Overview
Collaborating within and across teams to manage data and models has three official objectives. Task 2.1 covers organizing tabular, text and image data, choosing a preprocessing tool by scale (BigQuery SQL, Dataflow, Apache Spark or in-memory Python), building features in Feature Store and protecting PII. Task 2.2 covers prototyping in Agent Platform Workbench and Colab Enterprise with secure, shared setups, frameworks such as PyTorch, scikit-learn and JAX, and Model Garden models. Task 2.3 covers experiment environments, evaluating predictive and gen AI solutions with metrics and LLM-as-a-judge, and tracking artifacts and lineage.
At about 16% of the exam, this section focuses on reproducibility and teamwork rather than algorithms. Expect scenarios where several data scientists share data, notebooks and results under security and privacy rules.
How it's tested
Domain 2 questions describe a team, a data set with its size and sensitivity, and a collaboration problem. You choose the preprocessing tool, the notebook environment and its security settings, the place to store features, or the way to log and compare runs. Size, sensitivity and reproducibility are the deciding clues.
-
Tool by scale
Data volume and format are given, and you choose BigQuery SQL, Dataflow, Spark or pandas for preprocessing.
-
Secure collaboration
Several users share notebooks or data with compliance limits, and you pick the environment, access controls or de-identification step.
-
Compare and trace
Many training runs or gen AI prompts must be compared, and you choose the evaluation method or the tracking service.
A team must clean 40 TB of clickstream logs daily and the same logic must handle streaming events later. Pandas in a notebook cannot hold the data. A single Spark job could work but adds cluster management. Dataflow runs Apache Beam pipelines that scale automatically and use one code base for batch and streaming, so Dataflow is the expected choice.
Objectives 3
Official objectives from the Google Cloud outline, in outline order.
2.1 Exploring and preprocessing data for MLOrganizing tabular, text and image data, choosing preprocessing tools by scale, Feature Store and privacy.
- What it means
- This task covers preparing data for experimentation, training and serving. You organize different data types, choose BigQuery SQL, Dataflow, Apache Spark or in-memory Python based on scale and complexity, create and consolidate features in Feature Store, and protect sensitive data such as personally identifiable information.
- What to understand
- Match the tool to the data. Small samples fit in pandas; warehouse-scale joins and aggregations fit BigQuery SQL; large or streaming transformations fit Dataflow; existing Spark code fits Dataproc or serverless Spark. Feature Store on Agent Platform keeps feature definitions in one place, backed by BigQuery, and serves them online with low latency so training and serving read the same values. For privacy, use Sensitive Data Protection to find and mask, tokenize or redact PII before it reaches training data, and apply IAM and column-level controls in BigQuery.
- How it shows up
- A data volume, format or privacy requirement is given, and you choose the preprocessing tool, feature store pattern or de-identification method.
- Common confusion
- Loading a multi-terabyte table into pandas inside a notebook because the code is already written.
2.2 Model prototyping using notebooks (e.g., Gemini Enterprise Agent Platform Workbench and Colab Enterprise)Notebook security and collaboration, frameworks such as PyTorch, scikit-learn and JAX, and Model Garden models in notebooks.
- What it means
- This task covers building early models in managed notebooks. You apply collaboration and security practices, prototype with common frameworks in Agent Platform Workbench or Colab Enterprise, and pull foundation and open models from Model Garden into a notebook for experimentation.
- What to understand
- Workbench instances are customizable JupyterLab VMs with chosen machine types, GPUs and custom containers. Colab Enterprise offers managed, shareable notebooks with runtimes that start quickly and IAM-based sharing. Secure either one with a dedicated service account, least-privilege IAM, no public IP, VPC Service Controls and customer-managed encryption keys when required. Keep notebook code in a Git repository so work is reviewed and reusable. Model Garden lets you load open models such as Gemma or call Gemini from the notebook to test ideas before committing to tuning or training.
- How it shows up
- A team needs to share or secure notebooks, or test a foundation model quickly, and you pick the environment and controls.
- Common confusion
- Thinking notebook prototypes are ready for production; they still need to be moved into pipelines.
2.3 Tracking and running ML experimentsExperiment environments, predictive and gen AI evaluation including LLM-as-a-judge, and artifact and lineage tracking.
- What it means
- This task covers running experiments in a repeatable way. You choose an environment such as Agent Platform Experiments, Agent Platform Pipelines or Kubeflow Pipelines based on the framework, evaluate predictive and generative solutions with metrics and LLM-as-a-judge, and track model artifacts, versions and lineage with Experiments and ML Metadata.
