the three landings side by side, then a quick way to choose between them.
| Target role | What it answers | Day leans on | Fit for an ops background |
|---|---|---|---|
| Data analyst | what happened | SQL, dashboards, clear writing | quickest first step, lightest new theory |
| Data scientist | what will happen or what to do | statistics, experiments, modelling | longest runway, most brand new study |
| ML or MLOps engineer | make models run reliably | pipelines, serving, monitoring | closest to your job today, best paid landing |
Map your starting point before you open a textbook
Before studying anything, take an honest inventory. Anaconda has measured the same pattern for years across its State of Data Science reports: data preparation and cleaning eats between 38 and 45 percent of a working data scientist day, more than model training, selection and deployment combined. That single largest slice of the job is work you already do every time you reconcile timestamps across two monitoring systems that disagree about what happened. Four operational habits carry over almost untouched, and only two areas are genuinely new. The table below is the audit; read the last column to see where your months should go, and go deeper in transferable skills for infra engineers moving to data science.
| Skill you already have | Data science equivalent | New learning needed |
|---|---|---|
| Bash and scripting | Python for data, pandas | low, same reflexes, new library |
| SQL for ops queries | SQL for analysis | low, you already know it |
| Git, CI/CD, IaC | reproducibility, ML pipelines | low, direct transfer |
| Monitoring and alerting | model monitoring and drift | low to medium, new metrics |
| Capacity planning | forecasting and time series | medium, real modelling |
| None, this is new | statistics and inference | high, the core gap |
| None, this is new | machine learning models | high, but weeks not years |
A timeline you can run on evenings and weekends
Keep your current job while you make the change. The move is far cheaper funded and far less frightening when you practise on systems you already understand than when you have quit to study distributions in the dark. A realistic part time transition runs roughly 9 to 18 months at 6 to 10 hours a week, and the phases overlap rather than run in a clean line. The chart below is the shape I coach people toward; the full week by week version is laid out in the data science transition roadmap and timeline for infra engineers.
Build order, step by step, with a resource for each
Here is the whole path in one place. Work it top to bottom, one row at a time, and follow the linked part on this site for each step rather than scattering across random tutorials. This table is the reference to bookmark; the sections after it add the reasoning and the traps.
| Step | What you do | Follow this |
|---|---|---|
| 1. Pick the target role | separate analyst, scientist, ML engineer; aim at ML or MLOps | Role compared and mindset shift |
| 2. Audit your skills | list transferables against a real job description, name the gaps | Transferable skills |
| 3. Set the timeline | plan 9 to 18 months part time, in overlapping phases | Transition roadmap |
| 4. Python for data | move past automation scripting into pandas and numpy | Python for data work and pandas for awk and jq users |
| 5. Reproducible workflow | venvs, git and notebooks that someone else can run | Reproducible data work |
| 6. Get data in | pull from SQL, APIs and monitoring; treat telemetry as a dataset | Data into Python and telemetry as a dataset |
| 7. Statistics, the core gap | percentiles, distributions, inference and real versus noise | Statistics for infra, distributions, inference |
| 8. Machine learning foundations | first model, honest evaluation, feature engineering, trees | ML fundamentals, first model, evaluation, features, boosting |
| 9. Put models in production | serving, drift monitoring, pipelines, registry, your home turf | MLOps you already run, serving, drift, pipelines, registry |
| 10. Build infra native projects | anomaly detection, forecasting, log clustering, incident prediction | Anomaly detection, forecasting, log clustering, incident prediction |
| 11. Land the role | portfolio, resume, interviews, then choose a specialism | Portfolio and resume, interviews, architect path |
Statistics and machine learning, the stretch in the middle
Steps 7 and 8 are the only part of this path that is genuinely new, so size them honestly. Statistics is weeks of focused study to working competence, not a degree. Infrastructure work rewards determinism, and statistics asks you to reason about uncertainty instead, which is the real mental shift, not the arithmetic. Lean on ground you know: you already live in percentiles over averages every time you read a p99 latency, so start there rather than with abstract probability. The modelling toolkit that follows is a handful of algorithms you will reuse constantly, learnable by building rather than reading, which is why your first model goes from a metric threshold to a trained classifier on data you recognise.
One trap deserves a warning, because it is the trap the tutorials walk you into. The hard part of a model is not fitting it, it is not fooling yourself about how good it is. Data leakage, a test set that quietly contains information from the future, will hand you a model that scores 98 percent in your notebook and collapses in production. Read model evaluation without fooling yourself before you trust any score, and pair it with the general cross validation and leakage treatment from the Data Science series. This is also where your operational scepticism, that refusal to trust a green dashboard until you know why it is green, becomes your sharpest asset.
Production is the part you already run
Here is where most data scientists are weakest and you are strongest, so it is where you should show off. Serving a model, watching it for drift, wrapping it in a pipeline and versioning it in a registry are not data science inventions. They are version control, deployment, monitoring and rollback wearing new labels. A model that is right 92 percent of the time and runs reliably for a year beats a model that scores 96 percent in a notebook nobody can deploy, and you are one of the few career changers who understands that in your bones. Start with MLOps for infrastructure engineers, the operations you already run, then work through monitoring models for drift and decay, which maps one to one onto the alerting you already build. The batch and real time serving part from the Data Science series fills in the patterns you have not met yet.
Projects that double as proof on your own data
A career changer with three sharp projects on real infrastructure data beats one with ten Kaggle notebooks, because hiring managers weight evidence over credentials. You have an unfair source of projects: the telemetry you touch daily. Four of them practically write themselves. Anomaly detection on metrics turns your alerting instinct into a model, covered in anomaly detection on metrics and time series. Capacity forecasting turns the planning you do by gut into a forecast with an error bar, in capacity forecasting from telemetry. Log clustering at scale groups millions of noisy lines into a handful of real problems, in log analysis and clustering. And incident prediction, the one everyone oversells, gets an honest assessment of what it can and cannot do in incident prediction and AIOps.
Pick one, ship it end to end, and write up the decision and the failure honestly. A project that says here is what did not work and here is how I caught it reads as senior, and it is the exact story your operations background lets you tell better than a graduate can.
Landing the role and choosing a specialism
When the projects exist, the job hunt is a packaging problem, not a knowledge problem. Frame your operations decade as the asset it is rather than apologising for it, using the portfolio and resume guide for infra engineers, and prepare for the questions that actually get asked in data science interviews and specialisation. From there the fork is yours: deepen toward the platform and delivery side through the path to AI engineer, or toward breadth and system design through the data science architect path. Both build directly on the infrastructure foundation you refused to throw away.
Your first move this week
The full path, with a part for every step, lives in the Infra to Data Science complete guide. Start the CSV this week, and let the reading follow the building rather than the other way round.
References
- Anaconda, State of Data Science report, time spent on data preparation and cleaning
- Infra to Data Science, the Complete Guide
- Data Science From Zero to Architect, the Complete Guide
- AI Engineering From Zero to Production, the Complete Guide
About The Author
DrJha