Practice Test #1

A custom GKE microservice that polls AlloyDB is not CDC and will struggle to meet <10s latency and 99.9% delivery reliability at 2,500 changes/sec without significant custom engineering (deduplication, ordering, retries, backpressure, schema drift handling). Polling also increases load on the source and is operationally complex. It violates the requirement to use Google-managed services that automatically scale and handle schema changes without downtime.

BigQuery federated queries (via external connections) are for querying data in place, not continuously replicating it into BigQuery. They typically won’t meet the requirement to deliver changes with under-10-second end-to-end latency into BigQuery tables for downstream analytics/ML, and performance/cost can be unpredictable at high change rates. It also doesn’t provide a durable delivery pipeline with 99.9% delivery reliability guarantees for replicated data.

Database Migration Service is designed for migrations and continuous replication between databases (e.g., on-prem/MySQL/PostgreSQL to Cloud SQL/AlloyDB). It is not the standard managed service for streaming CDC directly into BigQuery as a target. Even if you could stage data elsewhere, it would add components and latency and would not match the requirement for automatic schema change handling and a direct, scalable, managed path into BigQuery.

Bigtable is a fully managed wide-column NoSQL database optimized for very high throughput and low-latency key/value access (time-series, IoT, analytics serving). It is not ANSI SQL relational and does not provide globally externally consistent, multi-region transactional semantics like Spanner. While it can replicate for availability, it is not the right fit for cross-region ACID transactions with strict consistency and online relational schema evolution requirements.

Firestore is a fully managed document database with strong consistency and multi-region replication options, and it can support global applications. However, it is not a relational ANSI SQL database and does not provide the same relational schema, joins, and SQL transaction model expected for an order reconciliation system described as relational. It also does not match the classic “external consistency + SQL + global transactions” requirement that points to Spanner.

Cloud SQL for MySQL is a managed relational database, but it is primarily regional. Cross-region designs typically rely on read replicas and asynchronous replication, which cannot guarantee RPO=0. Achieving concurrent multi-region writes with global external consistency is not a Cloud SQL capability; failover changes the primary and introduces operational complexity and potential data loss depending on replication mode. Meeting 99.99% globally with active-active writes is not its target use case.

Incorrect. Cloud Data Fusion is primarily an ETL and data integration platform, not a purpose-built transactional database migration service for low-downtime cutovers. A table-by-table pipeline approach does not naturally preserve commit ordering, transactional consistency, or continuous synchronization for a high-traffic payment ledger. It would also require substantial downtime or custom reconciliation logic to handle changes occurring during the migration. That makes it a poor fit for a 1.6 TB OLTP database with a strict sub-3-minute cutover requirement.

Incorrect. mysqldump is a logical export tool and is generally too slow for a 1.6 TB production database when downtime must remain under 3 minutes. Even if the export were compressed, the combined time for dumping, transferring, importing, and validating the data would be far longer than the allowed outage window. This method also places heavy load on the source and target during export and import, which increases operational risk. It is acceptable for smaller databases or maintenance windows with long downtime, but not for this scenario.

Incorrect. Exporting tables to CSV and importing them into Cloud SQL is a manual bulk-load pattern that is unsuitable for a transactional MySQL migration of this size and criticality. CSV workflows do not preserve full database semantics such as triggers, routines, foreign keys, and other schema-level objects without significant extra work. Rebuilding indexes and constraints after import would add even more time and complexity, making the outage far exceed 3 minutes. This approach is better suited to analytical data movement than to a payment ledger requiring low-downtime migration.

Setting a maintenance window helps control when maintenance occurs, but it does not prevent downtime on a single Cloud SQL instance because patching often requires a restart. Additionally, Cloud SQL maintenance windows are configured in UTC (not a local time zone), so “02:00–03:00 America/Chicago” is not how the setting is applied. Sequencing non-prod first is good practice, but it doesn’t solve production availability.

A read replica can offload read traffic, but it does not provide automatic, seamless continuity for a transactional (write-heavy) system during primary maintenance. Cloud SQL replicas are asynchronous; promoting a replica to primary is a disruptive operational event and typically requires connection endpoint changes and careful failover planning. This does not meet the requirement of avoiding downtime for users during operating hours.

Maintenance notifications and rescheduling are operationally helpful, but they don’t eliminate downtime. The requirement is to apply patches within 7 days and avoid downtime during 06:00–22:00 local time. Notifying users implies accepting downtime, which conflicts with the stated goal. Also, maintenance timing can be constrained by Google’s maintenance policies and the patch compliance window.

A clone of the primary is not the same as a read replica. Clones are independent copies and do not continue asynchronous replication from the primary after creation. Using a clone for read traffic would create data divergence over time and require more manual management. Therefore, it does not properly restore a managed read replica topology.

This option is incorrect because it relies on an assumed automatic recreation of the read replica in a healthy zone. Cloud SQL does not provide a documented guarantee that a single read replica will be automatically recreated elsewhere to preserve read availability after a zonal disruption. Verifying status is a reasonable operational check, but it does not itself ensure the read-only workload continues to function. The question asks what you should do to ensure availability, not merely what you should observe.

Restarting the primary is unnecessary and potentially harmful. The problem is with the replica being unreachable due to a zonal network disruption, not with the primary database process. Restarting the primary can interrupt write traffic and does not guarantee restoration of the replica or read service. It adds risk without addressing the actual failure domain.