Free PROFESSIONAL-DATA-ENGINEER Practice Test Questions and Answers (2026)

This solution proposes a denormalized, append-only model, which is the most idiomatic and cost-effective approach for data warehousing in BigQuery. Denormalization using nested and repeated fields minimizes the need for expensive JOIN operations, simplifying queries and improving performance. An append-only strategy, where new versions of records are added instead of updating existing ones, naturally preserves a complete history. Using an ingestion timestamp allows for easy point-in-time analysis, enabling the visualization layer to query for either the most current state of the data or its state at any point in the past. This directly addresses all requirements: maintaining history, simplicity, ease of use, and cost-effectiveness.

A. A normalized model in BigQuery often leads to expensive JOINs, which contradicts the cost-effective and simple-to-use requirements. Using snapshots for historical analysis is complex for tracking an entity's changes over time.

B. A normalized model is suboptimal for BigQuery. Keeping historical data only in Cloud Storage files makes it unavailable for direct querying from the data warehouse, failing the reporting requirement.

C. While denormalization is correct, updating the table in place (in-place DML UPDATEs) overwrites and destroys historical data within the table itself. Relying on snapshots for history is inefficient for reporting.

1. Google Cloud Documentation, BigQuery Best Practices: Schema Design. In the section "Denormalization," it states, "The recommended approach to denormalizing data in BigQuery is to use nested and repeated fields. This approach maintains the relationships between entities while co-locating the data for optimal query performance." This supports the use of a denormalized model with nested and repeated fields.

2. Google Cloud Documentation, Data warehouse design for BigQuery, "Handle data changes". This section describes implementing Slowly Changing Dimensions (SCDs). It recommends an append-only approach: "To keep a history of the changes, you can add date columns to your dimension table... This approach is often called Slowly Changing Dimension Type 2." This directly supports using an append-only model with a timestamp to track history.

3. Google Cloud Documentation, Introduction to designing schemas, "Denormalization, nested, and repeated fields". This page explains that denormalization is a common strategy for increasing read performance and that using nested and repeated fields is the recommended way to do this in BigQuery, which aligns with the chosen answer's approach for simplicity and performance.

BigQuery does not support a direct batch load operation from Google Cloud SQL. To ingest data from Cloud SQL into BigQuery, you must use an intermediate step. The standard and recommended pattern is to first export the data from Cloud SQL into a supported file format (e.g., CSV, Avro) and save it to Google Cloud Storage. From there, you can initiate a BigQuery load job to ingest the file. Alternatively, you can use federated queries to query Cloud SQL data directly from the BigQuery interface and then use an INSERT statement to materialize the results into a BigQuery table, but this is a query-based transfer, not a direct "load" source.

A. BigQuery supports uploading data directly from a local file system via the Google Cloud Console, which is suitable for smaller datasets.

B. BigQuery can load data from files (like CSVs) stored in Google Drive and can also create external tables that are directly linked to Google Sheets.

C. Google Cloud Storage is the primary, recommended, and most scalable source for batch loading large volumes of data into BigQuery.

1. Google Cloud Documentation - Exporting data for import into BigQuery: This Cloud SQL documentation explicitly states the required workflow: "To load data from Cloud SQL into BigQuery, you must first export it to Cloud Storage. ... After you export the data to Cloud Storage, you can load it into BigQuery." This confirms Cloud SQL is not a direct load source.

Source: Google Cloud, Cloud SQL for MySQL Documentation, "Exporting data for import into BigQuery".

2. Google Cloud Documentation - Introduction to loading data into BigQuery: This page lists the primary ways to load data. It prominently features loading from Cloud Storage and local files. It does not list Cloud SQL as a source for a load job.

Source: Google Cloud, BigQuery Documentation, "Introduction to loading data into BigQuery".

3. Google Cloud Documentation - Querying Cloud SQL data: This document describes federated queries, which allow BigQuery to query Cloud SQL data in place. It clarifies this is a query mechanism, not a load operation. It states, "BigQuery Cloud SQL federation enables BigQuery to query data residing in Cloud SQL in real time, without copying or moving data."

