CCNA Data Preparation And Ingestion Questions — Page 2 of 2

MCQhard

You are using Dataflow with autoscaling. What metric determines whether the worker pool scales up or down?

A.Work backlog and throughput

B.Network latency

C.Memory usage

D.User-defined triggers

AnswerA

Dataflow autoscaling evaluates the ratio of work backlog to throughput, scaling workers when backlog per worker exceeds the target. This directly satisfies the stem's constraint: autoscaling responds to outstanding work versus processing rate, not CPU or memory utilisation, which Dataflow does not use for horizontal worker scaling decisions.

Why this answer

Dataflow autoscaling is driven by the work backlog (unprocessed elements) and throughput relative to the current worker pool. The autoscaler compares the backlog and the rate of processing to decide whether to add or remove workers, targeting a steady-state where backlog stays low without over-provisioning. This is the documented signal for Streaming Engine and batch autoscaling.

Exam trap

GCP-ADP often tests whether candidates know Dataflow autoscaling is backlog/throughput-driven, not resource-metric-driven — the trap is picking memory or CPU as the scaling signal.

How to eliminate wrong answers

Option B is wrong because network latency is not an autoscaling signal — Dataflow scales on work volume, not network conditions. Option C is wrong because memory usage is a per-worker resource metric that can trigger OOM restarts but is not the primary autoscaling input; the autoscaler looks at backlog and throughput. Option D is wrong because user-defined triggers are a Beam windowing concept for emitting results, not a Dataflow worker-scaling mechanism.

Practice this question →

MCQhard

You need to ingest IoT data into BigQuery. You need to handle messages arriving in millions per second. Which service acts as the buffer?

A.Cloud SQL

B.Pub/Sub

C.Cloud Functions

D.BigQuery Streaming API

AnswerB

Pub/Sub decouples ingestion from processing by durably buffering messages across distributed partitions, absorbing millions of messages per second without backpressure. This satisfies the stem's buffering constraint: publishers write independently of BigQuery load, letting subscribers drain at their own pace and preventing data loss during traffic spikes.

Why this answer

Pub/Sub is Google Cloud's fully managed, globally scalable messaging service designed to decouple producers from consumers, and it can absorb millions of messages per second without backpressure. In an IoT ingestion architecture, Pub/Sub acts as the durable buffer that holds incoming telemetry until BigQuery (via Dataflow or the Storage Write API) can commit it. This decoupling prevents data loss during BigQuery load spikes and lets each side scale independently.

Exam trap

GCP-ADP often tests the difference between a buffer/ingestion layer and a processing or storage layer — candidates pick BigQuery Streaming API or Cloud Functions because they 'handle' the data, missing that only Pub/Sub provides durable, high-throughput buffering.

How to eliminate wrong answers

Option A is wrong because Cloud SQL is a relational OLTP database with strict connection limits and single-digit-thousand TPS ceilings, so it cannot buffer millions of messages per second. Option C is wrong because Cloud Functions is a stateless compute runtime with a 10-minute max timeout and per-instance concurrency limits — it processes events but does not durably buffer them. Option D is wrong because the BigQuery Streaming API is a write endpoint, not a buffer; it applies per-table streaming quotas and will reject or throttle traffic if the sink cannot keep up.

Practice this question →

MCQmedium

A financial analytics team ingests transaction events into Pub/Sub, then uses a Dataflow streaming pipeline with a GroupByKey transform to aggregate per-account totals. The team reports that late-arriving events—sometimes hours behind due to mobile connectivity—are being dropped and never counted in the aggregates. You need the pipeline to include those late events in the correct window. What should you configure?

A.Increase the Pub/Sub message retention duration to 7 days on the topic.

B.Switch the pipeline from streaming mode to batch mode with a longer maximum worker count.

C.Set a longer allowed lateness on the windowing strategy and emit late panes from the pipeline.

D.Add a Reshuffle transform after the GroupByKey to redistribute late records across workers.

AnswerC

Allowed lateness extends how long a window's state is retained after the watermark passes the window end. Setting it lets late events still be assigned to their correct window and emitted as late panes, so the mobile events arriving hours behind are counted. This is the intended Dataflow mechanism for bounded late data, and the team must also decide how to handle late panes downstream.

Why this answer

Allowed lateness is the Dataflow setting that governs how long window state persists after the watermark passes the window end. Extending it lets events arriving hours behind be assigned to the correct window and emitted as late panes. Pub/Sub retention, batch mode, and Reshuffle operate at different layers and do not change event-time windowing semantics.

Exam trap

The trap here is assuming that increasing Pub/Sub message retention will fix dropped late events, when the loss actually occurs in the Dataflow windowing layer governed by allowed lateness.

Practice this question →

Multi-Selecthard

Which THREE of the following are Dataflow windowing types?

Select 3 answers

A.Fixed

B.Session

C.Binary

D.Sliding

E.Random

AnswersA, B, D

Fixed windows partition unbounded data into equal, non-overlapping intervals based on event time, satisfying the stem's requirement for a genuine Dataflow windowing type. Apache Beam, which underpins Dataflow, defines fixed, sliding, and session windows as its core built-in types, so Fixed qualifies directly.

