Free
- Untimed practice across all four domains
- Every question explained against the objective it tests
- Domains by name and official weight
Databricks · Machine Learning Associate
The biggest domain on this exam isn't machine learning. It's Databricks.
Databricks Machine Learning — AutoML, Feature Engineering in Unity Catalog, the ML runtime and MLflow — is 38% of the form, more than Model Development's 31%. Strong modelling instincts get you a third of the way. 48 questions in 90 minutes across four domains, against a pass mark Databricks doesn't publish.
- Weighted to the official four-domain split, not the retired Spark ML outline still quoted on prep sites
- Platform items practised as their own domain, because that's where most of the marks are
- Python throughout — scikit-learn, Spark ML and the MLflow client
Free practice across all four domains. No card required.
Machine Learning AssociateDatabricks Machine LearningMedium
A model trained on a feature table scores well offline and badly in production. Investigation shows the production scoring job recomputes features with a separate notebook that a different team maintains. What addresses the cause?
Select one.
One choice only. Your earliest pick was dropped.
Where this question sits in the exam
Databricks Machine Learning is 38% of the Machine Learning Associate exam. Our mocks use the same mix.
0 of 1 selected
An original item written against the Databricks exam guide, last checked 19 September 2026. Not a real exam question.
How it's scored
The pass mark isn't just unpublished. It moves.
Databricks sets passing scores statistically and revises them as questions change. That's their stated reason for not printing one — so the figure you've seen quoted didn't come from them.
A readiness read still needs a line, so we use 80% weighted accuracy. It's ours, not theirs, and we set it high on purpose. Being too strict costs you a week of practice. Being too lenient costs you USD 200.
Machine Learning Associate48 questions · 90 min
Accuracy by domain, weighted as the exam weights them · assumed 80% — ours, not Databricks'
Databricks Machine Learning38%68%
ML Workflows19%74%
Model Development31%71%
Model Deployment12%Furthest behind54%
Model Deployment is your lowest score. Databricks Machine Learning is the bigger drag — a smaller gap, but on 38% of the form instead of 12%.
Recertifying here means the full current exam
The credential lasts two years. Databricks publishes a 25-question renewal exam for the Data Engineer Associate and none for this credential. Recertifying means sitting whatever version of the full exam is live then.
On a platform shipping this fast the guide will have moved. Features that did not exist when you certified are on the form, and topics you revised for may be gone.
What the questions look like
Four domains, and only two of them are about models
Platform behaviour, not ML theory
The 38% domain asks what AutoML produces, where a feature table's definition lives, and what MLflow records — questions a strong data scientist who has never used Databricks will get wrong.
Diagnosis from two numbers
Model Development items often hand you a training score and a held-out score and ask what the gap means. The answer is usually about how the pipeline was assembled rather than about the model.
Deployment by constraint
Batch, streaming and real-time are all correct deployments of the same model. The latency and caller in the stem decide which, and the smallest domain on the exam is where that gets tested.
The loop
Practice, sit, review, repeat
Step 1 is free. Steps 2 to 4 need a subscription.
certs.acedify.com/machine-learning-associate
Practice · untimed
MediumA model trained on a feature table scores well offline and badly in production. Investigation shows the production scoring job recomputes features with a separate notebook that a different team maintains.
What addresses the cause?
Retrain the model on the production feature computation instead of the feature table
Log the model with the feature table as a feature spec, so scoring looks features up from the same table training used
Add the production feature notebook's output as extra columns in the training set
Schedule the production notebook to run immediately before the scoring job
The cause is a second feature implementation, not a scheduling gap. Logging the model against the feature table so scoring reads the same table training used removes the second definition. Retraining on production features would align the two by adopting the unvalidated one. Domain: Databricks Machine Learning.
Click a step to see it.
What you get back
No official line. So we show you ours.
Your readiness is a range, because it's an inference from the questions you've answered rather than a result you've earned. For a certification the outcome is binary, so the question is whether the whole range sits above the line.
The difference here is whose line it is. AWS publishes 720 and we measure you against their number. Databricks publishes nothing, so we measure you against ours — 80% weighted accuracy, set high on purpose, drawn dashed and labelled everywhere it appears.
With only four domains the weights are lopsided rather than subtle: the platform domain is three times Model Deployment. A gap there costs three times as much.
When the range stops narrowing and stops rising, more drilling isn't buying you anything, and that's worth knowing before you book.
If the real pass mark is below ours, the worst case is that you sat more practice than you needed. We'd rather make that mistake than the other one.
Machine Learning Associate · 4 mocks completed
Where you stand
Borderline
70–90
assumed 80% — ours, not Databricks'
Your range crosses our assumed line. On today's evidence it's a coin flip, and we'll say so.
Three states
Try the three states. We won't show you a percentage chance of passing — nobody can calculate one honestly without years of real outcomes to calibrate against.
What we don't do
We won't show you a percentage chance of passing. Nobody can calculate one honestly without years of real outcomes behind it.
What it actually tests
Four domains, and the platform is the biggest
Databricks Machine Learning at 38% and Model Development at 31% are more than two-thirds of the exam between them. If you've done a lot of modelling and little Databricks, the larger of those two is the one you haven't studied.
| Domain | Weight | What it covers |
|---|---|---|
| Databricks Machine Learning | 38% · about 18 questions | The largest domain, and it is the platform rather than the modelling: AutoML, Feature Engineering in Unity Catalog, the ML runtime, and the parts of MLflow this exam expects — tracking, the model registry in Unity Catalog, and runs. |
| ML Workflows | 19% · about 9 questions | Exploratory analysis, cleaning and imputation, train/test splitting, and feature engineering decisions before a model is fitted. |
| Model Development | 31% · about 15 questions | Second-largest. Training, hyperparameter tuning, cross-validation, evaluation metrics and model selection. |
| Model Deployment | 12% · about 6 questions | The smallest domain. Batch, streaming and real-time inference, and moving a registered model into serving. |
Domain weights are published on the Databricks certification page. Last checked 19 September 2026.
The retired outline still quoted on prep sites had Spark ML at 33% and Scaling ML Models at 9%. Spark ML is no longer a named domain.
Exam facts
Null fields are omitted. We do not print TBA.
- Credential
- Databricks Certified Machine Learning Associate
- Level
- Associate
- Length
- 90 minutes
- Scored questions
- 48
- Unscored questions
- Present, count not published
- Question types
- Multiple choice
- Pass mark
- Not published
- Domains
- 4
- Fee
- USD 200 + local tax
- Delivery
- Online or test centre, proctored
- Test aids
- None
- Languages
- English, 日本語, Português BR, 한국어
- Prerequisite
- None
- Recommended experience
- Six months hands-on with the tasks in the exam guide
- Validity
- 2 years
- Retake wait
- 14 days between attempts, unlimited attempts, full fee each
- Recertification
- Full current exam — no renewal exam for this credential
Pricing
Practice free. Pay when you want to know where you stand.
Pro
- Timed mocks at 48 questions in 90 minutes
- Weighted accuracy across every mock you've sat
- Your readiness range, and how it moves
Everything on every exam, one subscription
One subscription covers the Databricks certifications, AWS, the Microsoft certifications, the Salesforce certifications, the Claude certifications, TOEFL, IELTS and the SAT.
The exam fee is separate and goes to Databricks. We're prep, not registration.
Start practising for freeFAQ
Machine Learning Associate questions
Find out which domain is furthest behind
Free practice across all four domains, no card. The timed mocks and the readiness range are there when you want them.
Start practising for freeAlso available for the Machine Learning Professional, Generative AI Engineer Associate and the rest of the Databricks track on the same subscription.