Source: Google Cloud, BigQuery Documentation, "Querying Cloud SQL data".

4. Google Cloud Documentation - Creating tables from data in Google Drive: This document confirms that Google Drive is a supported data source for creating BigQuery tables.

Source: Google Cloud, BigQuery Documentation, "Creating tables from data in Google Drive".

The most likely cause for a successful load job in BigQuery resulting in a byte-for-byte mismatch with the source CSV is that some rows in the source file were invalid and subsequently skipped. BigQuery's load job configuration includes the maxBadRecords property. If this is set to a value greater than 0, the job is permitted to skip malformed rows (e.g., rows with an incorrect number of columns) and still complete successfully. The job is considered successful, but the resulting table will not contain the skipped rows, leading to a discrepancy with the source file.

A. The CSV data loaded in BigQuery is not flagged as CSV.

If the data format was not correctly specified as CSV, the load job would likely fail due to parsing errors or load the data incorrectly (e.g., each line as a single string), which is a more severe issue than a byte-for-byte mismatch.

C. The CSV data loaded in BigQuery is not using BigQuery’s default encoding.

BigQuery expects UTF-8 encoding for CSV files. An encoding mismatch would typically cause parsing errors, leading to a failed job. For the job to succeed, it would have to be configured to skip these errored rows, which points back to the mechanism described in option B.

D. The CSV data has not gone through an ETL phase before loading into BigQuery.

The absence of a pre-load ETL (Extract, Transform, Load) process is a common pattern (known as ELT). It is a design choice and not a direct technical cause for data being skipped during a successful load operation.

---

1. Google Cloud Documentation, "Loading CSV data from Cloud Storage": Under the bq load command options, the documentation for the --maxbadrecords flag states: "The maximum number of bad records to ignore. The default is 0, which causes the job to fail if any invalid rows are found." This directly supports the explanation that setting this value allows a job to succeed while skipping invalid rows.

Source: Google Cloud Documentation, BigQuery > Guides > Load data > Loading CSV data from Cloud Storage.

2. Google Cloud Documentation, "JobConfigurationLoad" API Reference: The description for the maxBadRecords field in the JobConfigurationLoad resource confirms its function: "The maximum number of bad records that BigQuery can ignore when running the job... If the number of bad records exceeds this value, an invalid error is returned in the job result. The default value is 0, which requires all records to be valid." This explains the mechanism for a successful job with missing data.

Source: Google Cloud Documentation, BigQuery > Reference > REST > v2 > jobs.

3. Google Cloud Documentation, "Troubleshooting data loading errors": This guide discusses common errors, including "CSV parsing error". It notes that such errors occur when the data doesn't conform to CSV formatting rules. The resolution often involves either fixing the source data or using load options like allowJaggedRows or maxBadRecords to ignore the problematic rows.

Source: Google Cloud Documentation, BigQuery > Support > Troubleshoot > Troubleshooting data loading errors.

The most critical requirement is to combine the price data with other BigQuery data "as cheaply as possible." This refers primarily to query costs, which dominate overall costs in analytical workloads. Loading the small (100-row) price dataset into a native BigQuery table is the most performant and cost-effective solution for the join operations. Batch loads into BigQuery are free, and the storage cost for such a small table is negligible. By having the data as a native table, BigQuery's query engine can perform highly optimized, fast, and inexpensive joins against the large customer datasets. The other options introduce significant performance overhead or unnecessary complexity and cost.

B: Federated queries on Cloud Storage are generally slower and less efficient than queries against native BigQuery tables, especially for frequent join operations, which would increase overall query costs and time.

C: Using Cloud Dataflow to programmatically join data from BigQuery and Cloud Datastore is an expensive anti-pattern for this use case. It introduces high latency and significant processing costs for a task that a simple SQL join can handle efficiently.

D: Similar to option C, using Cloud Dataflow to join data from a file in Cloud Storage is unnecessarily complex and costly. This approach is designed for large-scale ETL, not for efficient analytical querying.

1. Google Cloud Documentation - Introduction to external data sources: "Performance. Queries that include external tables might be slower than queries that only read BigQuery managed tables. If query speed is a priority, load the data into BigQuery instead of setting up an external data source." This directly supports why option A is superior to B for performance, which impacts cost.

