Read this the night before or on exam morning. It assumes you’ve studied the notes. It’s here to refresh, not to teach.
Exam day
- Standard exam: about 40 to 50 multiple choice and multiple select questions in 2 hours. Fee $200. English and Japanese. Valid 2 years. No prerequisites; 3+ years of industry experience, 1+ on Google Cloud is recommended.
- Google reports pass or fail rather than a published passing score. Read every option; “choose TWO” needs two different pieces of one solution.
- Domain weights: designing data processing systems 22%, ingesting and processing 25%, storing data 20%, preparing and using data for analysis 15%, maintaining and automating workloads 18%.
- About 2.5 minutes per question. Flag long scenarios and return.
How to read a PDE question
| The question says | Pick |
|---|---|
| Minimal operations, serverless, fully managed | Dataflow, BigQuery, Pub/Sub, Cloud Run, Workflows; not self-managed clusters |
| Existing Spark or Hadoop jobs, no code changes | Dataproc (ephemeral clusters, Cloud Storage) |
| No-code, visual ETL | Cloud Data Fusion |
| SQL transformations with tests and version control | Dataform |
| Near real time, streaming | Pub/Sub plus Dataflow; BigQuery Storage Write API |
| Millisecond lookups at huge scale by key | Bigtable |
| Global, strongly consistent, relational | Spanner |
| Keys must stay with us / outside Google | CMEK / Cloud EKM |
| Share without copying | Analytics Hub, authorized views |
| Predictable cost for heavy BigQuery use | Editions with reservations |
| Most cost-effective | Cheapest option that still meets the stated requirement: ephemeral clusters, Spot, lifecycle classes, batch loads |
Elimination tricks
- Answers that need you to manage VMs, patch servers or write cron jobs are usually wrong when a managed service exists.
- Options that violate the hard constraint (residency, ACID, latency) are out whatever else they offer.
- BigQuery is never the answer for OLTP; Bigtable is never the answer for joins or multi-row transactions.
LIMITdoes not reduce bytes scanned; sharded tables are an anti-pattern.
Service pickers
| Need | Choose |
|---|---|
| Analytics warehouse | BigQuery |
| Lake files queried by many engines with fine-grained security | BigLake, Iceberg |
| Objects, raw landing, archive | Cloud Storage (Standard, Nearline 30d, Coldline 90d, Archive 365d) |
| Time series and IoT | Bigtable |
| Cache | Memorystore |
| Postgres with analytics | AlloyDB |
| Lift and shift MySQL, Postgres, SQL Server | Cloud SQL |
| Documents for apps | Firestore |
| Orchestrate Airflow DAGs | Composer |
| Simple API chaining | Workflows |
| Cron trigger | Cloud Scheduler |
| Catalog, quality, lineage | Dataplex |
Migration tools
| Need | Tool |
|---|---|
| Scheduled SaaS and warehouse loads into BigQuery | BigQuery Data Transfer Service |
| Database to Cloud SQL or AlloyDB with low downtime | Database Migration Service |
| Continuous CDC into BigQuery or Cloud Storage | Datastream |
| Objects from S3, Azure or on-prem | Storage Transfer Service |
| Hundreds of TB, slow network | Transfer Appliance |
| Teradata, Oracle SQL dialects | BigQuery SQL translator |
Streaming facts
- Pub/Sub: at-least-once by default; exactly-once delivery on pull subscriptions; ack deadline 10 s default (10 to 600); retention 7 days default, up to 31; ordering by key; dead-letter topic after N attempts; BigQuery and Cloud Storage subscriptions need no Dataflow.
- Beam windows: fixed, sliding, session (gap), global. Watermark estimates completeness; allowed lateness keeps windows open; data later than that is dropped. Accumulating vs discarding modes.
- Dataflow: drain finishes in-flight data; cancel drops it; update replaces a compatible running job. Streaming Engine, FlexRS, Flex Templates, hot-key fan-out.
- Load jobs are free; Storage Write API is the streaming path; legacy
insertAllis best effort.
BigQuery facts
- Partition: up to 10,000 partitions; time column, ingestion time or integer range. Cluster: up to 4 columns. Require partition filter.
- Time travel 7 days default (2 to 7); fail-safe 7 more; snapshots for longer; long-term storage after 90 days unmodified.
- Result cache 24 hours. Materialized views refresh incrementally; BI Engine accelerates dashboards in memory.
- Editions: Standard, Enterprise, Enterprise Plus; baseline plus autoscale slots; assignments by project, folder, organization and job type; batch queries wait for idle slots.
- Security: dataset IAM, authorized views, row access policies, policy tags (Fine-Grained Reader), dynamic masking (Masked Reader), CMEK, VPC Service Controls.
- Roles:
jobUserruns queries,dataViewerreads data. - Constraints (primary and foreign keys) are not enforced.
Storage design facts
- Bigtable: avoid timestamp-first keys; promote a field, salt, reverse; multi-cluster routing is eventually consistent; keep CPU under about 70%; Key Visualizer finds hotspots.
- Spanner: avoid monotonic keys; interleave; multi-region 99.999%.
- Cloud SQL: HA standby plus read replicas; cross-region replica for DR.
- Cloud Storage: Autoclass, lifecycle, Bucket Lock, dual-region with turbo replication (15 minutes).
Maintenance facts
- Prefer ephemeral Dataproc, Spot secondary workers, Cloud Storage not HDFS.
- Composer: idempotent tasks, logical date, retries,
trigger_rule='all_done'on cluster deletion. - Alert on Pub/Sub oldest unacked message age and Dataflow system lag.
- Quota errors: back off, batch, reservations, request an increase. Cost spikes:
INFORMATION_SCHEMA.JOBS, billing export, maximum bytes billed, budgets. - Recovery: time travel, snapshots, versioning, replicas, replay from raw data.