- What to understand
- Experiments logs parameters, metrics and artifacts per run so you can compare them side by side, and TensorBoard adds training curves. Predictive models are judged with metrics like precision, recall, AUC, RMSE or MAE chosen to match the business cost of errors. Generative output often has no single right answer, so the Gen AI evaluation service combines computed metrics such as ROUGE or BLEU with model-based judging, where a judge model scores fluency, groundedness or safety against a rubric, or compares two candidates pairwise. ML Metadata records which data, code and parameters produced each artifact.
- How it shows up
- Many runs or prompts need comparison, or an auditor asks what produced a model, and you choose the tracking or evaluation approach.
- Common confusion
- Using exact-match accuracy to grade open-ended generated text, which penalizes valid paraphrases.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Dataflow
Managed Apache Beam service for batch and streaming pipelines.
- Remember
- Autoscaling, one code base for both modes.
- Tested as
- Large-scale preprocessing scenarios.
- Trap
- Choosing it for a few megabytes of data.
-
Dataproc and serverless Spark
Managed Spark and Hadoop on Google Cloud.
- Remember
- Best when Spark code already exists.
- Tested as
- Migration and reuse scenarios.
- Trap
- Rewriting working Spark jobs without a reason.
-
Feature Store
Central registry and serving layer for ML features, backed by BigQuery.
- Remember
- Same feature values for training and online serving.
- Tested as
- Feature reuse and skew questions.
- Trap
- Treating it as a general database.
-
Point-in-time correctness
Joining features as they were when each label occurred.
- Remember
- Prevents future data leaking into training.
- Tested as
- Feature retrieval questions.
- Trap
- Joining on the latest value.
-
Sensitive Data Protection
Service to discover, classify and de-identify sensitive data.
- Remember
- Masking, tokenization and redaction of PII.
- Tested as
- Privacy scenarios.
- Trap
- Relying only on dropping a name column.
-
Agent Platform Workbench
Customizable JupyterLab instances on Agent Platform.
- Remember
- Choose machine type, GPUs and containers.
- Tested as
- Heavier prototyping needs.
- Trap
- Ignoring network and identity settings.
-
Colab Enterprise
Managed, shareable notebooks with IAM controls.
- Remember
- Fast start and easy sharing.
- Tested as
- Collaborative prototyping questions.
- Trap
- Confusing it with free consumer Colab.
-
Agent Platform Experiments
Logs runs with parameters, metrics and artifacts.
- Remember
- Compare runs side by side.
- Tested as
- Experiment tracking questions.
- Trap
- Using spreadsheets to compare runs.
-
ML Metadata and lineage
Records links between data, code, runs and models.
- Remember
- Answers what produced this model.
- Tested as
- Audit and reproducibility questions.
- Trap
- Assuming the model registry alone gives lineage.
-
LLM-as-a-judge
A model scores generated output against a rubric.
- Remember
- Pointwise scoring or pairwise comparison.
- Tested as
- Gen AI evaluation questions.
- Trap
- Skipping human spot checks of the judge.
-
Metric choice
Precision, recall, AUC, RMSE and similar measures.
- Remember
- Pick the metric that reflects business cost.
- Tested as
- Evaluation questions.
- Trap
- Using accuracy on imbalanced classes.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Preprocessing tools by scale
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| In-memory Python (pandas) | Single-machine processing in a notebook. | Fine for samples and small data. | Running it on full production data. |
| BigQuery SQL | Serverless SQL over warehouse tables. | Ideal for joins and aggregations at scale. | Exporting data just to transform it. |
| Dataflow | Apache Beam batch and streaming pipelines. | Handles streaming and complex transforms. | Overkill for one-off small jobs. |
| Apache Spark (Dataproc) | Distributed processing with Spark. | Reuse existing Spark code and libraries. | Choosing it when no Spark skills exist. |
Notebook environments
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Agent Platform Workbench | Configurable JupyterLab instance. | More control over hardware and environment. | Picking it only for quick sharing. |
| Colab Enterprise | Managed notebooks with built-in sharing. | Faster setup and collaboration. | Expecting deep VM customization. |
Evaluation methods
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Computed metrics | Formulas such as AUC, RMSE, BLEU or ROUGE. | Cheap and repeatable. | Missing meaning in open-ended text. |
| LLM-as-a-judge (pointwise) | Judge model scores one response on a rubric. | Captures quality aspects like groundedness. | Not validating the judge. |
| Pairwise comparison | Judge model picks the better of two responses. | Good for choosing between model or prompt versions. | Reading it as an absolute quality score. |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Pandas at scale
- What candidates pick
- Increase the notebook VM memory
- Why it looks right
- It is the quickest fix.
