Difficulty
Calculator
Interactive interview prep · 30 questions across the JD's stack — BigQuery, Dataflow/Beam, Dataproc/Spark, Pub/Sub, Cloud Storage, Composer & CI/CD, data modeling (SCD, Data Vault, dimensional), quality & lineage, and security/governance (IAM, VPC-SC, CMEK, row/column-level security) — with detailed explanations.
Press Check answer on each question to see whether you got it right — every check is timestamped and recorded in your Activity log below.
Each attempt is saved under its own name — load, rename, or delete them below. Difficulty / calculator tags are shared across all attempts. Numeric answers accept equivalent forms (e.g. 4/3 or 1.333).
The timer is stopped. Questions are hidden until you resume.
Medium Under BigQuery on-demand pricing, query cost is driven primarily by:
Answer: D — The number of bytes the query scans/processes
Medium A 50 TB events table is almost always filtered by event_date. The most effective way to cut scan cost is to:
Answer: B — Partition the table by event_date so queries prune to only the needed partitions
Medium Clustering a BigQuery table:
Answer: C — Sorts/co-locates rows by the clustering columns so filters and aggregations on them scan fewer blocks
Medium Why is SELECT * an anti-pattern in BigQuery?
Answer: C — BigQuery is columnar, so it reads every column's bytes — select only the columns you need to cut cost/scan
Medium A platform team needs predictable, capped monthly BigQuery spend for heavy workloads. The right approach is:
Answer: D — Capacity-based pricing — slot reservations / BigQuery Editions
Medium For analytical tables, BigQuery generally favors:
Answer: A — Denormalization with nested & repeated fields (STRUCT/ARRAY) over many joins
Medium The core value of Apache Beam (and Dataflow) is:
Answer: B — One unified programming model for both batch and streaming pipelines
Hard In a Beam streaming pipeline, a watermark is:
Answer: C — The system's estimate of how far event-time has progressed — used to close windows and reason about late data
Medium Which windowing strategy groups events by periods of inactivity (gaps) rather than fixed clock intervals?
Answer: D — Session windows
Medium Enabling Dataflow's Streaming Engine / Dataflow Shuffle primarily:
Answer: B — Moves shuffle and state off the worker VMs, giving smoother autoscaling and lower worker resource use
Medium You'd choose Dataproc over Dataflow mainly when:
Answer: C — You have existing Spark/Hadoop/Hive workloads or need the Spark ecosystem specifically
Medium Which choices reduce cost on a Dataproc Spark job? Select all that apply.
Answer: B, C, D
Medium By default, Pub/Sub delivery is:
Answer: C — At-least-once — duplicates are possible, so consumers should be idempotent (use ordering keys if order matters)
Medium To stop a repeatedly-failing (poison) message from blocking a Pub/Sub subscription forever, you configure:
Answer: C — A dead-letter topic with a max delivery-attempts threshold
Medium The canonical GCP streaming ingestion-to-warehouse pattern is:
Answer: C — Pub/Sub → Dataflow → BigQuery
Medium Cold, rarely-accessed data (queried maybe once a year) that must stay cheap to store belongs in which Cloud Storage class, managed via:
Answer: A — Archive class, transitioned automatically with Object Lifecycle Management rules
Medium Cloud Composer is:
Answer: A — Managed Apache Airflow for authoring and scheduling DAG-based orchestration
Medium Why should an Airflow/Composer task be idempotent?
Answer: A — So retries and backfills re-run safely without duplicating or corrupting data
Medium Cloud Workflows vs Cloud Composer — the typical split is:
Answer: C — Workflows = lightweight serverless sequencing of APIs/services; Composer = complex data DAGs with dependencies, scheduling, backfills
Medium A reusable, governed way to deploy pipeline/DAG changes is:
Answer: B — Git-based workflows with Cloud Build CI/CD — automated tests, then deploy DAGs/pipelines
Medium A Slowly Changing Dimension Type 2:
Answer: D — Preserves history by inserting a new row (with effective dates / a current flag) when an attribute changes
Medium In Data Vault modeling, the three core structures are:
Answer: D — Hubs (business keys), Links (relationships), Satellites (descriptive/historical context)
Medium A star schema consists of:
Answer: C — A central fact table referencing surrounding dimension tables
Medium Which are good automated data-quality checks to build into a pipeline framework? Select all that apply.
Answer: A, B, C, D
Medium For cataloging metadata, data discovery, and lineage across the GCP data estate, you'd use:
Answer: D — Dataplex / Data Catalog
Hard VPC Service Controls (VPC-SC) primarily protect against:
Answer: C — Data exfiltration — a service perimeter keeps data in managed APIs (e.g., BigQuery, GCS) from leaving to unauthorized projects/networks
Medium CMEK (Customer-Managed Encryption Keys) means:
Answer: B — You manage the encryption keys in Cloud KMS (rotation, disable, destroy) instead of using Google-managed default keys
Hard To stop most analysts from querying PII columns (e.g., SSN) in a shared BigQuery table while others can, you use:
Answer: C — Column-level security via policy tags (Dataplex/Data Catalog taxonomy)
Medium An IAM best practice for a data platform is:
Answer: C — Least privilege — grant narrow predefined/custom roles to service accounts, avoiding broad primitive roles like Owner/Editor
Medium The platform team's mandate is self-service and consistency. The right strategy is to:
Answer: D — Provide reusable, config-driven frameworks, blueprints, and templates that domain teams adopt — not hand-built one-off pipelines per use case