Google Professional Data Engineer: Current Exam Scope

Which Professional Data Engineer exam is current in 2026?

Google Cloud offers the Professional Data Engineer certification for professionals who collect, transform, store, deliver and operate data for business applications. The current issuer page distinguishes the standard examination from several renewal paths. A first-time candidate, or someone whose credential has expired, must take the standard exam. An eligible holder of an active certification may be able to renew through the standard examination, a shorter renewal examination or designated Google Skills work. Those routes are not interchangeable. A guide that quotes the renewal exam’s duration or domains as if they applied to every applicant could mislead a reader before any study begins.

As published on Google’s certification page, the standard Professional Data Engineer exam lasts two hours, includes approximately 40–50 multiple-choice and multiple-select questions, is available in English and Japanese, and costs US$200 before applicable taxes. Google states that there are no formal prerequisites, though it recommends at least three years of industry experience including more than a year working with Google Cloud solutions. The credential is valid for two years. A candidate should check the live issuer page again before booking because scheduling availability, taxes and administrative rules may change.

The current standard guide has five domains with approximate weights of 22, 25, 20, 15 and 18 percent. Its detailed topics include governance, migration, event-time streaming, analytical storage, AI/ML preparation, orchestration and cost management. These topics apply to the standard assessment; the separate renewal guide has a different structure and should not be substituted when planning an initial certification. The concept of Professional Data Engineer objectives should be read against the standard issuer document that is linked from Google’s current certification page.

This is an independent ExamLabs explanation of that public issuer material, not a Google-issued exam guide or a claim of access to confidential items. The purpose is to clarify current scope and avoid dated assertions about what an exam ‘always’ contains. Individual questions can cross domain boundaries, and the objective guide provides examples rather than an exhaustive list of every possible technology. Preparation should prioritize understanding the relationships among requirements and technologies.

Designing data processing systems: roughly 22 percent

The design domain tests whether a proposed solution meets business, security, compliance, reliability and portability constraints. IAM, organization policies, encryption, key management and personal-data handling belong here. Region selection and data sovereignty can constrain which projects, datasets and tables are permitted. A system that delivers excellent analytical response times but transfers restricted information to a prohibited region is not a correct design. Candidates should practice identifying the hard constraint that rules out an otherwise plausible service choice.

Reliability and fidelity cover cleaning and validation, monitoring and orchestration, disaster recovery, fault tolerance, and ACID versus availability decisions. Consider a migration from an on-premises booking database: historical data must be imported, continuing writes must be captured, and the final results must reconcile. A team may use managed transfer tools, Datastream or database migration capabilities, but the right option depends on source technology, downtime and change rate. The objective also includes migration planning and stakeholder needs, not only the command that starts a transfer.

Flexibility means planning for future requirements without creating needless complexity now. Data may need to remain portable across providers or regions, but that does not mean every application should be designed as a fully abstract multicloud platform. Cataloging, profiling and discovery are part of governance because they make data assets understandable to authorized consumers. Design questions should lead with ownership, trust and constraints, not with the largest number of deployable Google Cloud services.

A supporting explanation of Cloud SQL is relevant when the design needs a conventional managed relational database; it is not automatically the best answer to a globally distributed consistency requirement or a massive analytical scan. The scope includes choosing among competing data models and failure assumptions before choosing an implementation.

Ingestion and storage: approximately 25 and 20 percent

The ingestion-and-processing domain covers source and sink identification, transformation logic, networking and encryption, and building batch or streaming pipelines. Google lists processing services and frameworks such as Dataflow, Apache Beam, Dataproc, Spark, Hadoop, BigQuery, Pub/Sub, Kafka and Cloud Data Fusion as examples. No candidate should infer that one named framework must appear in every question. The decision turns on the source format, latency requirement, existing code, operational expertise and output contract.

Streaming includes event time, windowing and late-arriving data. A sensor that reconnects after an outage can send events that belong to an earlier time window. A correct system must decide whether a previously published aggregate can be revised and how duplicates are recognized. Batch jobs face their own partial-failure problems: if a job wrote half a partition before retry, rerunning without a deduplication or merge policy can corrupt the output. Deployment and operationalization also include CI/CD and orchestration, so the candidate should know how a data pipeline becomes repeatable rather than remaining a successful one-time script.