2. Google Cloud Documentation - BigQuery Pricing - Loading and extraction: "Loading data... Batch loading data into BigQuery is free of charge." This confirms that the data ingestion method proposed in option A does not incur a cost, making it highly economical.

3. Google Cloud Documentation - Optimize query performance - Optimize your join patterns: The best practices for optimizing joins, such as ordering the largest table first, are most effective when all data resides within BigQuery's managed storage. This allows the query execution engine to have full control over data distribution and shuffling, which is not possible with external tables, reinforcing the choice of a native table for optimal join performance.

The core issue is that a single query is attempting to access more than 1,000 tables, which is the hard limit in BigQuery. The current date-sharding approach (a new table for each day) results in approximately 1,095 tables for a three-year period, thus exceeding this limit.

The officially recommended solution for this scenario is to migrate from date-sharded tables to a single, time-partitioned table. By consolidating all daily logs into one table partitioned by date, a query over any time range will only ever reference this single table. BigQuery's query planner will then use the date filter to efficiently scan only the relevant partitions (days), resolving the 1,000-table limit issue while also improving performance and simplifying data management.

A. Convert all daily log tables into date-partitioned tables: This option is less precise and potentially misleading. The goal is to create a single partitioned table to consolidate the data, not to convert each of the 1,000+ daily tables into its own partitioned table.

C. Enable query caching so you can cache data from previous months: Query caching stores the results of successfully executed queries. It cannot prevent a query from failing in the first place due to exceeding fundamental service limits like the number of tables referenced.

D. Create separate views to cover each month, and query from these views: A view is a logical representation of a query. A query against multiple monthly views would still expand to reference all the underlying daily tables, ultimately hitting the same 1,000-table limit.

1. Google Cloud Documentation - Introduction to partitioned tables: "When you create tables that are sharded by date, you have to create a new table for each date. Another approach is to use time-partitioned tables. ... Partitioning is the recommended approach over table sharding." This source establishes that partitioning is the correct replacement for sharding.

2. Google Cloud Documentation - BigQuery Quotas and limits: Under the "Query jobs" section, the limit for "Maximum number of tables referenced per query" is listed as "1,000". This confirms the root cause of the user's problem.

3. Google Cloud Documentation - Migrating sharded tables to partitioned tables: This document provides a tutorial for the exact solution. The introduction states, "This tutorial shows you how to migrate data from multiple, date-sharded tables into a single time-partitioned table." This directly supports option B as the correct procedure.

The core requirements are to prevent duplicate processing of orders by different sales agents and to avoid adding complexity. Pub/Sub's exactly-once delivery feature is designed for this specific use case. By enabling this feature on the pull subscription, the Pub/Sub service guarantees that a message is delivered to a subscriber application only once. This prevents two different agents from receiving and processing the same order message. This solution is a native feature of Pub/Sub, requiring only a configuration change on the subscription, thereby adhering to the constraint of not adding complexity to the workflow.

A. Create a transactional database that monitors the pending messages.

This adds a new, complex component (a database) and application logic to manage message state, which violates the requirement to not add more complexity.

B. Create a new Pub/Sub push subscription to monitor the orders processed in the agent's system.

This is a reactive monitoring solution, not a preventative one. It doesn't stop duplicate processing from occurring; it only reports on it after the fact and adds complexity.

D. Use a Deduplicate PTransform in Dataflow before sending the messages to the sales agents.

Introducing a Dataflow pipeline is a significant architectural change and adds substantial complexity, which is inappropriate for a manual agent workflow and violates a key constraint.

1. Google Cloud Documentation - Pub/Sub, "Exactly-once delivery": "When you enable exactly-once delivery, Pub/Sub provides the following guarantees... A message is delivered once and only once to the subscriber clients... If a message is sent to a subscriber client and the acknowledgment for the message is lost, Pub/Sub resends the message. However, Pub/Sub does not resend the message if the acknowledgment was successfully processed by the Pub/Sub server." This directly supports using the built-in feature to prevent duplicate delivery.

