What you will be able to do
- Name the stages of the machine learning lifecycle and the Snowflake ML capability that supports each one
- Explain why an immutable dataset snapshot matters for reproducible training
- Identify data quality issues that exploration must surface before modeling begins
- Explain how a feature store prevents training-serving skew
- Describe where models are trained in Snowflake and how experiments compare training runs
Key concept
End-to-end ML lifecycle — Building a model is a chain of stages: collect and prepare data, explore it, engineer features, train and compare models, deploy, monitor, and version. Each stage feeds the next, so a mistake made early, such as leakage or features computed inconsistently, carries through to production.
1.The lifecycle as a chain of stages
Every exam question about the ML lifecycle assumes you know the order of the stages and what each one produces. Snowflake's ML overview lists them as capabilities on one platform. You prepare data, create and use features with the Snowflake Feature Store, and train models in notebooks on Container Runtime. Experiments evaluate the trained models against set metrics. ML Jobs operationalize the pipelines, the Snowflake Model Registry deploys models for inference at scale, and ML Observability and Explainability monitor the models in production. Running underneath all of these is ML Lineage, which traces artifacts from source data to features, datasets and models, for reproducibility, compliance and debugging.
The stages form a loop rather than a straight line. Monitoring finds drift, drift triggers retraining, and retraining produces a new model version. This page covers the first half of the loop, from raw data to a trained model.
| Lifecycle stage | Snowflake ML capability |
|---|---|
| Feature engineering | Snowflake Feature Store |
| Training | Snowflake Notebooks on Container Runtime, Snowflake ML Jobs |
| Comparing trained models | Experiments |
| Deployment | Snowflake Model Registry, Model Serving to Snowpark Container Services |
| Monitoring and explanation | ML Observability, ML Explainability |
| Tracing artifacts end to end | ML Lineage |
Checkpoint 1 of 7· Match them up
Match each Snowflake ML capability to the lifecycle job it does
Tap a term, then the definition that fits it.
The overview gives each capability one lifecycle job. Lineage is the only one that covers the whole chain rather than a single stage.
“Deploy your model for inference at scale with the Snowflake Model Registry”Source: docs.snowflake.com
Sources1
2.Data collection: getting data in and freezing it
Collection means getting the data a model needs into a place where it can be governed and queried. Snowflake Notebooks let you explore data that is already in Snowflake, or upload new data from local files, from external cloud storage, or from Snowflake Marketplace datasets. When the data has to keep arriving, for example new transactions every hour, the Feature Store supports automated, incremental refresh from both batch and streaming sources. You define a feature pipeline once and it keeps updating as new data lands.
Collection also has to be reproducible. If the training table changes after a model is trained, nobody can recreate the exact rows that model learned from. Snowflake Datasets solve this by giving you an immutable, file-based snapshot of the data, with a name and an optional version. A model trained on dataset version v1 can always be traced back to v1.
These sources describe the Snowflake mechanics of collection. They do not cover general collection methodology such as sampling design or how labels are gathered, so this lesson does not teach those topics.
Checkpoint 2 of 7· Check yourself
A team wants to be able to retrain last quarter's model on exactly the rows it originally used, even though the source tables have changed since then. What should they have created at training time?
A Dataset is an immutable snapshot, so the training rows stay fixed even after the source tables change. A refreshed table would have moved on, and the other two options do not store training data at all.
“Snowflake Datasets provide an immutable, file-based snapshot of data, which helps to ensure model reproducibility”Source: docs.snowflake.com
3.Visualization and exploration: find problems before you model
Exploratory data analysis (EDA) comes before any modeling. Its purpose is to find patterns and problems while they are still cheap to fix. Snowflake Notebooks give you a cell-based environment for Python, SQL and Markdown, where you can perform exploratory data analysis and develop models in the same interface. For charts, Altair and Streamlit are imported by default. You can also install matplotlib, plotly or seaborn from the notebook's Packages menu. If you work in SQL worksheets instead, Snowsight can turn query results into bar charts, line charts, scatterplots, heat grids and scorecards, which let you quickly spot patterns and outliers.
import plotly.express as px
px.bar(df, x='measurement', y='value', color='species')Exploration is a gate, not just a set of charts. Snowflake's Auto ML skill in CoCo treats it that way. Its four-step workflow puts mandatory quality gates (leakage detection, baseline scoring and EDA) ahead of any modeling. Leakage is the finding that matters most. If a column carries information that would not be available at prediction time, the model will look excellent in testing and fail in production. A baseline score gives you a number that any real model has to beat.