- Why it's wrong
- Vertical scaling hits limits and is costly; BigQuery or Dataflow processes the data in a distributed way.
Clue to watch for Data size is in terabytes or growing daily.
-
Drop the column
- What candidates pick
- Remove the name column to protect privacy
- Why it looks right
- Names are the obvious identifier.
- Why it's wrong
- Emails, addresses, IDs and free text can still identify people; use Sensitive Data Protection to find and de-identify them.
Clue to watch for Records contain free-text or many quasi-identifiers.
-
Latest feature value
- What candidates pick
- Join each label to the current feature value
- Why it looks right
- It is simple to query.
- Why it's wrong
- It leaks future information into training; point-in-time retrieval uses values as they were when the label occurred.
Clue to watch for Historical training data with time-varying features.
-
Public notebook
- What candidates pick
- Give the notebook a public IP for easy access
- Why it looks right
- It simplifies connectivity.
- Why it's wrong
- It widens the attack surface; private networking with VPC Service Controls meets security rules.
Clue to watch for Regulated data or an exfiltration concern.
-
Registry equals lineage
- What candidates pick
- Rely on the Model Registry to show training data used
- Why it looks right
- The registry stores model versions.
- Why it's wrong
- Versions alone do not record inputs; ML Metadata and Experiments capture data, parameters and runs.
Clue to watch for An auditor asks how a model was produced.
-
Accuracy on rare classes
- What candidates pick
- Report accuracy for a fraud model
- Why it looks right
- Accuracy is familiar.
- Why it's wrong
- With rare positives, a model predicting no fraud scores high accuracy; precision, recall or PR AUC show real performance.
Clue to watch for The positive class is a small share of data.
Answer framework
Work through these steps when a question involves shared data, notebooks or experiment results.
- Size and format the data first: small samples go to pandas, warehouse tables to BigQuery SQL, streaming or huge transforms to Dataflow, existing Spark jobs to Dataproc.
- Scan for sensitive fields and de-identify them with Sensitive Data Protection before any copy reaches a training or notebook environment.
- Define reusable features once in Feature Store and use point-in-time retrieval so training and serving read matching values.
- Prototype in Colab Enterprise for quick sharing or in Workbench for custom hardware, locked down with service accounts and private networking.
- Log every run to Experiments, pick metrics that reflect business cost, and use rubric-based or pairwise judging for generated output.
Practice question
Want more like this? The GCP PMLE practice test has questions for every domain.
Checklist
0 of 12
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Build one small project that touches every task. Load a public BigQuery dataset, transform it with SQL, then rewrite one step as a Dataflow template to see the difference in setup. Create a feature group in Feature Store from a BigQuery view and read it back online. Run Sensitive Data Protection on a table that contains fake emails and phone numbers and try masking and tokenization.
Next, open the same notebook in Colab Enterprise and in a Workbench instance, and note the differences in sharing, hardware and network settings. Train two small scikit-learn models, logging both to Agent Platform Experiments, and compare them in the console. Then run the Gen AI evaluation service on two prompt templates with a pointwise and a pairwise metric.
Spend extra time on tool selection by scale, because those questions are common and easy to lose. Keep a list of clue words such as streaming, terabytes, existing Spark code or regulated data, with the tool each one points to.
FAQ
How much of the GCP PMLE exam covers Collaborating within and across teams to manage data and models?
Collaborating within and across teams to manage data and models is about 16% of the exam, roughly eight to ten questions, and has three official objectives.
What is the difference between Workbench and Colab Enterprise?
Workbench gives a configurable JupyterLab instance with chosen hardware; Colab Enterprise gives managed notebooks that start fast and are easy to share.
Why use Feature Store instead of BigQuery tables alone?
It registers features for reuse, supports point-in-time retrieval for training and serves values online with low latency, which reduces training-serving skew.
How do I protect PII in training data?
Use Sensitive Data Protection to find and de-identify sensitive fields, restrict access with IAM and column-level security, and keep raw data out of notebooks.
What is LLM-as-a-judge?
A method where a model scores generated responses against a rubric or picks the better of two responses, used when there is no single correct answer.
Do I need to know Kubeflow Pipelines?
Yes at a conceptual level. The guide lists Kubeflow Pipelines with Experiments and Agent Platform Pipelines as experiment environments.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 2 knowledge
You have reviewed preprocessing tools, Feature Store, notebook security and experiment tracking. Try exam-style questions to check that you pick the right tool for each data size.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the GCP PMLE AI tutor Get any concept explained your way