ADP question patterns

BigQuery, Dataflow, Dataform, Workflows or a scheduled query rather than clusters or custom scripts.

Anything you have to size, patch or schedule yourself is less managed.

BigQuery SQL, Dataform for ELT, BigQuery ML for models.

Moving data out to another engine adds cost and complexity.

A predefined role such as BigQuery Data Viewer on the dataset, not Editor or Owner on the project.

Basic roles are broad and span many services.

Publish a listing in a data exchange; the subscriber gets a linked dataset.

One copy stays with the publisher, and access stays controlled.

Order the appliance, load it on site and ship it back.

Uploading over a slow link would take too long.

Create a scheduled transfer job (with an agent for on-premises sources).

It is managed, handles retries and checks integrity.

Set up a transfer configuration with a schedule.

Built-in connectors need no pipeline code.

Change data capture from the source into BigQuery or Cloud Storage.

It streams inserts, updates and deletes without custom code.

Continuous migration, then cut over.

It is the purpose-built migration tool.

bq load or scheduled load from Cloud Storage.

Load jobs are free; streaming inserts are billed.

Create a subscription that writes to the table.

No Dataflow job is needed.

Avro for row-based loads and streaming, Parquet for columnar analytics.

Both embed the schema; CSV and JSON have no types.

Transform with Dataflow or Data Fusion before loading.

ELT would land the raw sensitive values first.

Wide-column store keyed by row key.

Not a SQL warehouse; built for high-throughput key access.

Multi-region Spanner instance.

Cloud SQL scales vertically and by read replica only.

Managed Cloud SQL instance.

Lift-and-shift of a conventional relational app.

Document database.

Schema flexible, built for app backends.

Create the bucket in a dual-region location, optionally with turbo replication.

Multi-region doesn't let you pick the regions.

Partition by date, cluster by filter columns, select only needed columns.

Bytes scanned drives cost; LIMIT does not reduce it.

Use the validator or a dry run to see bytes processed.

It runs nothing and costs nothing.

Create a materialized view or summary table.

Results are precomputed and refreshed automatically.

Open a notebook in BigQuery Studio.

Managed Jupyter-style notebooks with IAM.

Connect BigQuery and build a report.

No modelling layer needed.

Model with LookML dimensions and measures.

A semantic layer keeps metrics consistent.

CREATE MODEL with logistic regression or boosted trees, ML.EVALUATE, ML.PREDICT.

No data export or ML code needed.

BigQuery ML k-means clustering.

Unlabelled data means unsupervised learning.

BigQuery ML time series model and ML.FORECAST.

It handles seasonality and holidays.

Create a connection, grant its service account access, create a remote model and call ML.GENERATE_TEXT.

No training and no data export.

Apache Beam pipeline or a Google template.

Serverless batch and streaming.

Run them on Dataproc (Managed Service for Apache Spark).

Dataflow needs Beam code.

Build the pipeline visually.

Low-code, prebuilt connectors.

SQLX definitions with assertions and a schedule.

ELT workflow inside BigQuery.

BigQuery scheduled query with an overwrite destination.

Composer would be overkill.

An Airflow DAG.

Built for complex multi-step orchestration.

Trigger a workflow from Cloud Scheduler.

No Airflow environment to run.

Eventarc trigger to Cloud Run, a Cloud Run function or Workflows.

Event-driven beats polling on a schedule.

Inspect the job graph, stage metrics and logs.

It shows per-stage lag and throughput.

Alerting policy on job or log-based metrics with a notification channel.

Logs alone don't notify anyone.

SetStorageClass and Delete conditions on object age.

Automatic, no scripts.

Archive storage with a retention policy.

Lowest storage price; 365-day minimum fits.

Set partition expiration on a date-partitioned table.

Old partitions are deleted automatically.

Enable it and disable ACLs.

One consistent, auditable model.

Generate a signed URL.

No Google account or IAM grant required.

Enable HA with a standby in another zone.

Automatic failover; but it does not replace backups.

Enable automated backups and PITR.

Replicas and HA copy the mistake.

Encrypt with a key in Cloud KMS.

Disabling the key makes data unreadable.

Customer-supplied keys, for Cloud Storage or Compute Engine disks only.

Google never stores the key; a lost key means lost data.

Inspect and de-identify with DLP.

Purpose-built for personal data.