Checkpoint 3 of 7· Put it in order
Put the Auto ML skill's four workflow steps in order
- 1.Build: feature engineering, model training and evaluation
- 2.Understand and configure: explore the data and propose task type, metric and trial plan
- 3.Quality gates: leakage detection, baseline scoring, EDA
- 4.Deliver: update the experiment manifest and present next steps
The skill fixes this sequence, and quality gates must pass before any modeling begins. Exploration findings such as leakage therefore get resolved before training starts.
“Quality gates: Mandatory data quality checks (leakage detection, baseline scoring, EDA) before any modeling begins.”Source: docs.snowflake.com
Checkpoint 4 of 7· Exam question
A binary churn classifier is evaluated on 1,000 customers and produces this confusion matrix: TP = 80, FP = 20, FN = 40, TN = 860. Which interpretation is correct?
Correct answer: B — Precision is 0.80 and recall is about 0.67, so 80% of flagged customers truly churned while one third of actual churners were missed.
- A. Incorrect. This swaps the two metrics. Precision divides true positives by predicted positives, and recall divides them by actual positives.
- B. Correct. Precision = 80/(80+20) = 0.80 and recall = 80/(80+40) ≈ 0.67, meaning 40 of the 120 real churners received no flag.
- C. Incorrect. 0.94 is the accuracy, (80+860)/1,000. Precision and recall are computed only from the positive-class cells.
- D. Incorrect. Recall uses actual churners in the denominator, which is the 80 true positives plus the 40 false negatives, not the flagged count.
4.Feature engineering: define once, use in training and serving
Feature engineering turns raw columns into the inputs a model learns from. The lifecycle risk here is training-serving skew: the features a model sees in production are computed differently from the features it was trained on. The Snowflake Feature Store is built to prevent this. It is an integrated place to define, manage, store and discover features. It supports both batch and low-latency online retrieval, and it keeps training and inference consistent because both read the same feature definitions.
Scale is the other concern. When the data is too large for one node, Snowflake's distributed APIs spread feature engineering and training across multiple nodes. The preprocessing functions live in snowflake.ml.modeling.preprocessing. In a well-run pipeline, preprocessing is registered once and reused everywhere, never re-implemented by hand for production.
Checkpoint 5 of 7· Check yourself
Which step best removes training-serving skew caused by preprocessing that was re-implemented separately for production?
Skew comes from two implementations of the same logic. A shared Feature Store definition gives training and inference one source of feature values.
“It supports both batch and low-latency online feature retrieval, and ensures consistency between training and inference to reduce training-serving skew.”Source: docs.snowflake.com
Checkpoint 6 of 7· Exam question
A fraud model scores card transactions where only 0.4% are fraudulent. A candidate model that labels every transaction as legitimate reports 99.6% accuracy on the test set. Which evaluation approach is MOST appropriate for judging whether the model is useful?
Correct answer: A — Review recall and precision on the fraud class from the confusion matrix, since they ignore the true negatives that inflate accuracy.
- A. Correct. With severe class imbalance, accuracy is dominated by true negatives. Recall and precision on the positive class show how many frauds are caught and how many alerts are genuine.
- B. Incorrect. Raising the threshold pushes the model toward predicting legitimate more often, which raises accuracy while catching even fewer frauds.
- C. Incorrect. R-squared is a regression metric for continuous targets and does not describe how well a binary classifier separates fraud from legitimate transactions.
- D. Incorrect. A matching average only shows calibration of the overall rate; a model can match it and still rank every individual transaction badly.