2. Google Cloud Documentation - Pub/Sub, "Create subscriptions": The process for enabling exactly-once delivery is shown as a simple checkbox or flag (--enable-exactly-once-delivery) during subscription creation. This confirms it is a low-complexity configuration change, not a major architectural addition. See the section "Create a subscription with exactly-once delivery".

3. Google Cloud Architecture Center - "Exactly-once processing with Cloud Dataflow and Pub/Sub": This document describes using Dataflow for deduplication in large-scale streaming pipelines. The complexity described (windowing, state management) illustrates why this approach (Option D) is overly complex for the scenario in the question, which involves manual agent interaction.

The gcloud ml-engine local train command is specifically designed to run a training job on a local machine. This feature allows developers to test and debug their training application using the Google Cloud SDK without submitting it to the cloud. It simulates the Cloud AI Platform (formerly ML Engine) environment locally, which is efficient for iterating on code before scaling up on the managed cloud service. Note that ml-engine is the former name, now aliased to ai-platform, but the command structure remains valid for exam purposes.

B. gcloud ml-engine jobs submit training: This command submits a training job to be executed on the managed Google Cloud AI Platform service, not on the local computer.

C. gcloud ml-engine jobs submit training local: This is an invalid command. The local argument is not used with the jobs submit training subcommand; it is its own subcommand (local train).

D. You can't run a TensorFlow program on your own computer using Cloud ML Engine. This is factually incorrect. The gcloud command-line tool for AI Platform explicitly provides a local training capability for development and testing purposes.

---

1. Google Cloud Documentation, gcloud ai-platform local train: "The gcloud ai-platform local train command runs your training application locally in an environment that is similar to that of a live Cloud AI Platform Training job." The documentation notes that ml-engine is the previous command group name.

Source: Google Cloud SDK Documentation, gcloud ai-platform local train command reference.

2. Google Cloud Documentation, "Running a training job": "To test your application before running it on AI Platform Training, you can run it locally with the gcloud ai-platform local train command." This guide clearly distinguishes between local training for testing and submitting a job to the cloud.

Source: Google Cloud, AI Platform Training Documentation, "Get started with training on AI Platform".

3. Google Cloud Documentation, "Using the gcloud command-line tool": This page describes the overall structure of gcloud commands for AI Platform, showing local as a distinct mode of operation from submitting jobs to the cloud service.

Source: Google Cloud, AI Platform Training Documentation, "Using the gcloud command-line tool".

To reduce the number of rows processed and control costs in BigQuery, you should minimize the amount of data scanned. Partitioning a table (e.g., by date) allows BigQuery to scan only the relevant partitions when a filter is applied on the partition column, a technique known as partition pruning. Similarly, splitting a large table into smaller, sharded tables (e.g., table2023, table2024) allows queries to target only the specific tables needed, thereby avoiding a full scan of the entire dataset. Both methods effectively reduce the number of rows processed.

B. The LIMIT clause is applied after data is scanned; it only restricts the number of rows returned, not the number of rows processed.

C. The LIMIT clause does not reduce the amount of data scanned from storage, which is the primary factor in BigQuery's on-demand pricing.

D. Using LIMIT does not affect the volume of data read by the query engine and therefore does not reduce the number of rows processed.

1. Google Cloud Documentation - Introduction to partitioned tables: "When you query a partitioned table, you can use a predicate filter on the partitioning column to reduce the amount of data that the query scans. This is known as partition pruning." This directly supports the use of partitions to reduce data processed.

2. Google Cloud Documentation - Best practices for controlling costs in BigQuery: Under the section "Avoid SELECT ", it states: "Applying a LIMIT clause to a SELECT query does not affect the amount of data read. You are billed for reading all bytes in the entire table, and the query counts against your free tier quota." This confirms that LIMIT is not a valid method for reducing processed data.

3. Google Cloud Documentation - To partition or to shard?: "Before partitioned tables were available, sharding tables was a common practice... If you query a subset of the shards, the cost of the query is lower because BigQuery only scans the tables specified in the query." This confirms that querying a subset of sharded tables reduces the data scanned.