Why this answer

In Google Cloud Dataflow (Apache Beam), the supported windowing types are fixed windows, sliding windows, and session windows, so options A, B, and D are correct. Fixed windows (A) divide a PCollection into non-overlapping, contiguous time intervals of a set duration, such as every 60 seconds. Session windows (B) group elements that arrive close together in time, separated by a gap duration of inactivity, which is useful for user-activity streams.

Sliding windows (D) are defined by a window duration and a slide period, allowing windows to overlap, such as a 60-second window that slides every 30 seconds. Binary (C) and Random (E) are not Beam/Dataflow windowing types; no such windowing functions exist in the Dataflow model.

Exam trap

The trap is confusing windowing types with other Beam concepts (e.g., triggers, side inputs) or inventing non-existent types like 'Binary' or 'Random' — candidates must recall the three canonical windowing types from Apache Beam.

Practice this question →

Multi-Selecthard

Which THREE steps are involved in an effective Dataflow pipeline development?

Select 3 answers

A.Deployment

B.Testing

C.Automated email alerts

D.Development

E.Manual data entry

AnswersA, B, D

Deployment is a required pipeline development step because a Dataflow pipeline must be built, tested and then released to production. Packaging the Apache Beam code and submitting it to the Dataflow service, whether via template or direct runner, satisfies the stem's demand for an end-to-end development lifecycle rather than coding alone.

Why this answer

The three correct steps are Deployment (A), Testing (B), and Development (D). Development (D) is where you author the pipeline code using the Apache Beam SDK, defining PTransforms, sources, and sinks before anything can be run. Testing (B) is essential to validate pipeline logic and correctness, typically via unit tests with TestPipeline and DoFnTester, and integration tests against representative data before production.

Deployment (A) is the step of running the pipeline on a runner such as Dataflow, whether via the gcloud command, the Dataflow template, or a CI/CD job, to execute it at scale. Automated email alerts (C) are an operational monitoring concern, not a core pipeline development step, and manual data entry (E) is not part of building a Dataflow pipeline at all.

Exam trap

GCP-ADP often tests whether candidates confuse operational concerns (alerts, monitoring) with the core development lifecycle steps — email alerts and manual entry are distractors that sound plausible but are not development phases.

Practice this question →

MCQhard

You need to ingest data from an external API that requires an OAuth token. Which tool is best for executing this periodically?

A.Pub/Sub

B.Cloud Functions

C.BigQuery Data Transfer Service

D.Storage Transfer Service

AnswerB

Cloud Functions suits this because it executes event-driven, serverless code on a schedule via Cloud Scheduler, letting you fetch a fresh OAuth token and call the external API each run. It satisfies the periodic execution constraint without managing servers, unlike batch-oriented Dataflow or orchestration-heavy Composer.

Why this answer

Cloud Functions is the best fit because it can run scheduled (via Cloud Scheduler), execute custom code to perform the OAuth token exchange, call the external API, and write results to a destination like BigQuery or Cloud Storage. It supports secrets via Secret Manager and scales automatically. This makes it ideal for periodic, code-driven API ingestion requiring authentication.

Exam trap

GCP-ADP often tests the misconception that Pub/Sub or Data Transfer Service can directly call arbitrary OAuth-protected APIs, when in fact you need a compute service like Cloud Functions to execute the authenticated request.

How to eliminate wrong answers

Option A is wrong because Pub/Sub is a messaging service, not an execution environment; it cannot itself perform OAuth token exchange or call an external REST API without a subscriber (e.g., a Cloud Function or Dataflow job) doing the work. Option C is wrong because BigQuery Data Transfer Service supports specific connectors (Google Ads, S3, Teradata, etc.) and cannot generically call an arbitrary OAuth-protected external API. Option D is wrong because Storage Transfer Service moves files between object stores (S3, GCS, HTTP) and does not handle OAuth-authenticated REST API calls or custom transformation logic.

Practice this question →

MCQmedium

You are using Dataprep to prepare data. How can you share your work with a team?

A.Copy the JSON definition

B.Share the Flow

C.Share the underlying GCS file

D.Export to a CSV

AnswerB

A Flow encapsulates the entire recipe, including transformations and inputs, so sharing it gives teammates the complete reproducible pipeline rather than just output data. This directly satisfies the requirement to share the preparation work itself.

Why this answer

In Google Cloud Dataprep, a Flow is the primary container that holds all your recipes, datasets, and transformation logic. Sharing the Flow directly with team members grants them access to the entire pipeline, allowing them to view, edit, and run the same data preparation steps. This is the built-in collaboration mechanism designed for team workflows, and it respects IAM permissions on the underlying project and datasets.

Exam trap

GCP-ADP often tests the distinction between sharing the transformation logic (Flow) versus sharing raw data or output, so candidates may incorrectly choose to share the underlying GCS file or export to CSV, missing that Dataprep's collaboration feature is specifically the Flow.

How to eliminate wrong answers

Option A is wrong because copying the JSON definition only gives a static snapshot of the Flow's configuration; it does not provide a live, collaborative sharing mechanism and requires manual import, which can lead to version drift. Option C is wrong because sharing the underlying GCS file only shares the raw data, not the Dataprep transformation logic, recipes, or Flow structure—team members would have no access to the preparation steps. Option D is wrong because exporting to CSV only produces the transformed output data, not the reusable Flow definition or recipes, so it cannot be used to share the actual work or enable further editing.

Practice this question →