Sources4
5.Training models and comparing them with experiments
You train a model in either a Snowflake Notebook or a Snowflake ML Job. Both run on Container Runtime, which comes with packages such as scikit-learn, numpy and scipy preinstalled. XGBoost and LightGBM are also available, and GPU images support PyTorch and TensorFlow. ML Jobs let teams that work in an external IDE send their code to the same runtime. To load training data, the DataConnector reads a Snowflake table into a pandas DataFrame, a PyTorch dataset or a TensorFlow dataset, and it uses the compute pool's distributed processing to load the data faster.
from snowflake.ml.data.data_connector import DataConnector
from snowflake.snowpark.context import get_active_session
import xgboost as xgb
session = get_active_session()
# Specify training table location
table_name = "TRAINING_TABLE"
# Load table into DataConnector
data_connector = DataConnector.from_dataframe(session.table(table_name))
pandas_df = data_connector.to_pandas()
label_column_name = 'TARGET'
X, y = pandas_df.drop(label_column_name, axis=1), pandas_df[label_column_name]
clf = xgb.Classifier()
clf.fit(X, y)Training is rarely a single run. You tune hyperparameters, either with Snowflake ML's distributed HPO or with open-source libraries such as hyperopt or optuna. Each candidate run produces a result, and experiments keep those results organized. A run can be logged to an experiment while it trains on Snowflake, or you can upload metadata and artifacts from training done earlier. You then compare all the runs in Snowsight and pick the model that goes forward to deployment.
Checkpoint 7 of 7· Check yourself
A data scientist has run twenty training configurations and needs to compare them against the same metrics before choosing one for production. Which Snowflake ML capability is designed for this?
Experiments record training results so that a collection of models can be compared and the best one selected. Observability watches models that are already in production.
“Use experiments to record the results of your model training, and evaluate a collection of models in an organized way.”Source: docs.snowflake.com
Sources4
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Re-implementing training-time preprocessing in production SQL is fine as long as the logic is equivalent.Why is that wrong?
Two separate implementations of the same logic drift apart and cause training-serving skew. Define the features once in a Feature Store so training and inference read the same definitions.
Covered in Feature engineering: define once, use in training and serving
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“ML Lineage is a capability to trace end-to-end lineage of ML artifacts from source data to features, datasets, and models.”
↩︎ The lifecycle as a chain of stages“supports automated, incremental refresh from batch and streaming data sources”
↩︎ Data collection: getting data in and freezing it“Snowflake ML is an integrated set of capabilities for end-to-end machine learning in a single platform on top of your governed data.”
↩︎ Key concept“ensures consistency between training and inference to reduce training-serving skew”
↩︎ Exam trap 1“Deploy your model for inference at scale with the Snowflake Model Registry”
↩︎ Checkpoint“ensures consistency between training and inference to reduce training-serving skew”
↩︎ Prediction“It supports both batch and low-latency online feature retrieval, and ensures consistency between training and inference to reduce training-serving skew.”
↩︎ Checkpoint“Use experiments to record the results of your model training, and evaluate a collection of models in an organized way.”
↩︎ Checkpoint - 2.
“upload new data to Snowflake from local files, external cloud storage, or datasets from the Snowflake Marketplace”
↩︎ Data collection: getting data in and freezing it“perform exploratory data analysis, develop machine learning models”
↩︎ Visualization and exploration: find problems before you model - 3.
“Charts let you quickly identify and understand patterns and outliers in data.”
↩︎ Visualization and exploration: find problems before you model - 4.
“Leverage distributed preprocessing functions in snowflake.ml.modeling.preprocessing.”
↩︎ Feature engineering: define once, use in training and serving“You can train a model within either a Snowflake Notebook or a Snowflake ML Job.”
↩︎ Training models and comparing them with experiments
Also cited
“Snowflake Datasets provide an immutable, file-based snapshot of data, which helps to ensure model reproducibility”
↩︎ Checkpoint“Quality gates: Mandatory data quality checks (leakage detection, baseline scoring, EDA) before any modeling begins.”
↩︎ Checkpoint