GCP Professional Data Engineer (PDE) - Exam Strategy¶

Format reminder¶

• 50-60 questions, 120 minutes
• Pass mark ~70-75% (not officially published)
• Multiple choice + multiple response
• Includes some case studies (Flowlogistic, MJTelco, etc.)

Top traps¶

1. Dataflow vs Dataproc vs Dataflow Prime:
2. Dataflow: serverless Apache Beam, the GCP-native streaming + batch
3. Dataproc: managed Hadoop/Spark on GCE clusters
4. Dataproc Serverless: serverless Spark

5. Dataflow Prime: enhanced Dataflow with vertical autoscaling PDE leans toward Dataflow for streaming.

6. BigQuery on-demand vs Editions:

7. On-demand: pay per TB scanned (100 GB free tier)

8. Editions (Standard / Enterprise / Enterprise Plus): slot reservations with autoscaling Switch when monthly spend exceeds break-even (~~$300-500).

9. Storage classes: Standard / Nearline (>30d) / Coldline (>90d) / Archive (>365d). Lifecycle policies for transitions.

10. Streaming inserts vs Storage Write API: Storage Write API is the modern recommendation - exactly-once, higher throughput, lower cost.

11. Partitioning + clustering in BigQuery:

12. Partitioning: divides table by date, integer range, or ingestion time

13. Clustering: sorts within partitions by up to 4 columns Both reduce data scanned (cost + speed).

14. Materialized views vs scheduled queries: MVs auto-refresh, queryable as tables. Scheduled queries write results to a destination on a cadence. Different use cases.

15. Pub/Sub vs Pub/Sub Lite:

16. Pub/Sub: global, autoscaling, pay per message

17. Pub/Sub Lite: regional/zonal, provisioned capacity, cheaper for high steady throughput Pub/Sub is the default modern answer.

18. Bigtable vs Spanner vs Firestore:

19. Bigtable: wide-column NoSQL, single-key access at petabyte scale (HBase API), no SQL
20. Spanner: global SQL, ACID

21. Firestore: document, mobile/web sync Don't confuse.

22. Apache Beam concepts: PCollection, PTransform, ParDo, GroupByKey, CoGroupByKey, windowing (fixed, sliding, session), watermarks, triggers. Tested often in PDE.

23. Dataform vs dbt: Dataform is GCP-native ELT (acquired into BigQuery). dbt is the third-party. PDE leans toward Dataform.

High-yield topics easy to miss¶

• BigLake tables (BigQuery query Cloud Storage / Iceberg / Delta)
• BigQuery Omni (BQ on AWS, Azure)
• BigQuery ML (in-warehouse ML)
• Dataflow templates (provided + custom)
• Dataform repositories and version control
• Data Catalog policy tags + taxonomies
• Sensitive Data Protection (formerly DLP) for PII discovery / de-id
• Vertex AI Feature Store + Pipelines
• Cloud Composer (Airflow) operator library

Time management¶

120 / ~55 = ~2.2 min/question. Pace: half done by minute 60. Leave 15 min for review.

When stuck¶

1. Identify whether the workload is batch or streaming - filters compute and storage choices.
2. Match data shape to storage: SQL relational → Cloud SQL / Spanner; analytics → BigQuery; key-value at scale → Bigtable; document → Firestore.
3. Default to managed > self-managed - Dataflow over self-Spark.
4. Eliminate "use Cloud Functions" for high-throughput - Functions is for low-volume glue.

Day-of logistics¶

120 min, ~55 questions. Bring two IDs.

After¶

Pass: Cert valid 2 years.

Fail: Most failures are on Pipeline Design (~25%) or Data Storage (~25%). Re-review Beam concepts, BigQuery cost optimization, and SCD patterns.

PDE patterns¶

• "Streaming with dedup + windowing" = Pub/Sub → Dataflow → BigQuery
• "Real-time analytics dashboard" = Pub/Sub → Dataflow → BigQuery (Storage Write API) → Looker
• "Batch ETL with retries + dependencies" = Cloud Composer
• "Cost-optimize BigQuery" = Editions reservations + partitioning + clustering + MVs
• "Globally consistent SQL transactional" = Spanner
• "Wide-column NoSQL at scale" = Bigtable
• "Sensitive data column-level access" = Data Catalog policy tags + BQ column ACL
• "Schema-on-read lake" = Cloud Storage + BigLake
• "Feature consistency train/serve" = Vertex AI Feature Store
• "ELT in BigQuery with version control" = Dataform