The most efficient and Google-recommended method for migrating an on-premises Redis database to Memorystore for Redis is by using the native backup and import functionality. This process involves creating a standard Redis Database (RDB) snapshot file from the source instance, uploading this file to a Cloud Storage bucket, and then using the gcloud redis instances import command to load the data into the new Memorystore instance. This approach is optimized for bulk data transfer, minimizing migration time, cost, and the engineering effort required compared to developing a custom solution or using intermediate steps.

A. This adds an unnecessary intermediate migration to a Compute Engine instance, which increases complexity, cost, and overall migration time.

B. Writing a custom shell script is a high-effort, error-prone approach that is significantly less efficient than using the native RDB import feature.

C. Using Dataflow is overly complex and expensive for a direct database migration. It is designed for large-scale data processing, not a lift-and-shift database transfer.

1. Google Cloud Documentation - Memorystore for Redis, "Importing data into a Redis instance": This document explicitly outlines the recommended procedure. "To import an RDB file into a Memorystore for Redis instance, you first upload the RDB file to a Cloud Storage bucket, and then you initiate the import operation. The source RDB file can come from any Redis environment..." This confirms the steps in option D are the official, documented method.

Source: Google Cloud Documentation, "Importing data into a Redis instance", Overview section.

2. Google Cloud Documentation - Memorystore for Redis, "Migrating Redis to Memorystore for Redis": This guide details migration strategies. For offline migration, it states: "The simplest way to migrate your data is to use the import/export feature for Memorystore for Redis. You can export your data from your existing Redis instance to an RDB file, upload it to Cloud Storage, and then import it into your new Memorystore for Redis instance." This directly supports option D as the simplest, recommended practice.

Source: Google Cloud Documentation, "Migrating Redis to Memorystore for Redis", Offline migration section.

3. Google Cloud Documentation - gcloud redis instances import: The command-line reference for gcloud shows the specific tool used for this task. The command gcloud redis instances import SOURCE INSTANCE --region=REGION requires a SOURCE argument which is the "Google Cloud Storage location of the RDB file." This confirms the tooling and process described in option D.

Source: Google Cloud SDK Documentation, gcloud redis instances import.

Hotspotting in Bigtable with time series data occurs when row keys are based on sequential timestamps, causing all new writes to target a single node. The preferred method to mitigate this is field promotion. This schema design technique involves prepending a high-cardinality, non-sequential value from the data—such as a metricid or deviceid—to the timestamp in the row key (e.g., [metricid]#[timestamp]). This design leverages the natural diversity of the promoted field to distribute write operations across many different row key prefixes, and therefore across multiple nodes in the cluster, while still keeping data for a single metric ordered by time.

B. Randomization: This approach destroys the lexicographical ordering of the row keys, making efficient time-based range scans—a critical requirement for time series analysis—impossible.

C. Salting: While salting (adding a computed prefix) is a valid technique to prevent hotspotting, field promotion is generally preferred as it uses the natural cardinality of the data, which can result in a simpler and more intuitive schema.

D. Hashing: Hashing the entire row key has the same drawback as randomization: it breaks the natural sort order required for time-range scans.

1. Google Cloud Documentation, "Schema design for time series data": This document explicitly warns against using a raw timestamp at the beginning of a row key. It recommends, as a primary solution, prepending a value with high cardinality. The example provided is [metricname]#[timestamp], which is a direct implementation of the field promotion technique. (See section: "Metric name, then timestamp").

2. Google Cloud Documentation, "Schema design best practices": In the section "Row keys that are related to timestamps," the documentation advises against using a raw timestamp as the row key to avoid hotspotting. The recommended example, [deviceid]#[reversedtimestamp], demonstrates promoting the deviceid field to the beginning of the row key to distribute the load.

3. Designing a schema for time series data on Google Cloud Bigtable, Google Cloud Blog (2022): This article reinforces the principle, stating, "The most common cause of hotspotting with time series data is using the timestamp as the row key... The solution is to add a prefix to the row key that will ensure that your writes are distributed across multiple tablets." It then presents [metricid]#[timestamp] as the first and most common pattern.