Storage selection is about access patterns and lifecycle. The standard guide explicitly lists BigQuery, BigLake, AlloyDB, Bigtable, Spanner, Cloud SQL, Cloud Storage, Firestore and Memorystore among possible managed services. A conventional relational order system, global strongly consistent ledger, large warehouse query and rapid row-key lookup have different needs. A discussion of Spanner can illustrate distributed relational consistency, but the requirement must justify its greater complexity; size alone is not enough.

The storage domain also addresses warehouse modeling, normalization decisions, data lake monitoring and governance, and platform architecture with Dataplex or catalog capabilities. A Cloud Storage bucket may be durable but still unusable as a trusted analytical product when nobody knows its schema or classification. The exam is testing selection and stewardship together. The correct design should say how consumers find and use the dataset, what data remains retained, and how the system prevents unauthorized or accidental exposure.

Analysis preparation: approximately 15 percent

Google’s standard guide explicitly includes preparing data for visualization, calculated fields, materialized views, BI Engine, query-performance troubleshooting, security, masking and Data Loss Prevention. An analytical query that runs quickly but omits late data may be incorrect for a finance report. A report that has the right total but exposes customer identifiers can violate its access contract. Prepare to identify which layer is responsible for the defect before proposing more compute capacity or another transformation.

The domain also includes preparing features for AI and machine learning, including BigQuery ML, and preparing unstructured inputs for embeddings and retrieval-augmented generation. The data engineer’s responsibility is to provide well-governed, correctly transformed and versioned data. That differs from claiming every data engineer must build and train a foundation model. A RAG pipeline must respect source-document access and handle updates or removal of content that should no longer be retrievable. This broadening of the exam reflects the growing variety of downstream consumers of curated data.

Sharing is a separate assessed consideration. Publishing datasets, analytical reports and visualizations requires access rules, trustworthy metadata and a known refresh contract. BigQuery sharing and Analytics Hub may help distribute approved data assets, but a sharing mechanism does not itself authorize a department to see protected records. The candidate should explain how policy, masking, data classification and auditing fit together. The topic becomes more concrete through BigQuery analytics when an example concerns a query’s cost, governance or consumer-facing model.

Distinguish data freshness from pipeline uptime. A pipeline can be green while a report is missing a source region. If downstream consumers depend on data quality, the publication process needs completeness checks and a way to mark an invalid or stale output. That is why visualization, AI preparation and sharing cannot be studied independently from ingestion and operational monitoring.

Maintaining and automating workloads: approximately 18 percent

The operations domain covers resource optimization, repeatable automation, workload organization, monitoring, troubleshooting and fault mitigation. Cloud Composer DAGs and service orchestration are examples, but the core issue is whether dependencies, retries and partial commits are handled safely. A good orchestration design prevents a failed upstream validation from publishing a false daily metric. It also records which output can be trusted during recovery. An ‘all tasks completed’ result has limited meaning without a business-data completeness signal.

BigQuery editions and reservations, interactive and batch jobs, and the choice between persistent and job-based Dataproc clusters illustrate cost and latency decisions. An always-on cluster can reduce startup delay while wasting resources when unused. A job-based cluster can lower idle cost but add provisioning time and operational complexity. Cost management requires measurable service obligations, not the assumption that the least expensive monthly estimate always wins. Query optimization should preserve correctness before any gain is called successful.

Google includes troubleshooting logs, quotas, error messages, billing anomalies and resource management. The candidate should identify the evidence that distinguishes a permissions failure, exceeded quota, missing data, corrupt input and slow query. Cloud Monitoring and Cloud Logging can be useful, but business data quality often needs custom indicators. A missing expected input file may be an incident even when infrastructure metrics appear healthy. Recovery strategy should address restart semantics, regions or zones and replication where warranted.

For current readiness, use the issuer’s standard guide line by line and support each objective with an example of a decision or verified test. A separate study sequence can help prioritize dependencies; it must not be mistaken for the official assessment structure. The current certification remains an assessment of engineering judgment across an entire data life cycle, not a product-identification quiz.